A novel graph-based framework for

state of health prediction of lithium-ion

battery Xing-Yan Yao
Visit to download the full and correct content document:
https://ebookmass.com/product/a-novel-graph-based-framework-for-state-of-health-pr
ediction-of-lithium-ion-battery-xing-yan-yao/
 Journal of Energy Storage 58 (2023) 106437

Contents lists available at ScienceDirect

Journal of Energy Storage

journal homepage: www.elsevier.com/locate/est

Research Papers

A novel graph-based framework for state of health prediction of

lithium-ion battery
Xing-Yan Yao a, *, Guolin Chen a, Michael Pecht b, Bin Chen c
a
Chongqing Engineering Laboratory for Detection Control and Integrated System, Chongqing Technology and Business University, Chongqing 400067, China
b
Center for Advanced Life Cycle Engineering, University of Maryland, College Park, MD 20742, USA
c
China Merchants Testing Vehicle Technology Research Institute Co., Ltd, Chongqing 400067, China

A R T I C L E I N F O A B S T R A C T

Keywords: State of health (SOH) plays a vital role in lithium-ion batteries (LIBs) safety, reliability and lifetime. Health
Lithium-ion battery (LIB) indicators (HIs) are a powerful approach to predict battery SOH. The existing methods for battery SOH prediction
State of health (SOH) according to HIs only consider the temporal features of HIs. The spatial features of interdependence between HIs
GraphSAGE
enrich predicational information especially for the limited data. This paper proposes a novel framework CL-
Graph neural network (GNN)
GraphSAGE to predict battery SOH based on graph neural network (GNN), which takes into both temporal
and spatial features of HIs. Firstly, the Pearson's correlation coefficients between HIs and SOH are obtained to
extract highly correlated HIs to build a graph. Subsequently, the temporal features are extracted by convolutional
neural network (CNN) and long short-term memory neural network (LSTM). Finally, the spatial features are
obtained by the graph sample aggregate (GraphSAGE) to propagate messages on a pre-defined graph structure,
which uncovers the deep information among HIs. The effectiveness of the proposed approach in predicting
battery SOH is verified by MIT, NASA datasets and the experimental datasets, and compared with CNN, LSTM
and graph convolutional network and graph attention network. The results show that the root mean square error
of the proposed approach CL-GraphSAGE can achieve 0.2 %, and the different datasets verify its feasibility.

1. Introduction
In response to extreme weather and environmental pollution, electric
vehicles are widely used in the world. Lithium-ion batteries (LIBs) are a
promising energy source for the electric vehicle due to their high energy,
low self-discharge rate, and long storage life [1,2]. Nevertheless, LIBs
performance degrades with cycle increases. The battery failure can lead
to serious or even safety accidents. Therefore, it is vital to monitor LIBs'
state and master the battery health. The state of health (SOH) of the
battery is a powerful tool to evaluate the battery current health condi­
tion and ensure battery safety and health management [3]. SOH is a
percentage of the maximum charge which can be released to the battery
rated capacity. However, SOH is difficult to predict since it cannot be
tested directly such as via terminal voltage. Moreover, the battery data
collected in the real world is a limitation. It is a challenge to obtain
precise SOH due to the lack of data and indirectly measured parameters
[4].
The methods for battery SOH prediction can be summarized as
model-based and data-driven [5]. Model-based methods include elec­
trochemical models [6,7], equivalent circuit models [8–11], and math­
ematical models [12–14]. The electrochemical model-based method
develops a model to describe the physical and chemical processes of a
battery and then evaluate the influence of the aging process on the SOH.
This method can be generally used for LIBs to obtain battery aging
trajectories. However, the electrochemical model-based method re­
quires lots of parameters, and SOH prediction performance is deter­
mined by the given parameters. In addition, the computational
complexity is high. The equivalent circuit model is based on battery
electrical characteristics combined with amounts of the dataset, and the
lithium battery is equivalent to a circuit. Compared with the

Abbreviations: LIBs, lithium-ion batteries; SOH, state of health; HIs, health indicators; CL-GraphSAGE, the proposed method in this paper; GNN, graph neural
network; GraphSAGE, graph sample aggregate; CNN, convolutional neural network; RNN, recurrent neural network; LSTM, long short-term memory neural network;
CC, the current curve of the constant current; CV, the current curve of the constant voltage; G = (V, E ), graph G, E and V are the edge and node set; RMSE, root
mean square error; MAPE, the average absolute percentage error; R2, coefficient of determination.
* Corresponding author.
E-mail address: yaoxingyan-jsj@163.com (X.-Y. Yao).

https://doi.org/10.1016/j.est.2022.106437
Received 6 June 2022; Received in revised form 14 August 2022; Accepted 14 December 2022
Available online 22 December 2022
2352-152X/© 2022 Elsevier Ltd. All rights reserved.
X.-Y. Yao et al.
electrochemical model, the computational complexity is less. Further­
more, mathematical models obtain the battery performance by large
amounts of experimental data [4]. The SOH prediction is achieved via
data analysis, fitting, trial and error, and statistical processing. In this
method, the battery physical mechanism is a black box, and the SOH
prediction is greatly affected by the experimental environment.
Data-driven methods achieve SOH prediction by utilizing battery
signals (current, voltage, temperature, internal resistance, etc.) or health
indicators (HIs) extracted from charging and discharging processes. This
method does not know the battery complex internal mechanism. With
the widespread application of artificial intelligence, machine learning
has been widely used in lithium battery SOH prediction [15]. Yang et al.
[16] employed Gaussian process regression to predict SOH by four
correlated HIs extracted from charging curves and analyzed the corre­
lation between HIs and SOH. Song et al. [17] extracted features to
describe the battery aging process and predict battery SOH. Yang et al.
[18] extracted HIs from voltage and temperature during discharge,
using a multiple kernel relevance vector machine that combined
Gaussian with sigmoid function to improve the accuracy of SOH pre­
diction. In addition, other machine learning methods such as random
forest [19], Markov chain [20], Monte Carlo [21] and ARIMA [22] are
also popular to use.
Although machine learning methods mentioned above achieved
great success in processing nonlinear relationships, they are still can be
improved to process historical data to predict battery SOH. Many re­
searchers applied recurrent neural network (RNN) [23], gate recurrent
unit (GRU) [24] and long-short term memory (LSTM) [25] to predict
SOH. In addition, some researchers have used neural ordinary differ­
ential equations and recurrent neural networks for the analysis of bat­
teries to predict SOH [26]. Kim et al. [27] used multiple features as the
input of LSTM and changed the input sequence length to reflect the
historical SOH. Cui et al. [28] proposed a dynamic spatial-temporal
attention-based gated recurrent unit (DSTA-GRU) model, which
considered both the time domain and space domain of features on SOH.
The methods above mentioned typically used the historical data of HIs to
establish the mapping between HIs and SOH. However, they ignore the
potential interdependence between HIs and the interdependence be­
tween HIs on the prediction results either. Furthermore, it is urgent to
deep mine hidden information from the limited battery data in the
application.
The graph is a special type of data structure that describes the rela­
tionship between different entities. For the problem of multi-variate
inputs, the graph uses variables as nodes, and the nodes are connected
through their hidden dependency. In recent years, graph neural network
(GNN) can effectively deal with complex graph structures and has
attracted the attention of researchers in various fields [29–31]. GNN
relies on the relationships between nodes to capture information in the
graph and optimizes the message-passing mechanism between graph
nodes to get the dependence between different variables.
GNN has already used to prediction task right now. Yu et al. [32]
used spatial-temporal graph convolutional networks to extract spatial-
temporal features from the time series of the graph structure to pre­
dict traffic flow. Khodayar et al. [33] proposed a novel graph convolu­
tional deep learning structure to predict short-term wind speed
combined with LSTM. Vyas et al. [34] proposed a dynamic GNN to
predict the soil moisture of multiple locations in a region over time.
Wang et al. [35] construct a graph structure on a sensor network with
Pearson Correlation Coefficients among sensors. In addition, they
combined graph convolutional network and temporal convolutional
networks to forecast the remaining useful life of the engine.
The data-driven approach can mine deep information in the limited
data without much prior knowledge. The CNN is a powerful deep
learning approach to extract local features. The LSTM has the advantage
of long-term memory in sequence modeling to overcome the problems of
gradient vanishing in the long sequence training. To deal with graph
structure data, GNN is a kind of algorithm that uses neural networks to
2
Journal of Energy Storage 58 (2023) 106437
learn graph structure data, extract and explore features in the graph
structure data. GNN is a promising method to deal with time series
prediction for data of spatial-temporal characteristics. The spatial
characteristic is reflected by deep mining the interdependence between
nodes of the features in the GNN, and the temporal characteristic can
also be taken into considering the historical data according to the LSTM.
As a kind of GNN, GraphSAGE can aggregate the information from nodes
of the graph, and the message passing is used to learn features in the
graph structure. Therefore, the fusion of CNN, LSTM and GraphSAGE
can mine deep information from both the spatial features and temporal
features of lithium-ion battery data to achieve SOH prediction. As far as
we know, little research on SOH using GNN is demonstrated.
In this paper, a novel framework CL-GraphSAGE for battery SOH
prediction is proposed. The HIs extracted from the charging and dis­
charging process are converted into a graph structure. The combination
of GNN and CNN-LSTM is employed for the node features and message
passing information to obtain the spatial-temporal features of the graph.
The CNN-LSTM mines the potential temporal features of historical input
of HIs data [36–38], which are used as node features of the graph
structure. GraphSAGE is used to carry out message passing in the graph
structure composed of HIs. Therefore, all the HIs can obtain comple­
mentary information from each other. The main contributions of this
paper are as follows:
• A novel framework CL-GraphSAGE is proposed to predict lithium-ion
battery SOH, which considers both the temporal and spatial features
of HIs.
• The HIs highly correlated with SOH are extracted to form a spatial-
temporal graph.
• The relationship between different HIs is deep mined and optimized
to enrich the information to predict SOH.
• Graph neural network (GNN) is the first time to use for battery
prediction problems.
The rest of the paper is organized as follows: Section 2 describes the
lithium-ion battery datasets and feature extraction. Section 3 describes
the CL-GraphSAGE SOH prediction method. Section 4 describes the re­
sults and discussions, followed by conclusions in Section 5.
2. Datasets and feature extraction
This section describes two public lithium-ion battery datasets used to
validate the method proposed in this paper. Different HIs are extracted
for these two datasets.
2.1. Datasets
NASA datasets are used to verify the proposed method [39]. The
B0005, B0006, and B0007 batteries are used. All battery charging and
discharging processes are carried out at room temperature, with a
nominal capacity of 2.0 Ah. During charging, the battery is first charged
with a CC at 1.5 A until the voltage reaches 4.2 V; it is then charged with
a CV until the current drops below 20 mA. During discharging, the
battery is discharged with CC at 2A until cut-off voltage. Repeat the
above steps until the battery reaches the end of its life (capacity loss 70
%). Fig. 1(a) shows the capacity curves of cycles.
The MIT lithium-ion battery dataset is currently the largest public
dataset used for battery degradation research [40]. The dataset consists
of 124 commercial LIBs from A123. The nominal capacity and voltage
are 1.1 Ah and 3.3 V, respectively. At 30 ◦ C, all batteries are charged
with a multistep fast-charging strategy and then discharged with a CC
process. The format of this strategy is “C1(Q1)-C2.” The battery is firstly
charged by current C1 until the state of charge reaches S1, and then it is
charged with current C2 until the state of charge reaches S2. The battery
is charged from 80 % to 100 % SOC at 1C CC-CV. The cutoff voltage is
3.6 V. Subsequently, the battery is discharged at 4C CC, and the battery
X.-Y. Yao et al. Journal of Energy Storage 58 (2023) 106437

Fig. 1. Capacity degradation curves: (a) NASA datasets; (b) MIT datasets.

life threshold is reduced to 80 % of the nominal capacity. The batteries • Voltage-related HIs. The HIs extracted from the voltage curve are
A0001, A0002, and A0003 are employed in the experiments. Fig. 1(b) similar to the current-related HIs. The areas covered under the
shows the capacity degradation of the specific cycle period. voltage curve of the CC process, CV process, and discharge process
are HIs. The maximum slope extracted from the CC process voltage
curve is also an HI.
2.2. Feature extraction
• Temperature-related HIs. During the charging and discharging pro­
cess, the temperature changes. The maximum temperature, mini­
The HIs used for data-driven approaches depend on the dataset.
mum temperature, and average temperature are taken as HIs. The
Different HIs are extracted from NASA and MIT datasets. All the
areas covered under the temperature curve of the CC process, CV
extracted HIs mentioned above are described in Table 1. The HIs
process CC-CV process and the entire charge and discharge cycle are
involved in this article are divided into four categories, which are
HIs.
voltage-related, current-related, temperature-related, and time-related
• Time-related HIs. The HIs include CC charging time, CV charging
[41–43].
time, charging time, and discharging time. In addition, the time for
the temperature to reach the peak during the discharge process is
• Current-related HIs. The areas covered under the current curve of the
also an HI.
CC process and the discharge process, the area covered under the
current curve of the CV process, the CC-CV process, and the
The HIs extracted are multidimensional features. Different HIs have
discharge process are as HIs. In addition, the maximum slope
different degrees of correlation with SOH. If all HIs are used as input, it
extracted from the CV process current curve is another HI.
will affect the prediction accuracy and easily lead to overfitting.
Therefore, given a threshold, the HIs with a lower correlation are
Table 1 eliminated, and the HIs with a higher correlation as the input of the
Four categories of HIs.
model are kept. Pearson correlation coefficient [44] is used to evaluate
HI type No. HIs the correlation between HIs and SOH.
Current-related B1 Area covered under current curves of the CC process The Pearson correlation coefficient calculates the linear correlation
B2 Area covered under current curves of CV process between HI and SOH in Eq. (1). Here, Xi and Yi are HI and SOH,
B3 Area covered under current curves of CC-CV process respectively. μX and μY are the average values, respectively. The absolute
B4 Maximum slope of current curves in CV process
B5 Area covered under current curves of discharge process
value of Pearson's correlation coefficient is in the range of [0,1]. The
Voltage-related B6 Area covered under voltage curves of CC process larger the absolute value, the stronger the correlation. In this paper, the
B7 Area covered under voltage curves of CV process 6 most relevant HIs with large absolute values are selected as the final
B8 Area covered under voltage curves of discharge process subset of HIs by the trials and errors of the experiments.
B9 Maximum slope of voltage curves in CC process ⃒ ⃒
Temperature- B10 Maximum temperature of charge and discharge cycles ⃒ ∑n ⃒
⃒ ⃒
related B11 Minimum temperature of charge and discharge cycle ⃒ (X − μ )(Y − μ ) ⃒
(1)
i=1 i X i Y
B12 Average temperature of charge and discharge cycle
ρ(X, Y) = ⃒√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ √ ̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
̅ ⃒
⃒ ∑n 2 ∑n 2 ⃒
B13 Area covered under temperature curves of charge and ⃒ i=1 (X i − μ X ) i=1 (Y i − μ Y ) ⃒
discharge cycle
B14 Area covered under temperature curves of the CC
process 3. Methodology
B15 Area covered under temperature curves of CV process
B16 Area covered under temperature curves of CC-CV This section describes in detail the proposed approach CL-
process GraphSAGE for the battery SOH prediction in this paper. The compo­
Time-related B17 Discharge time
B18 Charging time
nents in CL-GraphSAGE are depicted in detail, and all steps of SOH
B19 CC stage period prediction by CL-GraphSAGE are detailed descriptions.
B20 CV stage period
B21 CC stage end time
B22 Time to peak temperature during discharge 3.1. CL-GraphSAGE
B23 The time ratio of the CC process and the entire charging
process
Fig. 2 shows the proposed CL-GraphSAGE for battery SOH prediction
B24 The time for voltages increases from 3.7 V to 4.2 V
during the CC process framework. The CL-GraphSAGE includes CNN, LSTM and GraphSAGE.
B25 Time for currents decreases from 1.5 A to 0.3 A during The extracted HIs in Section 2 are used to build a graph structure. Then
the CV process the HI's temporal features are extracted by CNN and LSTM, and the
B26 Time for voltages decreases from 3.7 V to 2.7 V during spatial features are obtained by GraphSAGE. Finally, the SOH is ac­
discharging
quired by multilayer perceptron (MLP).

3
X.-Y. Yao et al. Journal of Energy Storage 58 (2023) 106437

Fig. 2. CL-GraphSAGE framework for battery SOH prediction.

The details of CL-GraphSAGE are as follows:

Table 2
Pseudocode of CL-GraphSAGE.
Step 1: Extracting and filtrating HIs. The HIs are extracted from the
original datasets. The HI subsets are derived from the Pearson cor­ 1: Input: The dataset , .
2: Calculating similarity scores between HIs and building graph structure.
relation coefficient in Eq. (1).
3: Training:
Step 2: Building a graph structure. The input of GraphSAGE is a 4: for in do:
graph. The HI subset in Step 1 needs to be converted into a graph 5: ℎ ← ( );
structure that consists of nodes and edges. Every filtrating HI is taken 6: ℎ ← (ℎ);
as one node of the graph. The similarity values between the HIs 7: ℎ ← ℎ (ℎ);
calculated by Eq. (2) are taken as edges of the graph. Therefore, the 8: ℎ ← ℎ (ℎ);
9: ℎ ← (ℎ);
graph is represented by nodes set, and the HIs sequence data are the (ℎ);
10: ←
values of the nodes. 11: Update model parameter.
Step 3: Obtaining HIs temporal features. The HIs sequence data of 12: end for
each node is used as the input of the CNN-LSTM hybrid model. The 13: Output: The SOH of .
HIs temporal features are the hidden state vector at the last moment
of LSTM output.
Step 4: Updating node messages. The GraphSAGE is used to pass
messages between graph nodes. The node message is updated by the the similarity values from the Eq. (2) [45]. xi and xj are the sample
center node, which receives all the connected node messages. All sequence at i-th and j-th nodes, respectively. xi = xi /d is the average
nodes are aggregated to obtain a global representation of the entire value of xi, and d is the number of elements of xi, and the range of i and j
graph. Finally, SOH is predicted by MLP. is [1, n].
( )
(xi − xi )T xj − xj
The pseudocode of the proposed CL-GraphSAGE is shown in Table 2. Aij = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅√(̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
)T ( )̅ (2)
(xi − xi )T (xi − xi ) xj − xj xj − xj

3.2. Graph structure of HIs
The determination of nodes and edges is the key to applying graph
neural network to battery SOH prediction. In this paper, HIs are regar­
ded as nodes in the graph, and HIs are connected as edges.
Assuming a graph G = (V, E ), where E and V = {v1, v2, …, vn} are
the edge and node set, respectively. Furthermore, the nodes consist of
node sets and the corresponding node values. Here, the HIs subsets are
the nodes of the graph, and the node values are obtained by the CNN-
LSTM hybrid model. |V| = n is the numbers of the HIs. For the i-th
node, vi ∈ V with T-dimension is denoted by xi. X = {x1, x2, …, xn} ∈ Rn×T
is the node feature vector.
For all the node vector V = {v1, v2, …, vn}, the HIs are connected by
After defining the nodes and edges to build the graph, the task of
SOH prediction is transformed to learn the mapping between HIs and
SOH by the given SOH value at the time t. Specifically, the HIs at
different timestamps are used as node data to generate temporal datasets
for the graph, and the SOH at the time t is used for prediction in the
graph.
3.3. Temporal feature of HIs
LSTM has been widely used in Lithium-ion battery SOH prediction.
LSTM overcomes the problem of gradient vanishing and captures the
long-term dependence of time series compared with RNN [46]. LSTM

4
X.-Y. Yao et al. Journal of Energy Storage 58 (2023) 106437

obtains the long-term dependence of time series data by using the ({ })

storage unit to store useful information and the gating mechanism to hkN (v) = Aggregate hk−j 1 , ∀j ∈ N (i) (3)
discard useless information. LSTM structure includes input gates, forget
( ( ))
gates, storage units, and output gates. The input gate saves the infor­ hki = σ W k Concat hk−i 1 , hkN (v) (4)
mation, the forget gate obtains the forgetting information, the output
gate represents the output information, and the storage unit keeps the Here, hki is the state of the i-th node in the k-th message passing, and
historical information until the current moment. N (i) is the adjacent nodes set. In Eq. (4), the Aggregate function takes the
CNN is a powerful tool to extract local hidden information of data by place of the Mean aggregation function, and Eqs. (3)–(4) are combined
multi-layer convolution operations [47]. As shown in Fig. 3, CNN is used and modified to obtain Eq. (5).
for extracting deep features from the historical information, which are ( ({ } { }))
the input of LSTM to mine temporal features of the sequence data. The hki = ReLU W • Mean hk−i 1 ∪ hk−j 1 , ∀j ∈ N (i) (5)
feature extraction of every HI time series is separated from each other,
and a one-dimensional CNN layer with one kernel of size 1 is used. HIs After obtaining the final state of all nodes, the readout function is
temporal features obtained by CNN and LSTM are used as the node used to aggregate all node representations into a global representation of
features of the graph. The temporal features of every HI constitute the the graph [48]. The global expression is obtained by one-time aggre­
temporal feature matrix X of the graph. gation of all inputs. Here, the Max function is to aggregate all nodes, and
MLP outputs the final predicted value ̂ y t as shown in Eq. (6). The number
3.4. GraphSAGE of neurons in the output layer is 1 when the SOH value is at a certain
moment. Huber loss [36] is used as the regression loss function. yt and ̂ yt
For the SOH prediction, there are two keys to building the GNN are the true value and predicted value, respectively. δ is the hyper­
model. One is to fuse the local information of nodes in the graph parameter in Eq. (7).
structure, and the other is to embed the updated nodes into vectors. In ( ({ }) )
y t = MLP Readout hki |i ∈ G
̂ (6)
this paper, GraphSAGE aggregates the information of adjacent nodes to
update nodes and obtains the interdependence between HIs. GraphSAGE ⎧
⎪ 1
is used to mine deep information between HIs of the edges in the graph. ⎪
⎪
⎪ (yt − ̂y t )2 , |yt − ̂
yt | ≤ δ
⎨ 2
Therefore, each HI node contains all information of both HI itself and L = ( ) (7)
adjacent HIs from the k-layer GraphSAGE. ⎪
⎪
⎪
1
The input of GraphSAGE is the graph G = (V, E ). The node features ⎩ δ • |yt − ̂y t | − 2 δ , otherwise
⎪

X ∈ Rn×d, that are temporal features, are obtained in Section 3.3, where n
is the number of HIs and d is the dimension of LSTM output. The simi­ 3.5. Evaluation metrics
larity between HIs is calculated to obtain the connection between nodes
of adjacency matrix A ∈ Rn×n in Section 3.2. The adjacency matrix A and The experiments are carried out by PyTorch 3.8, Intel Core i7-
the feature matrix X are used as the input of GraphSAGE. 10875H 2.30GHz CPU, and NVIDIA GEFORCE RTX 2060 GPU. Adam
Fig. 4 shows the aggregation process of GraphSAGE for the given is the optimizer, the learning rate is 8e-4, the δ in the Huber loss objective
graph. Fig. 4(b) displays the corresponding topological and node fea­ function is set to 1, the number of training epochs is 1500, and the batch
tures by the given graph structure in Fig. 4(a). Taking the red node A in size is 16. Root mean square error (RMSE), average absolute percentage
Fig. 4(a) as an example, GraphSAGE updates the information of the node error (MAPE), and coefficient of determination (R2) are used to evaluate
A by fusing the neighborhoods' information. To update nodes at k-layer, performances, which are as follows:
the information of the adjacent nodes is first aggregated to hkN (v) by the √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
aggregation function and then integrated and updated with the central 1∑ n
RMSE = (yi − ̂ y i )2 (8)
node state hk−
i
1
, which are shown in Eqs. (3)–(4). Then the aggregated n i=1
k
embeddings hN (i) are obtained by the neighborhood aggregation which
is fused with the node information of the previous layer hk− 1
. 1∑ n
|yi − ̂y i |
i MAPE = × 100% (9)
n i=1 yi

h1 h2 h3 ht

LSTM LSTM LSTM LSTM

x1 x2 x3 xt

LSTM
Dimension
HIs LSTM

.. ..
Dimension

LSTM
concat
..
HIs

LSTM

LSTM
HIs Feature matrix
LSTM

Fig. 3. Temporal feature matrix of HIs.

5
X.-Y. Yao et al. Journal of Energy Storage 58 (2023) 106437

A
(a) (b)
B
E
Aggregate
A
D C B
D

Aggregate Aggregate Aggregate

C
F
A F A A E

Fig. 4. Aggregation process of GraphSAGE: (a) the given graph; (b) aggregated embedding.

∑
n carried out for the comparative experiments. The training data are
y i )2
(yi − ̂
R2 = 1 − i=1
∑n
(yi − yi )2
i=1
4. Experiments
This section validates the performances of CL-GraphSAGE to predict
the SOH for different lithium-ion battery datasets. The three different
datasets are employed to discuss the performance by comparing with
different models.
4.1. Experiment for NASA dataset
According to HIs in Table 1 described in Section 2.2 and battery
characteristics in Fig. 1(a). HIs are extracted from the NASA dataset. The
correlation coefficient between HIs and SOH is obtained according to Eq.
(2) shown in Table 3.
The first 6 high-correlation HIs (B1, B3, B6, B21, B24, and B26) are
adopted as model inputs from Table 3. The similarity of HIs is calculated
by Eq. (2). Fig. 5 shows the similarities between nodes. Two nodes with a
similarity of 0.95 or more are connected into an edge. Taking B0005 as
an example, as shown in Fig. 5(a), the similarities between HIs are all
greater than 0.99. The nodes with a similarity of more than 0.95 are
connected. Therefore, the six HIs are all connected as edges of the graph,
and the graph structure is shown in Fig. 5(d). In Fig. 5(a)–(c), the sim­
ilarity of HIs is greater than 0.95, and the graph structures are fully
connected graphs.
The starting point is of great influence on the prediction results. The
data before the starting point are used for training the model, and the
remaining data are used for testing. Two kinds of training strategies are
chosen as the first 80 cycles and the first 100 cycles, respectively, and
(10) the corresponding remaining cycles are used as test data.
For the performance comparison, all the models are designed as
multi-input and single-output. The 6 HIs selected above as the input of
the models, and the battery SOH is the output of the models. The RNN is
a classic approach to process sequence data and predicts battery SOH.
RNN is a kind of neural network with sequence data as input. The output
not only depends on the input at the current time but also is related to
the input at the previous time. LSTM is a variant of RNN to improve the
performance of RNN by overcoming the gradient vanishing when
dealing with long-term memory. To mine the deep temporal features for
the HIs sequence data, the CNN is combined with LSTM to predict SOH.
In addition, GNN is used in the proposed CL-GraphSAGE to mine rele­
vant information between HIs to obtain the spatial features of HIs.
Through comparative verification, GNN has improved the SOH predic­
tion accuracy. The performance of three models RNN, LSTM, and CNN-
LSTM are compared with CL-GraphSAGE.
Fig. 6 shows the results starting at 81of the four models in B0005,
B0006, and B0007. Table 4 shows the performance evaluation of the
four models from the different starting points. The RMSE of RNN and
LSTM are both larger than 0.04, and the MAPE is larger than 4.5. The
RMSE of CNN-LSTM and CL-GraphSAGE are both less than 0.035, and
the MAPE are both less than 3.5. In some experiments, R2 of LSTM and
RNN is even less than 0, which indicates that LSTM and RNN have poor
adaptability and poor predictive ability in these cases. Therefore, the
performance of RNN and LSTM in these three batteries is worse than that
of CNN-LSTM and CL-GraphSAGE. As cycle numbers increase, the pre­
dicted result tends to be a straight line and cannot reflect the nonlinear
capacity degradation trend. However, in Fig. 6(c), the predicted result of
CNN-LSTM is consistent with that of RNN and LSTM, which is quite
different from the true SOH.

Table 3
Descending order of correlation coefficient of HIs.
Sample Sort in descending order (from left to right, top to bottom)

B0005 B26 0.998255 B6 0.996169 B3 0.996047 B21 0.996046

B1 0.996038 B24 0.996032 B25 0.994137 B14 0.989574
B2 0.988875 B23 0.988454 B15 0.988875 B9 0.894633
B4 0.615588 B16 0.311356
B0006 B26 0.993706 B3 0.991352 B21 0.990314 B6 0.990308
B24 0.989723 B1 0.988990 B25 0.988074 B14 0.986618
B23 0.986446 B9 0.961140 B2 0.938323 B15 0.929285
B4 0.628557 B16 0.201432
B0007 B26 0.998064 B21 0.992642 B3 0.992558 B1 0.992296
B6 0.992217 B24 0.988901 B23 0.982533 B14 0.980917
B2 0.961393 B15 0.883181 B9 0.694442 B4 0.635587
B25 0.362327 B16 0.288703

The bold in the table are the selected HIs which is the input of the model in this case study.

6
X.-Y. Yao et al.
Fig. 5. Node similarity and graph structure: (a) B0005; (b) B0006; (c) B0007; (d) graph of B0005 HIs.
Fig. 6. Prediction results at starting point of 81: (a) B0005; (b) B0006; (c) B0007.
However, the predicted results by CL-GraphSAGE can still capture
the tendency of capacity degradation, which indicates the proposed CL-
GraphSAGE is feasibility to the battery SOH prediction. The RMSE and
MAPE of CL-GraphSAGE are the lowest among the three experiments. In
battery B0007, the RMSE and MAPE are 0.019028 and 2.383225,
respectively, which are much lower than CNN-LSTM. The prediction
results of CL-GraphSAGE show good performance. Compared with the
other three models in Table 4, CL-GraphSAGE makes full use of the
spatio-temporal features of HIs. CL-GraphSAGE can mine the relation­
ship information between HIs and is better for SOH prediction.
Fig. 7 shows results starting at 101 of the four models in B0005,
B0006, and B0007. According to Fig. 7, the prediction results of the
three models are closer to the actual capacity degradation trend by using
more training data. The RMSE and MAPE of RNN and LSTM are larger
than those of CNN-LSTM and CL-GraphSAGE.
Table 4 shows the prediction errors of the four models at different
starting points for different lithium-ion battery samples. In Table 4, both
7
Journal of Energy Storage 58 (2023) 106437
RMSE and MAPE of CL-GraphSAGE are the lowest compared with CNN-
LSTM, which shows good prediction performance. Moreover, CL-
GraphSAGE predicts accurately when the starting point is 101
compared with the starting point at 81. The results show that regardless
of whether the prediction starting point is 81 or 101, CL-GraphSAGE is
the best choice.
4.2. Experiment of MIT dataset
According to HIs in Table 1 described in Section 2.2 and battery
capacity in Fig. 1(b), all HIs extracted from the MIT dataset and the
correlation coefficient between HIs and SOH are obtained according to
Eq. (2) are shown in Table 5. The first six high-correlation HIs (B1, B5,
B6, B8, B17, and B19) are selected as model inputs.
The method to construct the nodes and edges of the graph is the same
as in Section 4.1. Fig. 8 shows the similarities between nodes. The
similarity values are displayed in Fig. 8(a)–(c). The two nodes are
X.-Y. Yao et al. Journal of Energy Storage 58 (2023) 106437

Table 4 which is the lowest RMSE, MAPE with the highest R2.
Predicting errors at different starting points by different models. Experiment results show that the proposed method is better than the
Sample Model Starting RMSE MAPE R2 comparison model in terms of SOH prediction. Compared with the other
points three models, CL-GraphSAGE can obtain not only the temporal features
B0005 RNN 81 0.062645 7.892090 − 1.247296 of HIs using CNN and LSTM but also the spatial features between HIs.
LSTM 0.058334 7.331871 − 0.948624 Therefore, CL-GraphSAGE can make full use of the interrelationships
CNN-LSTM 0.011572 1.338180 0.923311 between different HIs to deep mine more useful potential features. The
CL- 0.008281 1.105863 0.960732 fluctuation range of the prediction error of CL-GraphSAGE is smaller for
GraphSAGE
RNN 101 0.040255 5.148115 − 1.002669
SOH prediction compared with the other three models.
LSTM 0.018398 2.179895 0.581690
CNN-LSTM 0.006289 0.720849 0.951118 4.3. Experiment of laboratory dataset
CL- 0.005808 0.625176 0.958309
GraphSAGE
B0006 RNN 81 0.046791 5.955632 0.131731 In this section, the lithium-ion battery dataset carried out in the
LSTM 0.041631 5.451267 0.312677 laboratory is used to validate the feasibility of CL-GraphSAGE. The ex­
CNN-LSTM 0.013927 1.802672 0.923077 periments are carried out in China Merchants Testing Vehicle Technol­
CL- 0.011475 1.470694 0.947783 ogy Research Institute Co., Ltd. The BAT-NEEFLCT-05500 from NEBULA
GraphSAGE
RNN 101 0.036882 4.796715 0.130402
Co., Ltd. is used to cycle the batteries. The data is collected by Agilent
LSTM 0.027935 3.581093 0.501154 34901A.
CNN-LSTM 0.013923 1.884062 0.876073 The batteries go through 4 steps in each cycle: 1) Discharging CC at
CL- 0.010859 1.331990 0.924613 58A until the voltage drop to 2.5 V. 2) The battery is rest for 1800 s. 3)
GraphSAGE
Charging CC at 38.86 A until the voltage increases to 4.25 V, then CV at
B0007 RNN 81 0.045730 5.199274 − 0.879691
LSTM 0.041061 4.671529 − 0.515507 4.25 V until the current drop to 2.9A. 4) Resting 1800 s.
CNN-LSTM 0.031836 3.398467 0.088939 Five strongly correlated HIs (B3, B5, B8, B17, B26) are taken as
CL- 0.019028 2.383225 0.674550 model input from the charge-discharge curves as CL-GraphSAGE input
GraphSAGE by using HIs extracted in Section 2.2. The predicted results are shown in
RNN 101 0.025090 2.852394 − 0.096398
LSTM 0.015771 1.766805 0.566800
Fig. 10. Fig. 10 shows that the CL-GraphSAGE well captures the degra­
CNN-LSTM 0.008994 0.956124 0.859108 dation of LIBs. The RMSE and R2 are 0.003688 and 0.986188, respec­
CL- 0.005314 0.606468 0.950810 tively. Both RMSE and R2 verify the effectiveness of CL-GraphSAGE.
GraphSAGE

4.4. Experiment of different graph models

connected when the similarity value is more than 0.95. All the graph
structures are fully connected graphs. Fig. 8(d) shows the graph struc­
tures of the battery B0005 HIs.
Compared with the NASA dataset, the MIT dataset has more cycles
and a larger amount of data. Different training method is employed in
the MIT dataset. A0001 is used for training, and A0002 and A0003 are
used for testing.
Fig. 9 shows the prediction results and errors. Table 6 shows the
performance evaluation of the four models. Compared with the other
three models, CL-GraphSAGE obtains the best prediction performance,
Instead of using GraphSAGE in the NASA dataset, combined with
CNN and LSTM, the graph convolutional network (GCN) and graph
attention network (GAT) which are CL-GCN and CL-GAT used to
compare with the CL-GraphSAGE. The results are shown in Fig. 11 and
Table 7. The results are similar to that of CL-GAT and CL-GraphSAGE.
However, the running time of CL-GCN, CL-GAT, and CL-GraphSAGE
are 169.21 s, 462.63 s, and 110.38 s, respectively. The execution time
of CL-GAT is more than four times that of CL-GraphSAGE. In terms of
time cost and prediction accuracy, CL-GraphSAGE is the best of the three
GCNs proposed in this paper to obtain the spatial features of HIs.

Fig. 7. Prediction results at starting point of 101: (a) B0005; (b) B0006; (c) B0007.

8
X.-Y. Yao et al. Journal of Energy Storage 58 (2023) 106437

Table 5
Descending order of correlation coefficient for HIs.
Sample Sort in descending order (from left to right, top to bottom)

A0001 B17 0.999098 B19 0.998555 B6 0.998443 B1 0.998323

B5 0.997899 B8 0.996969 B13 0.978474 B22 0.975399
B20 0.749903 B7 0.748421 B18 0.726796 B2 0.508062
B10 0.398186 B12 0.192002 B11 0.140236
A0002 B19 0.998027 B6 0.997991 B1 0.997727 B17 0.996563
B5 0.995133 B8 0.992152 B13 0.970764 B22 0.969931
B12 0.733495 B20 0.728244 B7 0.727605 B10 0.695812
B18 0.690713 B11 0.646046 B2 0.250485
A0003 B19 0.998105 B6 0.998042 B1 0.997935 B17 0.996740
B5 0.995852 B8 0.993183 B13 0.970857 B22 0.963337
B18 0.677841 B20 0.611967 B7 0.611051 B10 0.309357
B12 0.171201 B11 0.038053 B2 0.036999

The bold in the table are the selected HIs which is the input of the model in this case.

Fig. 8. Node similarity and graph structure: (a) A0001; (b) A0002; (c) A0003; (d) graph of A0001 HIs.

Fig. 9. SOH prediction results: (a) A0002; (b) A0003.

9
X.-Y. Yao et al. Journal of Energy Storage 58 (2023) 106437

Table 6 compared with RNN, LSTM, and CNN-LSTM.

Performance evaluation with different models. It is understood that there is little research on GNN in LIBs. This work
Sample Model RMSE MAPE R2 opens a new door for using GNNs to process various structural or un­
structured data of batteries. Naturally, there are some limitations to the
A0002 RNN 0.003048 0.264716 0.993135
LSTM 0.002468 0.215389 0.995500 method. The future work includes: (1) The correlated HIs extracted
CNN-LSTM 0.003383 0.266529 0.991545 based on partial charge-discharge curves will be developed in the future.
CL-GraphSAGE 0.002070 0.197313 0.996834 (2) The edges connected are fixed weight by calculating the Pearson
A0003 RNN 0.004226 0.353931 0.987291 correlation coefficient in the proposed approach, which are all fully
LSTM 0.002776 0.266655 0.994514
CNN-LSTM 0.003361 0.304047 0.991960
connected in the graph. Hence, the weight of the edges can be designed
CL-GraphSAGE 0.002114 0.184998 0.996818 to update automatically to determine which edge to connect adaptively.
(3) GNN can also be applied to other aspects of batteries, such as
remaining life prediction, thermal runaway warning, and fault
diagnosis.

CRediT authorship contribution statement

Xing-Yan Yao: Conception, Methodology, Validation, Formal anal­

ysis, Investigation, Writing - Original Draft, Writing - Review & Editing,
Visualization, Guolin Chen: Conceptualization, Methodology, Software,
Validation, Writing, Bin Chen: providing dataset, Michael Pecht: Re­
view, Supervision.

Declaration of competing interest

The authors declare that they have no known competing financial

interests or personal relationships that could have appeared to influence
the work reported in this paper.

Fig. 10. Prediction results of CL-GraphSAGE. Data availability

Data will be made available on request.

Acknowledgments

This work is partially supported by the National Natural Science

Foundation of China (51605061); Science and Technology Research
Project of Chongqing Municipal Education Commission
(KJQN201900808); Natural Science Foundation of Chongqing, China
(cstc2020jcyj-msxmX0736); Innovation and Entrepreneurship Demon­
stration Team of Chongqing Talent Plan (CQYC201903246); Startup
Project of Doctor Scientific Research (2016-56-04) and Open Grant of
Chongqing Engineering Laboratory for Detection Control and Integrated
System (KFJJ2021018). The valuable comments and suggestions from
Fig. 11. Results of the three GNN models. the anonymous reviewer are appreciated. Thanks for NASA and MIT
provide the battery datasets.
Table 7
Errors of the three GNN models. References
Method RMSE MAPE R2 [1] X. Hu, F. Feng, K. Liu, et al., State estimation for advanced battery management:
CL-GCN 0.020253 2.504092 0.493080 key challenges and future trends[J], Renew. Sust. Energ. Rev. 114 (2019), 109334.
CL-GAT 0.006454 0.505244 0.948513 [2] P.M. Attia, A. Grover, N. Jin, et al., Closed-loop optimization of fast-charging
protocols for batteries with machine learning[J], Nature 578 (7795) (2020)
CL-GraphSAGE 0.005808 0.625176 0.958309
397–402.
[3] X. Tang, C. Zou, K. Yao, et al., A fast estimation algorithm for lithium-ion battery
state of health[J], J. Power Sources 396 (2018) 453–458.
5. Conclusions [4] H. Tian, P. Qin, K. Li, et al., A review of the state of health for lithium-ion batteries:
research status and suggestions[J], J. Clean. Prod. 261 (2020), 120813.
[5] R. Xiong, L. Li, J. Tian, Towards a smarter battery management system: a critical
The current lithium-ion battery SOH predicted methods by health
review on battery state of health monitoring methods[J], J. Power Sources 405
indicators only take time features into account and ignore the rela­ (2018) 18–29.
tionship between different features especially for the limited data in the [6] J. Li, K. Adewuyi, N. Lotfi, et al., A single particle model with chemical/mechanical
application. In this paper, a novel GNN-based framework is proposed for degradation physics for lithium ion battery state of health (SOH) estimation[J],
Appl. Energy 212 (2018) 1178–1190.
SOH prediction, which considers both the temporal features and the [7] G.K. Prasad, C.D. Rahn, Model based identification of aging parameters in lithium
spatial features of HIs for the limited data. The validation and feasibility ion batteries[J], J. Power Sources 232 (2013) 79–85.
of the proposed CL-GraphSAGE are proved by public battery datasets [8] X. Hu, H. Yuan, C. Zou, et al., Co-estimation of state of charge and state of health
for lithium-ion batteries based on fractional-order calculus[J], IEEE Trans. Veh.
and laboratory experiments. The results show that the CL-GraphSAGE Technol. 67 (11) (2018) 10319–10329.
reached minimum error and predicts more accurately and precisely [9] Z. Guo, X. Qiu, G. Hou, et al., State of health estimation for lithium ion batteries
based on charging curves[J], J. Power Sources 249 (2014) 457–462.

10
X.-Y. Yao et al.
[10] J. Kim, B.H. Cho, State-of-charge estimation and state-of-health prediction of a li-
ion degraded battery based on an EKF combined with a per-unit system[J], IEEE
Trans. Veh. Technol. 60 (9) (2011) 4249–4260.
[11] J. Yang, B. Xia, W. Huang, et al., Online state-of-health estimation for lithium-ion
batteries using constant-voltage charging current analysis[J], Appl. Energy 212
(2018) 1589–1600.
[12] X. Feng, J. Li, M. Ouyang, et al., Using probability density function to evaluate the
state of health of lithium-ion batteries[J], J. Power Sources 232 (2013) 209–218.
[13] A. Guha, A. Patra, State of health estimation of lithium-ion batteries using capacity
fade and internal resistance growth models[J], IEEE Trans. Transp. Electrification
4 (1) (2017) 135–146.
[14] P. Singh, C. Chen, C.M. Tan, et al., Semi-empirical capacity fading model for SoH
estimation of li-ion batteries[J], Appl. Sci. 9 (15) (2019) 3012.
[15] Z. Deng, X. Hu, P. Li, et al., Data-driven battery state of health estimation based on
random partial charging data[J], IEEE Trans. Power Electron. 37 (5) (2021)
5021–5031.
[16] D. Yang, X. Zhang, R. Pan, et al., A novel gaussian process regression model for
state-of-health estimation of lithium-ion battery using charging curve[J], J. Power
Sources 384 (2018) 387–395.
[17] S. Song, C. Fei, H. Xia, Lithium-ion battery SOH estimation based on XGBoost
algorithm with accuracy correction[J], Energies 13 (4) (2020) 812.
[18] Y. Yang, J. Wen, Y. Shi, et al., State of health prediction of lithium-ion batteries
based on the discharge voltage and temperature[J], Electronics 10 (12) (2021)
1497.
[19] K.S.R. Mawonou, A. Eddahech, D. Dumur, et al., State-of-health estimators coupled
to a random forest approach for lithium-ion battery aging factor ranking[J],
J. Power Sources 484 (2021), 229154.
[20] R. Xiong, J. Cao, Q. Yu, Reinforcement learning-based real-time power
management for hybrid energy storage system in the plug-in hybrid electric vehicle
[J], Appl. Energy 211 (2018) 538–548.
[21] R. Xiong, Y. Zhang, J. Wang, et al., Lithium-ion battery health prognosis based on a
real battery management system used in electric vehicles[J], IEEE Trans. Veh.
Technol. 68 (5) (2018) 4110–4121.
[22] Z. Yun, W. Qin, W. Shi, et al., State-of-health prediction for lithium-ion batteries
based on a novel hybrid approach[J], Energies 13 (18) (2020) 4858.
[23] P. Venugopal, State-of-health estimation of li-ion batteries in electric vehicle using
IndRNN under variable load condition[J], Energies 12 (22) (2019) 4338.
[24] Y. Fan, F. Xiao, C. Li, et al., A novel deep learning framework for state of health
estimation of lithium-ion battery[J], J. Energy Storage 32 (2020), 101741.
[25] Z. Deng, X. Lin, J. Cai, et al., Battery health estimation with degradation pattern
recognition and transfer learning[J], J. Power Sources 525 (2022), 231027.
[26] S. Pepe, J. Liu, E. Quattrocchi, et al., Neural ordinary differential equations and
recurrent neural networks for predicting the state of health of batteries[J],
J. Energy Storage 50 (2022), 104209.
[27] S.J. Kim, S.H. Kim, H.M. Lee, State of health estimation of Li-ion batteries using
multi-input LSTM with optimal sequence length[C], in: 2020 IEEE 29th
International Symposium on Industrial Electronics (ISIE), IEEE, 2020,
pp. 1336–1341.
[28] S. Cui, I. Joe, A dynamic spatial-temporal attention-based GRU model with healthy
features for state-of-health estimation of lithium-ion batteries[J], IEEE Access 9
(2021) 27374–27388.
11
Journal of Energy Storage 58 (2023) 106437
[29] Z. Wu, S. Pan, G. Long, Connecting the dots: multivariate time series forecasting
with graph neural networks[C], in: Proceedings of the 26th ACM SIGKDD
International Conference on Knowledge Discovery & Data Mining, 2020,
pp. 753–763.
[30] H. Xu, Y. Huang, Z. Duan, Multivariate Time Series Forecasting With Transfer
Entropy Graph[J], arXiv preprint arXiv:2005.01185, 2020.
[31] Y. Wang, Z. Duan, Y. Huang, et al., MTHetGNN: a heterogeneous graph embedding
framework for multivariate time series forecasting[J], Pattern Recogn. Lett. 153
(2022) 151–158.
[32] B. Yu, H. Yin, Z. Zhu, Spatio-temporal Graph Convolutional Networks: A Deep
Learning Framework for Traffic Forecasting[J], arXiv preprint arXiv:1709.04875,
2017.
[33] M. Khodayar, J. Wang, Spatio-temporal graph deep neural network for short-term
wind speed forecasting[J], IEEE Trans. Sustain. Energy 10 (2) (2018) 670–681.
[34] A. Vyas, S. Bandyopadhyay, Semi-supervised Soil Moisture Prediction Through
Graph Neural Networks[J], arXiv preprint arXiv:2012.03506, 2020.
[35] M. Wang, Y. Li, Y. Zhang, et al., Spatio-temporal graph convolutional neural
network for remaining useful life estimation of aircraft engines[J], Aerosp. Syst. 4
(1) (2021) 29–36.
[36] B. Zraibi, C. Okar, H. Chaoui, et al., Remaining useful life assessment for lithium-
ion batteries using CNN-LSTM-DNN hybrid method[J], IEEE Trans. Veh. Technol.
70 (5) (2021) 4252–4261.
[37] X. Song, F. Yang, D. Wang, et al., Combined CNN-LSTM network for state-of-charge
estimation of lithium-ion batteries[J], IEEE Access 7 (2019) 88894–88902.
[38] G. Zou, Z. Yan, C. Zhang, Transfer learning with CNN-LSTM model for capacity
prediction of lithium-ion batteries under small sample[C], in: Journal of Physics:
Conference Series 2258(1), IOP Publishing, 2022, p. 012042.
[39] B. Sahaand, K. Goebel, Battery Data Set, NASA Ames Prognostics Data Repository
[J], NASA Ames Research Center, 2007.
[40] K.A. Severson, P.M. Attia, N. Jin, et al., Data-driven prediction of battery cycle life
before capacity degradation[J], Nat. Energy 4 (5) (2019) 383–391.
[41] Z. Fei, F. Yang, K.L. Tsui, et al., Early prediction of battery lifetime via a machine
learning based framework[J], Energy 225 (2021), 120205.
[42] P. Guo, Z. Cheng, L. Yang, A data-driven remaining capacity estimation approach
for lithium-ion batteries based on charging health feature extraction[J], J. Power
Sources 412 (2019) 442–450.
[43] Y. Li, D.I. Stroe, Y. Cheng, et al., On the feature selection for battery state of health
estimation based on charging–discharging profiles[J], J. Energy Storage 33 (2021),
102122.
[44] G. Chandrashekar, F. Sahin, A survey on feature selection methods[J], Comput.
Electr. Eng. 40 (1) (2014) 16–28.
[45] L. Qiao, L. Zhang, S. Chen, et al., Data-driven graph construction and graph
learning: a review[J], Neurocomputing 312 (2018) 336–351.
[46] J. Niu, C. Liu, L. Zhang, Remaining useful life prediction of machining tools by 1D-
CNN LSTM network[C], in: 2019 IEEE Symposium Series on Computational
Intelligence (SSCI), IEEE, 2019, pp. 1056–1063.
[47] T.Y. Kim, S.B. Cho, Predicting residential energy consumption using CNN-LSTM
neural networks[J], Energy 182 (2019) 72–81.
[48] J. Gilmer, S.S. Schoenholz, P.F. Riley, Neural message passing for quantum
chemistry[C], in: International Conference on Machine Learning. PMLR, 2017,
pp. 1263–1272.
Another random document with
no related content on Scribd:
 “It’s more’n Ah be, though,” said he dubiously. “Ah found him
standing over t’ corpse loike this ’ere, wi’ this in his hand.” He
produced the revolver from his pocket. “And in t’ other hand he
gotten letter ye spoake of, lass, that ye said would enreage him.”
“And what did he say? Did you accuse him?”
“He said he didn’t do it, an’ Ah, why, Ah didn’t believe him.”
“But I do,” said Freda calmly.
“Weel, but who could ha’ done it then?” asked he, hoping that she
might have a reason to give which would bring satisfaction to his
mind also.
But in Freda’s education faith and authority had been put before
reason, and her answer was not one which could carry conviction to
a masculine understanding.
“My father,” she said solemnly, “could not commit a murder.”
“Weel, soom folks’ feythers does, why not your feyther? There was
nobody else to do it, an’ t’ poor feller couldn’t ha’ done it hissen, for
he was shot in t’ back.”
“I will never believe my father did it,” said Freda.
“Happen he’ll tell ye he did.”
Freda shook her head.
“I have been very foolish,” she said at last, “to listen to all the
things I have heard said against him. And perhaps it is as a
punishment to me that I have heard this. He was good and kind
when I was a baby: how can he be bad now? And if he has done bad
things since then, the Holy Spirit will come down into his heart again
now, if I pray for him.”
“Amen,” said Barnabas solemnly.
This farmer had no more definite religion himself than that there
was a Great Being somewhere, a long way off behind the clouds,
whom it was no use railing at, though he didn’t encourage honest
industry as much as he might, and whom it was the parson’s duty to
keep in good humour by baptisms, and sermons, and ringing of the
church-bells. But he had, nevertheless, a belief in the more lively
religion of women, and thought—always in a vague way—that it
brought good luck upon the world. So he took off his hat reverently
when the girl was giving utterance to her simple belief, and then he
led the horse past the dead body, and jumping up into the cart
beside her, took up the reins.
 CHAPTER VI.
After a little more jolting along the highroad they turned to the right
up one less used, and soon came in full sight of the Abbey ruins.
Just a jagged dark grey mass they looked by the murky light of this
dull evening, with here and there a jutting point upwards, the outline
of the broken walls softened by the snow.
Freda sat quite silent, awestruck by the circumstances of her
arrival, and by the wild loneliness of the place. A little further, and
they could see the grey sea and the high cliffs frowning above it.
Barnabas glanced down at the grave little face, and made an effort
to say something cheering.
“It bean’t all so loansome as what this is, ye know. Theer’s t’ town
o’ t’other soide o’ t’ Abbey, at bottom of t’ hill. And from t’ windows o’
Capt’n Mulgrave’s home ye can see roight oop t’ river, as pretty a
soight as can be, wi’ boats a-building, an’ red cottages.”
“Oh!” said Freda, in a very peaceful voice, “I don’t mind the
loneliness. I like it best. And I have always lived by the sea, where
you could hear the waves till you went to sleep.”
“Aye, an’ you’ll hear ’em here sometimes; fit to split t’ owd cliffs
oop they cooms crashing in, an’ soonding like thoonder. Ye’ll have a
foin toime here, lass, if ye’re fond of t’ soond o’ t’ weaves.”
“My father has a yacht too, hasn’t he?”
“Aye, an a pretty seeght too, to see it scoodding along. But it goes
by steam, it isn’t one of yer white booterflies. That sort doan’t go fast
enough for t’ Capt’n.”
Freda was no longer listening. They were on the level ground now
at the top of the hill. To the right, the fields ran to the edge of the cliff,
and there was no building in sight but a poor sort of farm-house, with
a pond in front of it, and a few rather dilapidated outhouses round
about. But on the left hand hedged off from the road by a high stone
wall, and standing in the middle of a field, was the ruined Abbey
church, now near enough for Freda to see the tracery left in the
windows, and the still perfect turrets of the East end, and of the
North transept pointing to heaven, unmindful of the decay of the old
altars, and of the old faith that raised them.
Barnabas looked at her intent young face, the great burning eyes,
which seemed to be overwhelmed with a strange sorrow.
“Pretty pleace, this owd abbey of ours, isn’t it?” said he with all the
pride of ownership.
“It’s beautiful,” said Freda hoarsely, “it makes me want to cry.”
Now the rough farmer could understand sentiment about the old
ruin; considering as he did that the many generations of Protestant
excursionists who had picknicked in it had purged it pretty clear of
the curse of popery, he loved it himself with a free conscience.
“Aye,” said he, “there’s teales aboot it too, for them as loikes to
believe ’em. Ah’ve heard as there were another Abbey here, afore
this one, an’ not near so fine, wheer there was a leady, an Abbess,
Ah think they called her. An’ she was a good leady, kind to t’ poor,
an’ not so much to be bleamed for being a Papist, seeing those were
dreadful toimes when there was no Protestants. An’ they do seay
(mahnd, Ah’m not seaying Ah believe it, not being inclined to them
soart o’ superstitious notions myself) they say how on an afternoon
when t’ soon shines you can see this Saint Hilda, as they call her,
standing in one of t’ windows over wheer t’ Communion table used
for to be.” Perceiving, however, that Freda was looking more
reverently interested than was quite seemly in a mere legend with a
somewhat unorthodox flavour about it, Barnabas, who was going to
tell her some more stories of the same sort, changed his mind and
ended simply: “An’ theer’s lots more sooch silly feables which
sensible fowk doan’t trouble their heads with. Whoa then, Prince!”
The cart drew up suddenly in a sort of inclosure of stone walls. To
the right was an ancient and broken stone cross, on a circular flight
of rude and worn steps; to the left, a stone-built lodge, a pseudo-
Tudor but modern erection, was built over a gateway, the wrought-
iron gates of which were shut. In front, a turnstile led into a
churchyard. Barnabas got down and pulled the lodge-bell, which
gave a startingly loud peal.
“That yonder,” said he, pointing over the wall towards the
churchyard, in which Freda could dimly see a shapeless mass of
building and a squat, battlemented tower, “is Presterby Choorch. An’
this,” he continued, as an old woman came out of the lodge and
unlocked the gate, “is owd Mary Sarbutt, an she’s as deaf as a
poast. Now, hark ye, missie,” and he held out his hand to help Freda
down, “Ah can’t go no further with ye, but ye’re all reeght now. Joost
go oop along t’ wall to t’ left, streight till ye coom to t’ house, an’ pull
t’ bell o’ t’ gate an’ Mrs. Bean, or happen Crispin himself will coom
an’ open to ye.”
The fact was that Barnabas did not for a moment entertain the
idea that Captain Mulgrave would have the heart to leave his newly-
recovered daughter, and the farmer meant to come up to the Abbey-
house in a day or two and let him know quietly that he had nothing to
fear from him as long as he proved a good father to the little lass.
But just now Barnabas felt shy of showing himself again, and he
shook his head when Freda begged him to come a little way further
with her. For a glance through the gates at the house showed her
such a bare, gaunt, cheerless building that she began to feel
frightened and miserable.
“Noa, missie, Ah woan’t coom in,” said Barnabas, who seemed to
have grown both shyer and more deferential when he had landed the
young lady at the gates of the big house; “but Ah wish ye ivery
happiness, an’ if Ah meay mak’ so bawld, Ah’ll shak’ honds wi’ ye,
and seay good-bye.”
Freda with the tears coming, wrung his hand in both hers, and
watched him through the gates while he turned the horse, got up in
his place in the cart and drove away.
“Barnabas! Barnabas!” she cried aloud.
But the gates were locked, and the old woman, without one word
of question or of direction, had gone back into the lodge. Freda
turned, blinded with tears, and began to make her way slowly
towards the house.
Nothing could be more desolate, more bare, more dreary, than the
approach. An oblong, rectangular space, shut in by high stone walls,
and without a single shrub or tree, lay between her and the building.
Half-way down, to the right, a pillared gateway led to the stables,
which were very long and low, and roofed with red tiles. This bit of
colour, however, was now hidden by the snow, which lay also, in a
smooth sheet, over the whole inclosure. Freda kept close to the left-
hand wall, as she had been told to do, her heart sinking within her at
every step.
At last, when she had come very near to the façade of the house,
which filled the bottom of the inclosure from end to end, a cry burst
from her lips. It was shut up, unused, deserted, and roofless. What
had once been the front-door, with its classic arch over the top, was
now filled up with boards strengthened by bars of iron. The rows of
formal, stately Jacobian windows were boarded up, and seemed to
turn her sick with a sense of hideous deformity, like eye-sockets
without eyes. The sound of her voice startled a great bird which had
found shelter in the moss-grown embrasure of one of the windows.
Flapping its wings it flew out and wheeled in the air above her.
Shocked, chilled, bewildered, Freda crept back along the front of
the house, feeling the walls, from which the mouldy stucco fell in
flakes at her touch, and listening vainly for some sound of life to
guide her.
 CHAPTER VII.
Freda Mulgrave was superstitious. While she was groping her way
along the front of the dismantled house, she heard a bell tolling
fitfully and faintly, and the sound seemed to come from the sea. She
flew instantly to the fantastic conclusion that Saint Hilda, in heaven,
was ringing the bell of her old church on earth to comfort her in her
sore trouble.
“Saint Hilda was always good to wanderers,” she thought. And the
next moment her heart sprang up with a great leap of joy, for her
hand, feeling every excrescence along the wall, had at last touched
the long swinging handle of a rusty bell.
Freda pulled it, and there was a hoarse clang. She heard a man’s
footsteps upon a flagged court-yard, and a rough masculine voice
asked:
“Who’s that at this time of night?”
“It is Captain Mulgrave’s daughter. And oh! take me in; I am tired,
tired.”
The gate was unbolted and one side was opened, enabling the girl
to pass in. The man closed the gate, and lifted a lantern he carried
so as to throw the light on Freda’s face.
“So you’re the Captain’s daughter, you say?”
“Yes.”
Freda looked at him, with tender eyes full of anxiety and inquiry.
He was a tall and rather thickset man with very short greyish hair
and a little unshaved stubble on his chin. Her face fell.
“I thought——” she faltered.
“Thought what, miss?”
There was a pause. Then she asked:
“Who are you?”
She uttered the words slowly, under her breath.
“I am your servant, ma’am.”
“My servant—you mean my father’s?”
“It is the same thing, is it not?”
 “Oh, then you are Crispin Bean!”
The man seemed surprised.
“How did you know my name?”
“They told me about you.”
“What did they tell you?”
“That you were a ‘rough-looking customer.’ ”
The man laughed a short, grim laugh, which showed no
amusement.
“Well, yes; I suppose they were about right. But who were ‘they’?”
“The people at Oldcastle Farm.”
The man stopped short just as, after leading her along a wide,
stone-paved entry, between the outer wall and the side of the house,
he turned into a large square court-yard.
“Oh!” he said, and lifting his lantern again, he subjected the young
lady to a second close scrutiny. “So you’ve been making friends with
those vermin.”
Freda did not answer for a moment. Presently she said, in a stifled
voice:
“I am not able to choose my friends.”
“You mean that you haven’t got any? Poor creature, poor creature,
that’s not far from the truth, I suppose. That father of yours didn’t
treat you over well, or consider you over much, did he?”
Freda grew cold, and her crutch rattled on the stones.
“What do you mean? ‘Didn’t treat me well’!” she whispered. “He
will, I am sure he will, when he sees me, knows me.”
“Oh, no, you’re mistaken. He’s dead.”
Freda did not utter a sound, did not move. She remained
transfixed, benumbed, stupefied by the awful intelligence.
“It isn’t true! It can’t be true!” she whispered at last, with dry lips.
“Barnabas saw him to-day—just now.”
“He was alive two hours ago. He went out this afternoon, came in
in a great state of excitement and went up to his room. Presently I
heard a report, burst open the door, and found him dead—shot
through the head.”
“Dead!” repeated Freda hoarsely.
She could not believe it. All the dreams, which she had cherished
up to the last moment in spite of disappointments and disillusions, of
a tender and loving father whom her affection and dutiful obedience
should reconcile to a world which had treated him harshly, were in a
moment dashed to the ground.
“Dead!”
It was the knell of all her hopes, all her girlish happiness. Forlorn,
friendless, utterly alone, she was stranded upon this unknown corner
of the world, in a cheerless house, with no one to offer her even the
comfort of a kindly pressure of the hand. The man seemed sorry for
her. He stamped on the ground impatiently, as if her grief distressed
and annoyed him.
“Come, come,” he said. “You haven’t lost much in losing him. I
know all about it; he never went to see you all these years, and didn’t
care a jot whether you lived or died, as far as any one could see.
And it’s all nonsense to pretend you’re sorry, you know. How can you
be sorry for a father you don’t remember?”
“Oh,” said Freda, with a sob, “can’t you understand? You can love
a person without knowing him, just as we love God, whom we can
never see till we die.”
“Well, but I suppose you love God, because you think He’s good to
you.”
“We believe He is, even when He allows things to happen to us
which seem cruel. And my father being, as I am afraid he was, an
unhappy man, was perhaps afraid of making me unhappy too. And
he did send for me at last, remember.”
“Yes, in a fit of annoyance over something—I forget what.”
“How do you know that he hadn’t really some other motive in his
heart?” said Freda, down whose cheeks the tears were fast rolling.
“He was a stern man, everybody says, who didn’t show his feelings.
So that at last he grew perhaps ashamed to show them.”
“More likely hadn’t got any worth speaking of,” said the man
gruffly.
“It’s not very nice or right of you to speak ill of your master, when
he’s de-ad,” quavered Freda.
“Well, it’s very silly of you to make such a fuss about him when
he’s de-ad,” mimicked the man.
Although he spoke without much feeling of his late master, and
although he was somewhat uncouth of speech, manner and
appearance, Freda did not dislike this man. As might have been
expected, she confounded bluntness with honesty in the
conventional manner. Therefore she bore even his little jibes without
offence. There was a pause, however, after his last words. Then he
asked, rather curiously:
“Come, honestly, what is your reason for taking his part through
thick and thin like this? Come,” he repeated, getting for the moment
no answer, “what is it?”
Freda hesitated, drying her eyes furtively.
“Don’t you see,” she said, tremulously, “that it is my only
consolation now to think the very, very best of him?”
The man, instead of answering, turned from her abruptly, and
signed to her with his hand to follow him. This she did; and they
passed round one side of the court-yard under a gallery, supported
by a colonnade, and entering the house, went through a wide, low
hall, into an apartment to the right at the front of the building. It was a
pretty room, with a low ceiling handsomely moulded, panelled walls,
and an elaborately carved wooden mantelpiece, which had been a
good deal knocked about. The room had been furnished with solid
comfort, if without much regard to congruity, a generation or so back;
and the mahogany arm-chairs having been since shrouded in
voluminous chintz covers with a pattern of large flowers on a dark
ground, the room looked warm and cheerful. Tea was laid on the
table for two persons. Freda’s sharp eyes noted this circumstance at
once. She turned round quickly.
“Who is this tea for?” she asked.
“Captain Mulgrave’s death was not discovered until it was ready.”
“But it was laid for two. Was it for you also?”
“Yes.”
Freda’s face fell.
“You think it was derogatory to his dignity to have his meals with
me?”
“Oh, no, no indeed,” said Freda blushing. “I knew at once, when
you said you were a servant, that it was only a way of speaking. You
were an officer on board his ship, of course?”
“Yes,” said he.
 “But I had hoped,” said Freda wilfully, “that he had expected me,
and had tea made ready for me and him together.”
“Ah!” said the man shortly. “Sit down,” he went on, pointing
brusquely to a chair without looking at her, “I’ll send Mrs. Bean to
you; she must find a room for you somewhere, I suppose.”
“For to-night, yes, if you please. Mrs. Bean—that is your wife?”
He nodded and went out, shutting the door.
Freda heard him calling loudly “Nell, Nell!” in a harsh, authoritative
voice, as he went down the passage.
She thought she should be glad to be alone, to have an
opportunity to think. But she could not. The series of exciting
adventures through which she had passed since she left the quiet
convent life had benumbed her, so that this awful discovery of her
father’s sudden death, though it agitated her did not impress her with
any sense of reality. When she tried to picture him lying dead
upstairs, she failed altogether; she must see him by-and-by, kiss his
cold face; and then she thought that she would be better able to pray
that she might meet him in heaven.
It seemed to her that she had been left alone for hours when a
bright young woman’s voice, speaking rather querulously, reached
her ears. Freda guessed, before she saw Mrs. Bean, that her
father’s fellow-officer or servant (she was uncertain what to call him)
had married beneath him. However, when the door opened, it
revealed, if not a lady of the highest refinement, a very pleasant-
looking, plump little woman, with fair hair and bright eyes, who wore
a large apron but no cap, and who looked altogether like an
important member of the household, accustomed to have her own
way unquestioned.
“Dear me, and is that the little lady?” she asked, in a kind,
motherly voice, encircling the girl with a rounded arm of matronly
protection. “Bless her poor little heart, she looks half-perished.
Crispin,” she went on, in a distant tone, which seemed to betray that
she and her husband had been indulging in a little discussion, “go
and put the kettle on while I take the young lady upstairs. Come
along, my dear. I’ll get you some hot water and some dry clothes,
and in two-twos I’ll have you as cosy as can be.”
 Mrs. Bean looked a little worried, but she was evidently not the
woman to take to heart such a trifle as a suicide in the house, as
long as things went all right in the kitchen, and none of the chimneys
smoked. Crispin, who seemed to have little trust in her discretion,
gave her arm a rough shake of warning as she left the room with the
young lady. Mrs. Bean, therefore, kept silence until she and her
charge got upstairs. Then she popped her head over the banisters to
see that Crispin was out of hearing, and proceeded to unbend in
conversation, being evidently delighted to have somebody fresh to
speak to.
 CHAPTER VIII.
“Oh,” began Mrs. Bean, with a fat and comfortable sigh, “I am glad to
have you here, I declare. Ever since the Captain told me, in his short
way, that you were coming, I’ve been that anxious to see you, you
might have been my own sister.”
“That was very good of you,” said Freda, who was busily taking in
all the details of the house, the wide, shallow stairs, low ceilings, and
oaken panelling; the air of neglect which hung about it all; the
draughts which made her shiver in the corridors and passages. She
compared it with the farm-house she had just left, so much less
handsome, so much more comfortable. How wide these passages
were! The landing at the top of the staircase was like a room, with a
long mullioned window and a wide window-seat. But it was all bare,
cold, smelling of mould and dust.
“Isn’t this part of the house lived in?” asked Freda.
“Well, yes and no. The Captain lives in it—at least did live in it,”
she corrected, lowering her voice and with a hasty glance around.
“No one else. This house would hold thirty people, easy, so that
three don’t fill it very well.”
“But doesn’t it take a lot of work to keep it clean?”
“It never is kept clean. What’s the good of sweeping it up for the
rats?” asked Mrs. Bean comfortably. “I and a girl who comes in to
help just keep our own part clean and the two rooms the Captain
uses, and the rest has to go. If the Captain had minded dust he’d
have had to keep servants; I don’t consider myself a servant, you
know,” she continued with a laugh, “and I’m not going to slave myself
to a skeleton for people that save a sixpence where they might
spend a pound.”
It would have taken a lot of slaving to make a skeleton of Mrs.
Bean, Freda thought.
They passed round the head of the staircase and into a long
gallery which overlooked the court-yard. It was panelled and hung
with dark and dingy portraits in frames which had once been gilt.
 “Does any one live in this part?” asked Freda, shivering.
Mrs. Bean’s candle threw alarming shadows on the walls. The
mullioned window, which ran from end to end of the gallery, showed
a dreary outlook of dark walls surrounding a stretch of snow.
“Well, no,” admitted her guide reluctantly. “The fact is there isn’t
another room in the house that’s fit to put anybody into; they’ve been
unused so long that they’re reeking with damp, most of them; some
of the windows are broken. And so I thought I’d put you into the
Abbot’s room. It’s a long way from the rest of us, but it’s had a fire in
it once or twice lately, when the Captain has had young Mulgrave
here. It’s a bit gloomy looking and old fashioned, but you mustn’t
mind that.”
Freda shivered again. If the room she was to have was more
gloomy than the way to it, a mausoleum would be quite as cheerful.
“The Abbot’s room!” exclaimed Freda. “Why is it called that?”
“Why, this house wasn’t all built at the same time, you know.
There’s a big stone piece at this end that was built earliest of all. It’s
very solid and strong, and they say it was the Abbot’s house. Then in
Henry the Eighth’s time it was turned into a gentleman’s house, in
what they call the Tudor style. They built two new wings, and carried
the gallery all round the three sides. A hundred and fifty years later a
banqueting room was built, making the last side of the square; but it
was burnt down, and now there’s nothing left of it but the outside
walls of the front and sides.”
This explained to Freda the desolate appearance the house had
presented as she approached it. The deep interest she felt in this,
the second venerable house she had been in since her arrival in
England, began to get the better of her alarm at its gloominess. But
at the angle of the house, where the gallery turned sharply to the
right, Mrs. Bean unlocked a door, and introduced her to a narrow
stone passage which was like a charnel-house.
“This,” said Mrs. Bean with some enthusiasm, “is the very oldest
part; and I warrant you’ll not find such another bit of masonry, still
habitable, mind, in any other house in England!”
Was it habitable? Freda doubted it. The walls of the passage were
of great blocks of rough stone. It was so narrow that the two women
could scarcely walk abreast. They passed under a pointed arch of
rough-hewn stone, and came presently to the end of the passage,
where a narrow window, deeply splayed, threw a little line of murky
light on to the boards of the floor. On the right was a low and narrow
Gothic doorway, with the door in perfect preservation. Mrs. Bean
opened it by drawing back a rusty bolt, and ushered Freda, with
great pride, into a room which seemed fragrant with the memories of
a bygone age. Freda looked round almost in terror. Surely the Abbot
must still be lurking about, and would start out presently, in dignified
black habit, cowl and sandals, and haughtily demand the reason of
her intrusion! For here was the very wide fireplace, reaching four feet
from the ground, and without any mantelshelf, where fires had
burned for holy Abbot or episcopal guest four hundred years ago.
Here were the narrow windows deeply splayed like the one outside
from which the prosperous monks had looked out over their wide
pasture-lands and well-stocked coverts.
Even in the furniture there was little that was incongruous with the
building. The roughly plastered walls were hung with tapestry much
less carefully patched and mended than the hangings at Oldcastle
Farm. The floor was covered by an old carpet of harmoniously
undistinguishable pattern. The rough but solid chairs of unpolished
wood, with worn leather seats; the ancient press, long and low,
which served at one end as a washhand-stand, and at the other as a
dressing-table; a large writing-table, which might have stood in the
scriptorium of the Abbey itself, above all, the enormous four-poster
bedstead, with faded tapestry to match the walls, and massive
worm-eaten carvings of Scriptural subjects: all these combined to
make the chamber unlike any that Freda had ever seen.
“There!” said Mrs. Bean, as she plumped down the candlestick
upon the writing table, “you’ve never slept in a room like this before!”
“No, indeed I haven’t,” answered Freda, who would willingly have
exchanged fourteenth century tapestry and memories of dead
Abbots for an apartment a little more draught-tight.
“Ah! There’s plenty of gentlemen with as many thousands as the
Captain had hundreds, would give their eyes for the Abbot’s guest
chamber in Sea-Mew Abbey. Now I’ll just leave you while I fetch
some hot water and some dry clothes. They won’t fit you very well,
you being thin and me fat, but we’re not much in the fashion here.
Do you mind being left without a light till I come back?”
Freda did mind very much, but she would not own to it. Just as
Mrs. Bean was going away with the candle, however, she sprang
towards her, and asked, in a trembling voice:
“Mrs. Bean, may I see him—my father?”
The housekeeper gave a great start.
“Bless me, no, child!” she said in a frightened voice. “Who’d ever
have thought of your asking such a thing! It’s no sight for you, my
dear,” she added hurriedly.
Freda paused for a moment. But she still held Mrs. Bean’s sleeve,
and when that lady had recovered her breath, she said:
“That was my poor father’s room, to the right when we reached the
top of the stairs, wasn’t it?”
Again the housekeeper started.
“Why, how did you know that?” she asked breathlessly.
“I saw you look towards the door on the left like this,” said Freda,
imitating a frightened glance.
Mrs. Bean shook her head, puzzled and rather solemn.
“Those sharp eyes of yours will get you into trouble if you don’t
take care,” she said, “unless you’ve got more gumption than girls of
your age are usually blest with. We womenfolks,” she went on
sententiously, “are always thought more of when we don’t seem
over-bright. Take that from me as a word of advice, and if ever you
see or hear more than you think you can keep to yourself, why, come
and tell me—but nobody else.”
And Mrs. Bean with a friendly nod, and a kindly, rough pat on the
cheek which was almost a slap, left the girl abruptly, and went out of
the room.
But this warning, after all the mysterious experiences of the last
two days, was more than Freda could bear without question. She
waited, stupefied, until she could no longer hear the sound of Mrs.
Bean’s retreating footsteps, and then, with one hasty glance round
her which took in frowning bedstead, yawning fireplace and dim
windows, she groped her way to the door, which was unfastened,
and fled out along the stone passage. Her crutch seemed to raise
strange echoes, which filled her with alarm. She hurt herself against
the rough, projecting stones of the wall as she ran. The gallery-door
was open: like a mouse she crept through, becoming suddenly afraid
lest Mrs. Bean should hear her. For she wanted to see her father’s
body. A horrible suspicion had struck her; these people seemed
quite unconcerned at his death; did they know more about it than
they told her? Had he really shot himself, or had he been murdered?
She thought if she could see his dead face that she would know.
Tipity-tap went her crutch and her little feet along the boards of the
gallery. The snow in the court-yard outside still threw a white glare
on the dingy portraits; she dared not look full at them, lest their eyes
should follow her in the darkness. For she did not feel that the
dwellers in this gloomy house had been kith and kin to her. She
reached the landing, and was frightened by the scampering of mice
behind the panelling. Still as a statue she stood outside the door of
her father’s room, her heart beating loudly, her eyes fixed on the faint
path of light on the floor, listening. She heard no sound above or
below: summoning her courage, she turned the handle, which at first
refused to move under her clammy fingers, and peeped into the
room.
A lamp was burning on a table in the recess of the window, but the
curtains were not drawn. There was a huge bed in the room, upon
which her eyes at once rested, while she held her breath. The
curtains were closely drawn! Freda felt that her limbs refused to
carry her. She had never yet looked upon the dead, and the horror of
the thought, suddenly overpowered her. Her eyes wandered round
the room; she noted, even more clearly than she would have done at
a time when her mind was free, the disorder with which clothes,
papers and odds and ends of all sorts, were strewn about the
furniture and the floor. On two chairs stood an open portmanteau,
half-filled. She could not understand it.
Just as, recovering her self-command, she was advancing towards
the bed, with her right hand raised to draw back the curtain, she
heard a man’s footsteps approaching outside, and turned round in
terror. The door was flung suddenly open, and a man entered.
“Who’s in here?” he asked, sharply.
“It is I,” said Freda hoarsely, but boldly. “I have come to see my
father. And I will see him too. If you don’t let me, I shall believe you
have killed him.”
She almost shrieked these last words in her excitement. But the
intruder, in whom she recognised the man she knew as Crispin
Bean, took her hand very gently and led her out of the room.
 CHAPTER IX.
Freda was so easily led by kindness that when, not heeding her
passionate outburst, Crispin pushed her gently out of the room, she
made no protest either by word or action. He left her alone on the
landing while he went back to get a light, and when he rejoined her, it
was with a smile of good-humoured tolerance on his rugged face.
“So you think I murdered your father, do you, eh?” he said, as he
turned the key in the lock and then put it in his pocket.
“Why don’t you let me see him?” asked she, pleadingly.
“I have a good reason, you may be sure. I am not a woman, to act
out of mere caprice. That’s enough for you. Go downstairs.”
Freda obeyed, carrying her crutch and helping herself down by the
banisters.
“Why don’t you use your crutch?” called out Crispin, who was
holding the lamp over the staircase head, and watching her closely.
“If you can do without it now, I should think you could do without it
always?”
He spoke in rather a jeering tone. At least Freda thought so, and
she was up in arms in a moment. Turning, and leaning on the
banisters, she looked up at him with a gleam of daring spirit in her
red-brown eyes.
“It’s a caprice, you may be sure,” she answered slowly. “I am not a
man, to act upon mere reason.”
Crispin gave a great roar of derisive laughter, shocking the girl,
who hopped down the rest of the stairs as fast as possible and ran,
almost breathless, into the room she had been in before. Mrs. Bean
was bringing in some cold meat and eggs, and she turned, with an
alarmed exclamation at sight of her.
“Bless the girl!” she cried. “Why didn’t you wait till I came to you? I
have a bundle of dry clothes waiting outside, and now you’ll catch
your death of cold, sitting in those wet things!”
“Oh, no, I sha’n’t,” said Freda, “we were not brought up to be
delicate at the convent, and it was only the edge of my dress that
was wet.”
Mrs. Bean was going to insist on sending her upstairs again, when
Crispin, who had followed them into the room, put an end to the
discussion by drawing a chair to the table and making the girl sit
down in it.
“Have you had your tea?” asked Freda.
“I don’t want any tea,” said he gruffly. “I’ve got to pack up my
things; I’m going away to-night.”
“Going away!” echoed Freda rather regretfully.
“Well, why shouldn’t I? I’m sure you’ll be very happy here without
me.” And, without further ceremony, he left the room.
Mrs. Bean made a dart at the table, swooped upon a plate and a
knife which were not being used, and with the air of one labouring
under a sudden rush of business, bustled out after him.
There was a clock in the room, but it was not going. It seemed to
Freda that she was left a very long time by herself. Being so tired
that she was restless, she wandered round and round the room, and
thought at last that she would go in search of Crispin. So she opened
the door softly, stepped out into the wide hall, and by the dim light of
a small oil lamp on a bracket, managed to find her way across the
wide hall to the back-door leading into the court-yard. This door,
however, was locked. To the left was another door leading, as Freda
knew, into Mrs. Bean’s quarters. This also was locked. She went
back therefore to the room she had left, the door of which she had
closed behind her. To her astonishment, she found this also locked.
This circumstance seemed so strange that she was filled with alarm;
and not knowing what to do, whether to call aloud in the hope that
Mrs. Bean or Crispin would hear her, or to go round the hall, trying all
the doors once more, she sat down on the lowest steps of the
staircase listening and considering the situation.
A slight noise above her head made her turn suddenly, and
looking up she saw peering at her through the banisters of the
landing, an ugly, withered face. Utterly horrorstruck, and convinced
that the apparition was superhuman, Freda, without a word or a cry,
sank into a frightened heap at the bottom of the stairs, and hid her
eyes. She heard no further sound; and when she looked up again,
the face was gone. But the shock she had received was so great that