You are on page 1of 4

Artificial Intelligence for Preventive Maintenance

Rajat Shukla
Department of Computer Applications
IPEM Group of Institutions
A-13/1, South Side G.T Road, Industrial Area, Ghaziabad, Uttar Pradesh 201010

shuklarajat265@gmail.com

Abstract— This paper explores the application of data analytics for known inputs and outputs is fed to the ANN. Then, an optimization algorithm
predictive maintenance in manufacturing sector. By leveraging minimizes the value of the ( ̂ ) , where ̂ and are the estimated
machine learning and real-time sensor data, we demonstrate the
and actual outputs. Through several iterations, the algorithm tweaks the values of
effectiveness of predictive maintenance in minimizing downtime and
the weights until the MSE falls below a determined threshold.
reducing operational costs. Case studies illustrate practical
implementation and benefits, contributing to improved equipment Figure 1. A schematic drawing of an artificial neural network
efficiency and lifespan.

1. Introduction
Tasks include applications such as decisionmaking and expert systems, pattern
recognition, classification of data, visual and audio recognition, stock prediction,
medical diagnosis and text and speech translation[2–5]. AI usually uses artificial
neural networks (ANN) to function in an intelligent-like manner[6].

The output from each neuron becomes an input to any neuron in the following layer
or even as a feedback to itself. During the training process of an ANN, data with

Recently, AI has been increasingly used in predicting the requirement and planning 2. Method
of maintenance operations for machines[9,10]. This field is now known to as
predictive maintenance (PM). Predictive maintenance is a technique where In this work, a dataset from the public domain is used. It was obtained from the UC
maintenance is conveniently scheduled according to the condition of a machine in Irvine Machine Learning Repository [13]. The data is synthetic and represents
terms of its degree of degradation and probability of failure. This is in contrast to simulated measurements of sensors connected to an industrial machine. The use of
preventive maintenance where maintenance is conducted periodically or according such synthetic data was because real ones are not publically available nor easily
to time intervals. The problem with the conventional preventive maintenance obtained. The dataset contains 10,000 records of different measurements. The
approach is that the time intervals at which maintenance operations are performed measurements are: (1) ambient air temperature [K], (2) process temperature [K], (3)
are calculated according to statistical estimates based on standard operating rotational speed [rpm], (4) torque [Nm], (5) a measurement representing tool wear,
conditions. Usually, mean-time-to-failure (MTTF) is calculated from observations (6) the type of product being produced represented by one of the letters (L, M or H)
of samples of standard operations. Deviations from these standard conditions are and (7) finally failure represented by ones or zeros for a working machine. The
frequent in real life and cause failure to occur either earlier or later than the expected process is shown in figure 2.
MTTF. Premature failure results in sudden breakdown and in some cases can be
catastrophic. On the other hand, if maintenance is conducted prematurely, this Figure 2. Flow diagram of Predictive maintenance process
means a waste of some useful life and resources.PM allows the full use of useful
lifetime of the machine and prevents the waste of resources on needless
maintenance operations. For the application of PM, the condition of the machine
monitored through appropriate sensors and data is collected. Data is then analyzed
and used to train an artificial neural network that can effectively predict the
probability of failure of the machine. This ANN can then be deployed to an
embedded system within the machine. The system supervises the condition of the
machine through sensors readings and informs the system administrator about
current machine condition and maintenance can be appropriately made when
needed. PM has been increasingly used
in heavy industries such as steel where the cost of a downtime of a machine is high
and can reach up to 60% of total production cost [11]. Moreover, it has been
estimated that around one third of the budget for maintenance was wasted on
unnecessary or improper maintenance operations. Considering that in the US alone,
the estimated costs of maintenance amounted for $50 billion in 2016 [12]. This
shows the importance of implementing PM in order to reduce manufacturing costs.
The rapid development of measuring instrumentation, computer systems, data
storage, cloud computing and fast internet connections have all facilitated the
application of PM in industry. However, technical difficulties arise during the
application of AI to predictive maintenance. These difficulties may result in
undesirable outcomes to the point that renders the whole application nonusable. In
this work, we aim at exploring the whole process through the comprehensive study
of a case where AI is applied to a maintenance dataset to explore the process and 3. Results and discussions
show the potentials and benefits of implementing predictive maintenance.
3.1 Data analysis and visualization

In order to develop an efficient AI model, at first, we should examine the data for
indications of correlations between variables and whether these affect the output
response. First, we analyze the data using boxplots and histograms as shown in
figure 4, It is clear that temperature values have limited range ~8 K and the values
are roughly uniformly distributed. Speed is skewed to the left (lower end) and is
centered on 1540 rpm, with some outliers towards the higher end. Torque is
normally distributed with outliers at both lower and higher ends. As per wear, it is
uniformly distributed along the full range; however, large values are less frequent
than others are. Hence, our initial estimation is to expect speed and torque to have
the major effect followed by wear on machine failure.

Figure 5. Higher failures for high torque and low


3.2 Modeling
This section presents the modeling of an ANN using the dataset. An ANN is trained
to classify the condition (operational/failed) of the machine based on the six
variables (air temperature, process temperature, rotational speed, torque, wear and
type of product). An ANN feed forward neural network classifier was used. Since
the dataset was unbalanced, i.e. the number of failure incidents was approximately
30 times smaller than operational ones; any artificial classifier will be biased to
predict operational incidents with higher accuracy than failed ones. The metric for
judging the accuracy of prediction is the confusion matrix chart as shown in figure
6. that shows the results of training the final ANN after many attempts. The
confusion matrix represents the true class (actual measurements) in the rows
placement and predicted outcomes are in columns. Using validation dataset, the
final ANN was able to predict correctly the no-failure condition for 1924 cases
compared to 21that were actually failed but were wrongfully predicted as non-
failure. The right hand side of figure 7 shows a summary of the percentages of each
corner representing combination between true and predicted values. As expected, a
99.5% true positive rate (TPR) was obtained in classifying non-failures compared to
68.7% true negative rate (TNR) for classifying true failures.

Figure 3. Boxplots and histograms for each


variable
By examining machine failure occurrences, there are exactly 339 failure incidents
out of the 10,000 records. In terms of pure machine operations terms, this translates
to a3.39% probability of failure, however, this also means that there is unbalance
between the numbers of records representing operational versus failed machines.
Consequently, this will affect the affect the training process of the artificial neural
network as will be discussed later in section 3.2. Descriptive statistics are presented
in Table 1. It is clear that mean values of temperatures of air and process were
almost equal during operational and failed conditions. Considering the remaining
variables: speed, torque and wear, it is clear that there was a noticeable difference
between their respective mean values for healthy and failed conditions. In addition,
one can notice that failure incidents increase when the machine is operated at high
torque and low speed values. This is shown in figure 5. This observation agrees with
our former initial guess that torque and speed are influential factors to failure;
however, other factors are also important. Since it is not easy to role of each factor
by simple data visualization, we use artificial intelligence to reveal these hidden
relations between input variables and the output response.
Figure 5. Results of confusion matrix
Table 1. Descriptive statistics
3.3 Performance evaluation
Failure 0 1
Group Count 9661 339 This section describes the measures that were taken in order to improve the
Mean (Air_temp) [K] 300.0 300.9 capability of the ANN for predicting failure incidents. i.e., TNR. Two more ANNs
Mean(Process_temp) 310.0 310.3 have been studied as shown in Table 2, where NN_1 represents the first network
[K] that was discussed earlier; NN_2 and NN_3 refer to an ANN with two and three
Mean (Speed) [rpm] 1540.3 1496.5 hidden layers. In addition, the dataset was partitioned into two sets: training that was
Mean (Torque) [Nm] 39.6 50.2 80% of the original dataset and the remaining 20% was for validating the results of
the ANN. The validation dataset was held out from the ANN during the training
phase in order to ensure that the ANN will be able to predict the failure condition
based only on its own learning experience. Since the original dataset contained non-
failure records in the top rows, we ensured that the dataset be shuffled before
portioning it into separate training and validation datasets. This was done to make
sure that the data in each dataset was homogeneous as much as possible. Finally, the
values of predictors were standardized so that the learning of the ANN will not be
affected by the difference in scales between predictor variables. As shown in Table
2, these actions increased the prediction ability in terms of TPR and TNR.

3.4 Prediction

Having the ANN developed, it can now be used to predict the condition of the
machine based on predictive measurements. As an example of the potentials of the
ANN, a dataset is used. This dataset includes 100 records that were extracted from
the original dataset where the machine had failed. Feeding these records into our
ANN, it predicted that 77 have failed, while it wrongfully predicted 23 as
operational/non-failures. However, the results also include predictions for the
probabilities of failure for each record. The values of probability of failure for each
of the 23 records are shown in figure 7. Our ANN considered a probability of failure
below 0.5 as operational and given it has a 77% success in predicting failure; we can
consider a lower threshold as for example, 0.1. In other words, the ANN will be
given data from the sensors monitoring the machine and warnings to system Figure 7. Decision tree
administrator are shown when a prediction is higher than a predetermined threshold.
Consequently, further investigation will be conducted in order to take corrective
4. Conclusions
actions.
This paper presented a case study for applying artificial intelligence techniques to
the rather new field of research, which is predictive maintenance. A synthetic
dataset published in the public domain that represents 10,000 records of sensors
measurements and failure incidents of an artificial machine was used for
demonstration. Artificial neural networks were used to classify the condition of the
machine as failed or operational. We can conclude based on our current study that
the inherent complexities with applying artificial intelligence to the field of
predictive maintenance are about obtaining a high quality data. We opted to use a
synthetic dataset for this reason. Experiments can be setup to produce useful and
reliable data; however, they are expensive and time consuming. High fidelity
simulations can provide a good source of data; however, they also require high
quality inputs so they can be relied upon. Those models are generally referred to as
digital twins of the real physical machines. Finally, it can be concluded that the
application of artificial intelligence in the field of maintenance will lead to financial
benefits and systems that are more efficient.
Figure 6. Probability of failure for each wrongfully
predicted record References
[1] Joshi A V. 2020 Machine learning and artificial intelligence (Cham: Springer)
3.5 Alternative predictive models
[2] Virvou M, Tsihrintzis G A and Tsoukalas L H Advances in Artificial
Finally, we briefly give an overview of other models that can be used in predicting Intelligence-based Technologies Selected Papers in Honour of Professor vol 1
the conditions of machines. It should be noted that, this overview is limited and
serves only to give a general idea of the capabilities of predictive modeling in the [3] Ognjanovic I 2020 Artificial intelligence in healthcare vol 274
field of maintenance. For more information on predictive models, one can refer to
[11,14–19]. [4] Hassanien A E 2020 Advances in Intelligent Systems and Computing 723
The decision tree model presents inputs and an output in a flowchart-like structure, Mohamed Mostafa Editors The International Conference on Advanced Machine
which makes it easy to interpret. Figure 8 shows a view of the decision tree Learning Technologies and Applications ( AMLTA2018 )
predictive model [20]. This type of predictive models gives a clear and intuitive
indication to the user. Another form of predictive models is logistic regression [21]. [5] Partridge D 2013 Artificial intelligence and software engineering: Understanding
It is clearly shown that the polynomial equation represents Failure occurrences as a the promise of the future
function of process variables.
[6] Jackson P C 2019 Introduction to artificial intelligence (Mineola New York:
Dover Publications Inc.)

[7] Kamel H 2012 Neural Network Application for Structure Design Optimization
of Thin-Wall Structures ASME 2011 International Mechanical Engineering
Congress and Exposition, IMECE 2011 vol 9 (American Society of Mechanical
Engineers Digital Collection) pp 41–5

[8] Anon File:Artificial neural network.png - Wikimedia Commons

[9] Hu C, Fan H and Wang Z 2022 Residual Life Prediction and Optimal
Maintenance Decision for a Piece of Equipment

[10] Filomena M, Justyna S and Erika T Innovations in Mechanical Engineering

[11] Mobley R K 2002 Introduction to Predictive Maintenance


[12] Thomas D S 2018 The Costs and Benefits of Advanced Maintenance in
Manufacturing (Gaithersburg, MD)

[13] Matzka S 2020 UCI Machine Learning Repository: AI4I 2020 Predictive
Maintenance Dataset Data Set

[14] Mathworks C 2020 Predictive Maintenance Toolbox TM User ’ s Guide R


2020 a

[15] Alexandre J, Marques L, Nauber F, Gois José B, Simon X-N and Fong J 2021
Predictive Models for Decision Support in the COVID-19 Crisis

[16] Lughofer E and Sayed-Mouchaweh M 2019 Prologue: Predictive maintenance


in dynamic systems

[17] Mathworks 2019 Predictive Maintenance: Extracting Condition Indicators with


MATLAB

[18] Killeen P, Ding B, Kiringa I and Yeap T 2019 IoT-based predictive


maintenance for fleet management Procedia Comput Sci151 607–13

[19] Çinar Z M, Nuhu A A, Zeeshan Q, Korhan O, Asmael M and Safaei B 2020


Machine learning in predictive maintenance towards sustainable smart
manufacturing in industry 4.0 Sustain12

[20] Kotsiantis S B 2011 Decision trees: a recent overview Artif Intell Rev 2011
39439 261–83 [21] Pampel F C 2000 Logistic Regression: A Primer Sage Univ
Press132 98

You might also like