Professional Documents
Culture Documents
A R T I C LE I N FO A B S T R A C T
Keywords: To avoid the sudden failure of mechanical structures under repeated loading and to supplement conventional
Failure detection methods of managing life time, we propose a failure detection technique under random fatigue loading using
Fatigue loading machine learning and dual sensing on a symmetric structure. The state of a shackle, which is used to connect the
Machine learning cargo to the hoist efficiently, under fatigue loading was collected using two strain sensors of a dual system. The
Dual sensing
strains were preprocessed and labeled as normal or abnormal. Logistic regression machine learning was em-
Prognostics and Health Management (PHM)
ployed to determine the decision boundary line. Then, we gathered the decision boundary lines of each ex-
periment for determining the time of failure, and we verified every experiment with the most conservative
decision boundary line. The results indicate that failure was detected before the crack occurred and the time to
notice maintenance could be controlled.
⁎
Corresponding author.
E-mail address: dcbaek@kimm.re.kr (D.-C. Baek).
https://doi.org/10.1016/j.ijfatigue.2018.05.004
Received 26 December 2017; Received in revised form 2 May 2018; Accepted 3 May 2018
Available online 04 May 2018
0142-1123/ © 2018 Elsevier Ltd. All rights reserved.
D.-W. Jang et al. International Journal of Fatigue 114 (2018) 57–64
the shackle under repeated loading (Fig. 1). Repeated loading was ap-
plied using a tensile tester (Landmark 100 kN, MTS, USA), and strain
data was collected by a data logger. Repeated loading has sequences
with a frequency of 8 Hz. Data was measured discretely to reduce
computer memory (2, 4, 6, or 10 s in 60 s).
There are standard load spectra for hoists such as ASME HST-4, FEM
9.511, and FEM 9.683 [31–33]. Because shackles are usually used as
connecting components of hoists and a weight, it is better to use the
load spectrum for hoists to evaluate the fatigue life of shackles. How-
ever, the goal of this paper is not to evaluate fatigue life, but to predict
failure using strain signals. Therefore, standard load spectra were only
used as reference, and the load spectra for an accelerated fatigue test
were used.
The details of the loading sequences are presented in Table 1 and
Fig. 2. If the load falls below 0, the shackle and jig can move apart and
cause misalignment and wear. Therefore, the shackle was preloaded
with 980 N and every loads were applied as positive values. Every
loading sequence has the same load valley; however, they have dif-
ferent upper load peaks characterized by the maximum peak load for
each load sequence of 13 kN, 16.7 kN, and 20.6 kN. The load spectra of
three types of loading sequences are presented from Fig. 2(d)–(f). The
Fig. 1. A dual sensing system for a shackle. Two strain gauges were installed on maximum peak loads of 12.40 kN, 13.23 kN, and 13.72 kN are cate-
the outer left and outer right surfaces of the shackle. gorized as Test set I; those of 16.7 kN, Test set II; and those of 20.6 kN,
Test set III. Because it was an accelerated test, the maximum peak load
and some load magnitudes exceed the working load limit of the shackle,
which is 9.8 kN. Furthermore, the maximum peak load in Test set III
exceeds a yield load of 19.6 kN. It included a severe condition to vali-
Table 1
date the algorithm with overloading.
Details of loading sequence.
To distinguish between normal data and abnormal data, the surface
Test set Experiment # Maximum peak load Measuring scheme [s/min] of the shackle was observed with a CCD camera (WISE-MN28V,
[N]
WISECAM, Republic of Korea) to determine whether cracks occurred.
I 1 12,740 2 When a crack was identified in the video image, the time was recorded.
2 4 The crack occurred at a specific location, i.e., the inner curved surface
of the shackle’s leg. Cameras could not observe the inner curved surface
3 13,230 2
of the shackle’s leg perpendicularly owing to geometrical limitations;
4 2
5 2 however, a slightly tilted side view could be observed. The detectable
crack size in the video image was about 1 mm in the titled side view. It
6 13,720 2 occurred randomly in one of the two legs. For the convenience of
processing, we set the strain data from the cracked leg as strain 1, which
II 7 16,660 10
8 10
represents the x-axis in Figs. 4–6. We classified data as normal before
9 6 cracks were detected, and abnormal after cracks were detected. Fig. 3
shows the strain data in one experiment.
III 10 20,580 10 Initially, the two strain plots showed the same trend of variation with
11 10
different magnitudes. Although the shackle is symmetric and the load is
12 10
applied perpendicularly, the two strains have different magnitudes. The
shackle is symmetric in the macro scale; however, it is not perfectly sym-
metric at the micro scale, because material distribution is not uniform and
abnormal. Logistic regression [28–30] machine learning was carried the geometry is also asymmetric. Another reason is that strain gauges were
out to determine the decision boundary line for 12 experimental sets. not perfectly aligned. When the cyclic load was continuously applied, the
Then, we gathered 12 decision boundary lines for determining the time values in one part of the strain plot (black) were reduced, and cracks were
of failure from various experiments and verified every experiment with generated and broken because they originated from the inner side of the
the most conservative decision boundary line. As a result, it was ver- shackle such that the outer surface where the strain gauge is located re-
ified that failure was detected before the crack occurred and the ceived a compressive load. On the other hand, the other strain plot (red) did
prognostic distance could be controlled. not change until the shackle was broken.
If the graph is drawn with one strain data on the x-axis and the other
strain data on the y-axis, the graph is shown in Fig. 4. Since the two
2. Experiments
strain data plots in the initial and intermediate stage showed the same
trend with different sizes, data before 4000 s was plotted on the narrow
We measured the dual strain data of a shackle under repeated
area on the top of the graph. The corresponding color is from red to
loading. Strain gauges were installed on the outer left and outer right of
58
D.-W. Jang et al. International Journal of Fatigue 114 (2018) 57–64
Fig. 2. (a–c) Every loading sequence has the same lower load valley; however, they have different upper load peaks that are characterized by the maximum peak load
for which each load sequence is 13, 16.7, and, 20.6 kN. In addition, the load spectra of three types of loading sequence are presented from (d) to (f).
59
D.-W. Jang et al. International Journal of Fatigue 114 (2018) 57–64
Fig. 4. Strain vs. strain from the dual sensing system. The color of the data
shows time and its meaning is on the right side of the figure. (For interpretation
of the references to colour in this figure legend, the reader is referred to the web
version of this article.)
straight line near the maximum value. As shown in Fig. 5, the hysteresis
loop is observed. After the maximum load is applied, the data does not
go back along the loading path. However, data does not move along the
linear path as well and its shape is distorted. This implies that there is
plastic deformation, but it is not the main reason for non-linearity on
the top right part of the strain-versus-strain graph. One possible reason
is the non-linear geometrical behavior such as stretching of the legs of
the shackle.
When learning includes the curved section (since the classification
model was characterized by a polynomial), the learning time increased Fig. 6. Graph of normalized and linearized strain from the dual sensing system
and the modification of the decision boundary line becomes compli- (a) and its magnified graph (b).
cated. Therefore, normalization was performed again with the straight-
line section from (0.2, 0.2) to (0.5, 0.5) in Fig. 5 to use only the linear
Data learning, as mentioned earlier in the introduction, was carried
part. When data from the linear part was used, the decision boundary
out using logistic regression.
line obtained through logistic regression was linearly formed such that
Logistic regression is a regression model for classifying data into two
the learning time was shortened and the control of the decision
types [28–30]. With the logistic regression model, the computer
boundary line became simple.
Fig. 5. Graph of the normalized strain from the dual sensing system and one cycle of the magnified graph.
60
D.-W. Jang et al. International Journal of Fatigue 114 (2018) 57–64
Fig. 7. Labeled normalized strain 1 (x-axis) versus normalized strain 2 (y-axis) plot and its decision boundary line by logistic regression for 12 experiments.
attempts to find weights and bias of hypothesis that have the highest typical examples are the mean of squared error and a cross entropy. The
possibility to divide data correctly and this process is called learning. cross entropy is a function to measure the distance of two probability
The hypothesis is a mathematical function. Weights and bias are coef- distributions and is usually used as a cost function in the logistic re-
ficients and the constant of a function. In the learning process, the gression model [34].
computer changes weights and the bias by reducing the cost function; In this paper, a linear function that has one weight and one bias was
61
D.-W. Jang et al. International Journal of Fatigue 114 (2018) 57–64
62
D.-W. Jang et al. International Journal of Fatigue 114 (2018) 57–64
Table 3 Table 4
Ratio of anomaly detection time and crack detection time from modified al- Ratio of anomaly detection time and crack detection time from modified al-
gorithm with the criterion which is several cycles earlier than actual crack gorithm with modified decision boundary; Δ is the difference between the
detected time; Δ is the difference between the anomaly detection time and crack anomaly detection time and crack detection time.
detection time.
Experiment # Conservative method 2–1 Conservative method 2–2 Modified
Experiment # Conservative method 1–1 500 s Conservative method 1–2 1000 s Modified intercept of decision intercept of decision boundary line
earlier classification earlier classification boundary line as 90% of original as 50% of original value
value
Anomaly Anomaly detection Anomaly Anomaly detection
detection time/actual crack detection time/actual crack Anomaly Anomaly detection Anomaly Anomaly detection
time [s] detection time [%] time [s] detection time [%] detection time/actual crack detection time/actual crack
([Δ]) ([Δ]) time [s] detection time [%] time [s] detection time [%]
([Δ]) ([Δ])
1 50,600 82.3 (10900) 33,500 54.5 (28000)
2 42,900 95.8 (1900) 42,100 94.0 (2700) 1 54,900 89.3 (6600) 35,600 57.9 (25900)
3 17,000 80.2 (4200) 16,000 75.5 (5200) 2 43,600 97.3 (1200) 42,100 94.0 (2700)
4 28,200 77.0 (8400) 22,400 61.2 (14200) 3 17,700 83.5 (3500) 16,000 75.5 (5200)
5 15,900 75.7 (5100) 14,500 69.0 (6500) 4 30,900 84.4 (5700) 23,100 63.1 (13500)
6 18,700 78.2 (5200) 17,900 74.9 (6000) 5 16,900 80.5 (4100) 14,700 70.0 (6300)
7 6000 93.8 (400) 5520 86.3 (880) 6 19,100 79.9 (4800) 17,900 74.9 (6000)
8 4080 72.9 (1520) 3780 67.5 (1820) 7 6180 96.6 (220) 5580 87.2 (820)
9 6900 86.3 (1100) 6000 75.0 (2000) 8 4260 76.1 (1340) 3780 67.5 (1820)
10 1870 50.5 (1830) 1560 42.2 (2140) 9 7740 96.8 (260) 6120 76.5 (1880)
11 2940 91.9 (260) 550 17.2 (2650) 10 2160 58.4 (1540) 1560 42.2 (2140)
12 550 17.2 (2650) 480 15.0 (2720) 11 2950 92.2 (250) 2880 90.0 (320)
12 1560 48.8 (1640) 480 15.0 (2720)
63
D.-W. Jang et al. International Journal of Fatigue 114 (2018) 57–64
64