1. Introduction
Plastic and composite tanks are widely used in the modern automotive industry. These tanks are fuel tanks, Adblue tanks, water tanks, hydraulic tanks, etc. They offer several advantages over conventional steel tanks, such as lower weight, higher corrosion resistance, better crash performance, etc. These tanks can be used to store highly hazardous substances that can contaminate the environment easily. Therefore, they need to be checked regarding whether they leak or not. There are several ways of detecting leakage of tanks in the automotive industry. As the tanks are big in volume, some leakage test methods are far too expensive, time consuming, and impractical. The easiest, most economical, and most used method is the manual visual based leakage inspection. The tank is immersed in water and pressurized. If bubbles are seen evolving from the tank, the tank is leaking and is to be rejected. The automotive industry has been reluctant to set standards to the bubbles seen in order for the tank to be rejected. The most secure way for the automotive industry was to set the standard as “one should see no bubbles underwater.” The current practice of having a human dependency on critical leakage detection brings a few concerns regarding the reliability due to the tiresome nature of the monotonous test observation. Moreover, manual visual inspection is also time consumptive and results in an undetermined duration of the complete inspection.
Most of the techniques used for fault detection without manual intervention can be divided into three major categories: model-based [
1,
2], signal-based [
3], and knowledge-based. Fault detection systems usually employ artificial neural networks (ANNs) or other classifiers to better identify detection rates. Such intelligent systems consist of three main parts: data acquisition, feature extraction, and data classification. In terms of feature extraction, four major classes of signal processing are used: time-domain [
4], frequency domain [
5,
6], enhanced frequency [
7,
8,
9,
10], and time-frequency analysis [
11,
12]. Deep learning methods have outstanding performances in image classification, computer vision, and fault detection. Convolution neural network (CNN) structure is a type of deep neural network [
13,
14]. Furthermore, the most popular approach for detecting water leak is to use acoustic sensors or a pressure transducer attached to the surface of a pipe [
15,
16,
17,
18]. Although many of the works mentioned above have achieved good results in fault detection, there is still plenty of room for improvement. For instance, in some studies, the classifier was trained for a specific type of data, which means it may achieve high accuracy on similar data while performing poorly with another type of data. Additionally, when analyzing a highly complex system, the choice of suitable feature functions requires considerable machinery expertise and abundant mathematical knowledge. As deep learning works best with unstructured data, most researchers try to apply deep learning as anomaly detection.
The main reason for comparing these techniques is to show fast Fourier transform’s (FFT’s) efficiency in distinguishing crack in underwater tanks. In a faulty tank, bubbles burst in different frequency components, necessitating application of a method investigating these signals in the frequency-domain. Time-domain features consider the signal in several batches and then assigns a value to each of them, resulting in missing some of the information. The results show that time-domain features achieve less accuracy in detecting the fault. The proposed work is based on acoustic emission measurements and a machine learning technique to develop an intelligent fault diagnosis system. This work can be employed to find cracks in industrial components such as immersed underwater tanks.
This paper is organized as follows:
Section 2 describes the mathematical review of FFT, wavelet, and statistical features and the methodology of the proposed fault detection in tanks by applying the 1D-CNN method.
Section 3 explains the experimental setup and the way data were obtained. In
Section 4, the experimental results and analysis are given. Finally, the conclusion is given in
Section 5.
4. Experimental Results and Analysis
A better model on the training data is not necessarily a model that will do better on data it has never seen before. The first plots of validation accuracy and loss showed over-fitting in our model (
Figure 10). In precise terms, the following plots were over-fitting in the FFT, followed by the 1D_CNN technique. The training accuracy increased linearly over time until it reached nearly 100%, whereas the validation accuracy stalled at 70–72%. The validation loss reached its minimum after about ten epochs and then started increasing, whereas the training loss kept decreasing linearly until it reached nearly zero. The number of epochs was a hyper-parameter that defined the number of times the learning algorithm worked through the entire training dataset.
Clearly, after the tenth epoch, we had over-optimized the training data and ended learning representations specific to the training data and thus did not generalize to data outside of the training set. In the other two methods, over-fitting before tuning the hyper-parameters was also visible. They are shown in the following plots (
Figure 11 and
Figure 12).
To find the optimal configuration, we needed to regularize the model and tune the hyper-parameters that neither under-fit nor over-fit.
It was required to modify the model, train repeatedly, and evaluate the validation data (not the test data at this point). Several modification items were:
Adding dropout layers;
Trying different architectures, e.g., adding or removing layers;
Adding L1 and/or L2 regularization;
Trying different hyper-parameters (such as the number of units per layer or the optimizer’s learning rate) to find the optimal configuration;
Optionally, iterating on feature engineering, e.g., adding new features or removing features that were not informative.
Thanks to L2 regularization and adding dropout layers, our plots were no longer over-fitting; the training curves were closely tracking the validation curves, as is obvious in the figures below from
Figure 13,
Figure 14 and
Figure 15. We obtained an accuracy of about 87% or 88% in the model of FFT_1D-CNN, about 52% in the model of time-domain features_1D-CNN, and, finally, 54% in the model of wavelet_1D-CNN.
In this model, we used adaptive moment estimation (Adam) as the main efficient optimization algorithm to update network weights iteratively based on training data. Adam has superiority compared with the other optimizers; it is straightforward to implement, computationally efficient, has few memory requirements, is invariant to diagonal rescale of the gradients, is well suited for problems that are large in terms of data and/or parameters, is appropriate for non-stationary objectives as well as for problems with very noisy/or sparse gradients, and hyper-parameters have intuitive interpretation and typically require little tuning.
We also needed to specify the loss function to evaluate a set of weights. In this case, we used logarithmic loss, which is defined in Keras as binary cross-entropy for a binary classification problem. It is worth mentioning that, in all the mentioned methods, after tuning the hyper-parameters, evaluation was performed by computing the trained model’s accuracy on a held-out validation dataset. Finally, the best hyper-parameter combination in terms of validation accuracy could be tested on a held-out test dataset. A summary of the results regarding changing some hyper-parameters in the 1D-CNN model is mentioned in the
Table 5.
Stochastic gradient descent (SGD) was not a good optimizer compared with Adam optimizer. In other words, Adam optimizer outperformed every other optimization algorithm. Furthermore, using a smaller dropout value of 20% provided a good starting point. It is worth mentioning that a value of dropout that was too low had minimal effect, and a value too high resulted in under-learning by the network. As specified in
Table 6, the total number of layers used in the proposed 1D-CNN architecture was 12.
The summary of the parameters used in the FFT_1D-CNN method is described in
Table 7.
“None” in the above table means that it did not have a pre-defined number. For example, it could be the batch size we used during training and made flexible by not assigning any value. Because of this reason, we could change the size of our batch. The model inferred the shape from the context of the layers. The numbers of epochs and batches were also highly dependent on the types and the quantity of input raw data. It means that fewer epochs may lead to the best result for one method, and, for the other, more epochs are required. For the proposed model, various epochs and batches were tested where the epoch’s best number was 100.
In this experiment, the performance of applied methods was analyzed in terms of accuracy. It was observed that 1D-CNN based on FFT had higher prediction accuracy (86.42%) than the other signal processing techniques presented in
Table 8. It is interesting to note that, in higher epochs, the accuracy of 1D-CNN based on wavelet overtook the accuracy of 1D-CNN based on time domain features, while FFT-1D-CNN showed its superiority in all different conditions. As mentioned earlier, we did not evaluate our model with the test data for a number of times. We only evaluated once the best model gained from the training and the validation datasets on the test dataset. In the following, the results of evaluation on the test model are mentioned.
Author Contributions
Conceptualization, M.R. and A.A.; methodology, M.R. and A.A.; software, M.R. and A.A.; validation, M.R.; A.A.; M.A. and J.H.; formal analysis, M.R. and A.A.; investigation, M.R.; A.A.; M.A. and J.H.; resources, M.R..; data curation, M.R. and A.A.; writing—original draft preparation, M.R.; writing—review and editing, M.R.; A.A.; M.A. and J.H.; visualization, M.R. and A.A.; supervision, M.R. and A.A.; project administration, A.A.; funding acquisition, M.A. and J.H. All authors have read and agreed to the published version of the manuscript.
Funding
This research received no external funding.
Acknowledgments
The authors would like to thank the team of Smart Infrastructure and Industry Research Group at Manchester Met University for all kinds of support for conducting research and preparing the manuscript.
Conflicts of Interest
The authors declare no conflict of interest.
References
- Giantomassi, A.; Ferracuti, F.; Iarlori, S.; Ippoliti, G.; Longhi, S. Electric motor fault detection and diagnosis by kernel density estimation and kullback–leibler divergence based on stator current measurements. IEEE Trans. Ind. Electron. 2014, 62, 1770–1780. [Google Scholar] [CrossRef]
- Gao, Z.; Cecati, C.; Ding, S.X. A survey of fault diagnosis and fault-tolerant techniques—Part I: Fault diagnosis with model-based and signal-based approaches. IEEE Trans. Ind. Electron. 2015, 62, 3757–3767. [Google Scholar] [CrossRef]
- Dai, X.; Gao, Z. From model, signal to knowledge: A data-driven perspective of fault detection and diagnosis. IEEE Trans. Ind. Inform. 2013, 9, 2226–2238. [Google Scholar] [CrossRef]
- Zhou, W.; Habetler, T.G.; Harley, R.G. Bearing fault detection via stator current noise cancellation and statistical control. IEEE Trans. Ind. Electron. 2008, 55, 4260–4269. [Google Scholar] [CrossRef]
- Schoen, R.; Habetler, T.; Kamran, F.; Bartfield, R. Motor bearing damage detection using stator current monitoring. IEEE Trans. Ind. Appl. 1995, 31, 1274–1279. [Google Scholar] [CrossRef]
- Kliman, G.B.; Premerlani, W.J.; Yazici, B.; Koegl, R.A.; Mazereeuw, J. Sensor-less, online motor diagnostics. IEEE Comput. Appl. Power 1997, 10, 39–43. [Google Scholar] [CrossRef]
- Pons-Llinares, J.; Antonino-Daviu, J.A.; Riera-Guasp, M.; Bin Lee, S.; Kang, T.-J.; Yang, C. Advanced induction motor rotor fault diagnosis via continuous and discrete time–frequency tools. IEEE Trans. Ind. Electron. 2014, 62, 1791–1802. [Google Scholar] [CrossRef]
- Arthur, N.; Penman, J. Induction machine condition monitoring with higher order spectra. IEEE Trans. Ind. Electron. 2000, 47, 1031–1041. [Google Scholar] [CrossRef]
- Benbouzid, M.; Vieira, M.; Theys, C. Induction motors’ faults detection and localization using stator current advanced signal processing techniques. IEEE Trans. Power Electron. 1999, 14, 14–22. [Google Scholar] [CrossRef]
- Li, D.Z.; Wang, W.; Ismail, F. An enhanced bi-spectrum technique with auxiliary frequency injection for induction motor health condition monitoring. IEEE Trans. Instrum. Meas. 2015, 64, 2679–2687. [Google Scholar] [CrossRef]
- Eren, L.; Devaney, M.J. Bearing damage detection via wavelet packet decomposition of the stator current. IEEE Trans. Instrum. Meas. 2004, 53, 431–436. [Google Scholar] [CrossRef]
- Ye, Z.; Wu, B.; Sadeghian, A. Current signature analysis of induction motor mechanical faults by wavelet packet decomposition. IEEE Trans. Ind. Electron. 2003, 50, 1217–1228. [Google Scholar] [CrossRef]
- Kiranyaz, S.; Ince, T.; Gabbouj, M. Real-Time patient-specific ECG classification by 1-D convolutional neural networks. IEEE Trans. Biomed. Eng. 2015, 63, 664–675. [Google Scholar] [CrossRef]
- Ince, T.; Kiranyaz, S.; Eren, L.; Askar, M.; Gabbouj, M. Real-Time motor fault detection by 1-D convolutional neural networks. IEEE Trans. Ind. Electron. 2016, 63, 7067–7075. [Google Scholar] [CrossRef]
- Cataldo, A.; Cannazza, G.; De Benedetto, E.; Giaquinto, N. A new method for detecting leaks in underground water pipelines. IEEE Sens. J. 2011, 12, 1660–1667. [Google Scholar] [CrossRef]
- Giaquinto, N.; D’Aucelli, G.M.; De Benedetto, E.; Cannazza, G.; Cataldo, A.; Piuzzi, E.; Masciullo, A. Criteria for automated estimation of time of fight in tdr analysis. IEEE Trans. Instrum. Meas. 2016, 65, 1215–1224. [Google Scholar] [CrossRef]
- Knapp, C.; Carter, G. The generalized correlation method for estimation of time delay. IEEE Trans. Acoust. Speech Signal Process. 1976, 24, 320–327. [Google Scholar] [CrossRef]
- Hero, A.; Schwartz, S. A new generalized cross correlator. IEEE Trans. Acoust. Speech Signal Process. 1985, 33, 38–45. [Google Scholar] [CrossRef]
- Li, B.; Chow, M.-Y.; Tipsuwan, Y.; Hung, J. Neural-network-based motor rolling bearing fault diagnosis. IEEE Trans. Ind. Electron. 2000, 47, 1060–1069. [Google Scholar] [CrossRef]
- Pandiyan, V.; Tjahjowidodo, T. In-process endpoint detection of weld seam removal in robotic abrasive belt grinding process. Int. J. Adv. Manuf. Technol. 2017, 93, 1699–1714. [Google Scholar] [CrossRef]
- Kwak, J.-S.; Ha, M.-K. Neural network approach for diagnosis of grinding operation by acoustic emission and power signals. J. Mater. Process. Technol. 2004, 147, 65–71. [Google Scholar] [CrossRef]
- Al-Habaibeh, A.; Gindy, N. A new approach for systematic design of condition monitoring systems for milling processes. J. Mater. Process. Technol. 2000, 107, 243–251. [Google Scholar] [CrossRef]
- Niu, Y.M.; Wong, Y.S.; Hong, G.S. An intelligent sensor system approach for reliable tool flank wear recognition. Int. J. Adv. Manuf. Technol. 1998, 14, 77–84. [Google Scholar] [CrossRef]
- Betta, G.; Liguori, C.; Paolillo, A.; Pietrosanto, A. A DSP-based FFT-analyzer for the fault diagnosis of rotating machine based on vibration analysis. IEEE Trans. Instrum. Meas. 2002, 51, 1316–1322. [Google Scholar] [CrossRef]
- Rai, V.; Mohanty, A. Bearing fault diagnosis using _t of intrinsic mode functions in Hilbert-huang transform. Mech. Syst. Signal Process. 2007, 21, 2607–2615. [Google Scholar] [CrossRef]
- Pandiyan, V.; Caesarendra, W.; Tjahjowidodo, T.; Tan, H.H. In-process tool condition monitoring in compliant abrasive belt grinding process using support vector machine and genetic algorithm. J. Manuf. Process. 2018, 31, 199–213. [Google Scholar] [CrossRef]
- Sohn, H.; Park, G.; Wait, J.R.; Limback, N.P.; Farrar, C.R. Wavelet-based active sensing for delamination detection in composite structures. Smart Mater. Struct. 2003, 13, 153–160. [Google Scholar] [CrossRef]
- Hou, Z.; Noori, M.; Amand, R.S. Wavelet-Based approach for structural damage detection. J. Eng. Mech. 2000, 126, 677–683. [Google Scholar] [CrossRef]
- Yang, H.-T.; Liao, C.-C. A de-noising scheme for enhancing wavelet-based power quality monitoring system. IEEE Trans. Power Deliv. 2001, 16, 353–360. [Google Scholar] [CrossRef]
- LeCun, Y. Deep learning and convolutional networks. In Hot Chips 27 Symposium (HCS); IEEE: New York, NJ, USA, 2015; pp. 1–95. [Google Scholar]
- Sohaib, M.; Islam, M.; Kim, J.; Jeon, D. Leakage Detection of a Spherical Water Storage Tank in a Chemical Industry Using Acoustic Emissions. Appl. Sci. 2019, 9, 196. [Google Scholar] [CrossRef]
Figure 1.
Flowchart of the proposed methods. 1D-CNN: 1D convolution neural network; FFT: fast Fourier transform.
Figure 2.
Comparison of FFT spectrum for healthy and faulty states.
Figure 3.
Wavelet output of a healthy signal.
Figure 4.
Wavelet output of a faulty signal.
Figure 5.
Comparison of training and test data in terms of accuracy in wavelet_1DCNN.
Figure 6.
1D-CNN structure.
Figure 7.
Schematic illustration of the experimental setup.
Figure 8.
Raw healthy signal (no bubble).
Figure 9.
Raw faulty signal (bubble).
Figure 10.
Accuracy and loss model of FFT_1D-CNN.
Figure 11.
Accuracy and loss model of time-domain features 1D-CNN.
Figure 12.
Accuracy and loss model of wavelet_1D-CNN.
Figure 13.
Accuracy and loss model of FFT_1D-CNN.
Figure 14.
Accuracy and loss model of time-domain features_1D-CNN.
Figure 15.
Accuracy and loss model of Wavelet_1D-CNN.
Table 1.
Previous research efforts in anomaly detection using three signal-processing techniques.
Signal Analysis | Feature | Sensor Used | Process |
---|
Time-domain | Root mean square (RMS) | Acoustic emission (AE) and power | Monitoring grinding operation [21] |
Skewness | Vibration | Condition monitoring for milling [22] |
Kurtosis | AE | Tool flank wear recognition [23] |
Frequency- domain | FFT | Vibration | Fault diagnosis of the rotating machine [24] |
FFT | Vibration | Bearing fault diagnosis [25] |
FFT | Vibration, AE, force | Tool wear monitoring [26] |
Wavelet | Morlet wavelet | Piezoelectric sensor | Delamination detection [27] |
Daubechies-4 | Vibration | Structural damage detection [28] |
Daubechies-4 | Power | Power quality monitoring [29] |
Table 2.
Statistical time-domain features.
Features | Expression |
---|
Mean value | $\overline{x}=\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$N$}\right.{\displaystyle {\displaystyle \sum}_{i=1}^{N}}{x}_{i}$ |
Standard deviation | $SD=\sqrt{\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$N$}\right.{\displaystyle {\displaystyle \sum}_{n=1}^{N}}{\left({x}_{i}-\overline{x}\right)}^{2}}$ |
Kurtosis | K = $\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$N$}\right.\sum _{i=1}^{N}\frac{{\left({x}_{i}-\overline{x}\right)}^{4}}{{\sigma}^{4}}$ |
Skewness | S = $\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$N$}\right.\sum _{i=1}^{N}\frac{{\left({x}_{i}-\overline{x}\right)}^{3}}{{\sigma}^{3}}$ |
Root mean square | RMS = $\sqrt{\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$N$}\right.\sum _{i=1}^{N}{x}_{i}{}^{2}}$ |
Crest factor | C = $\raisebox{1ex}{$\mathrm{max}value$}\!\left/ \!\raisebox{-1ex}{$RMS$}\right.$ |
Peak-to-peak (PPV) value | PPV = max value$-$ min value |
Table 3.
Statistical features values.
Label | Mean | STD | Skewness | Kurtosis | RMS | Peak-to-Peak | Crest-Factor |
---|
1 | −0.00303 | 0.932716 | 0.001772 | −1.55262 | 0.932721 | 3.166601 | 1.694493 |
1 | 0.001391 | 0.934791 | 0.002414 | −1.55813 | 0.934792 | 3.135015 | 1.649781 |
1 | 0.004578 | 0.942283 | −0.00174 | −1.50837 | 0.942295 | 3.411634 | 1.81576 |
1 | −0.00252 | 0.936973 | 0.005309 | −1.52431 | 0.936976 | 3.431415 | 1.81483 |
0 | 5.40 ×10^{−5} | 0.813413 | −0.00043 | −1.56246 | 0.813413 | 2.636973 | 1.594723 |
0 | 0.001712 | 0.813213 | 0.002127 | −1.57254 | 0.813215 | 2.619106 | 1.583342 |
0 | 0.001307 | 0.815626 | 0.001999 | −1.57703 | 0.815627 | 2.609215 | 1.582571 |
0 | 0.000724 | 0.812949 | 0.002658 | −1.57934 | 0.812949 | 2.607301 | 1.604661 |
Table 4.
Statistical time-domain features.
Variables | Quantity |
---|
Data | 127 samples |
Number of buckets | 120 |
Each bucket | (100,000, 1) |
Full data array | (15,240, 100,000) |
Full label array | (15,240, 1) |
Train data | (2172, 1000) |
Validation data | (1070, 1000) |
Test data | (1598, 1000) |
Table 5.
Impact of the different parameters and methods on the accuracy of the model.
Parameters Methods | Activation Function (Tanh) | Dropout (0.5) | Having 2 Dense Layers | Having 2 Convolution Layers | Optimizer Stochastic Gradient Descent (SGD) |
---|
FFT_1D-CNN | 74.53 | 81.66 | 83.85 | 83.55 | 56.45 |
Wavelet_1D-CNN | 47.68 | 47.68 | 47.68 | 47.68 | 47.68 |
Time-domain features_1D-CNN | 52.88 | 53.69 | 53.94 | 56.45 | 48.19 |
Table 6.
Structure of the 1D-CNN model.
Layer | Name | Specification |
---|
1 | Convolution | 2×2×1 |
2 | Relu | N/A |
3 | Max pooling | 2×2 |
4 | Flatten | 998 |
5 | Dense | 128 |
6 | Sigmoid | N/A |
7 | Dense | 64 |
8 | Sigmoid | N/A |
9 | Dropout | 20% |
10 | Fully Connected | 1 |
11 | Sigmoid | N/A |
12 | Classification | Binary cross-entropy |
Table 7.
Impact of the different parameters on the methods.
Layers | Methods |
---|
FFT | Wavelet | Time Domain |
---|
Shape Param | Shape Param | Shape Param |
---|
Conv 1D | (None, 999, 2) 6 | (None, 6, 2) 6 | (None, 999, 2) 6 |
MaxPooling1 | (None, 449, 2) 0 | (None, 2) 0 | (None, 449, 2)0 |
Flatten | (None, 998) 0 | (None, 6) 0 | (None, 998)0 |
Dense | (None, 128) 127,872 | (None, 128) 896 | (None, 128) 127,872 |
Dense | (None, 64) 8256 | (None, 64) 8256 | (None, 64) 8256 |
Dropout | (None, 64) 0 | (None, 64) 0 | (None, 64) 0 |
Output (Dense) | (None, 1) 65 | (None, 1) 65 | (None, 1) 65 |
Total Params | 136199 | 9223 | 136199 |
Table 8.
The fault detection performance of the proposed methods on the test dataset.
Methods | Accuracy (%) |
---|
Epoch = 50 | Epoch = 100 |
---|
1 | FFT_1D-CNN | 84.23 | 86.42 |
2 | Wavelet_1D-CNN | 52.88 | 54.57 |
3 | Time-domain Features_1D-CNN | 54.32 | 53.94 |
| Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. |
© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).