A Study on the Fault Diagnosis Technology and Application for Industrial Equipment using Deep Learning

Abstract

With the recent progress of the 4th industrial revolution, many manufacturing industries focus on implementing a smart factory that can maximize the effect of increasing productivity and reducing labor costs. For maintaining stable operation of the smart factory, countermeasures must be prepared according to the increased uncertainty and complexity of the advanced system. To do this, it is necessary to extend the life of each equipment by utilizing an appropriate maintenance strategy. In this dissertation, deep learning-based algorithms are studied for reliable fault diagnosis of bearings and induction motors of a rotating equipment as well as circulating fluidized bed combustion boiler (CFBC) boiler tubes. First, a bearing fault diagnosis technology is introduced based on a convolutional neural network (CNN). To improve classification performance of the CNN, a normalized bearing characteristic component (NBCC) is used generated by extracting bearing defect frequencies from spectrum of the acoustic emission signal. In addition, importance weights of features are extracted by using gradient-weighted class activation mapping (Grad-CAM) to enable CNN interpretation. Experimental results show that the proposed method achieves high classification accuracy and CNN successfully learns bearing’s characteristic frequency for each type of the bearing failure. To further improve performance of the deep learning-based fault diagnosis method, two-dimensional CNN-based fault diagnosis methods have been studied by converting one-dimensional signal into two-dimensional data and learning it. In this study, a CNN-based fault diagnosis method is proposed using stacked envelope spectral image (SESI) to find a 2D representation of the AE signal based on the fault characteristic frequency of the bearing. SESI is designed to include general-purpose fault characteristics of bearings by extracting and stacking bearing fault frequencies from the envelope spectrum. When learning a 2D CNN using SESI, the learned CNN can directly learn fault frequency of the bearing. To verify performance of the proposed method, Experimental results show that fault diagnosis for each type of bearing with high performance is possible by mutually learning the data acquired from two different testbeds and performing a diagnostic test. In addition to SESI, another new method, called “defect signature wavelet image (DSWI)”, is established to construct the 2-D fault diagnosis representation of multiple bearing defects from 1-D acoustic emission signals. This technique starts by applying envelope analysis to extract the envelope signal. A novel strategy is propounded for the deployment of the continuous wavelet transform with damage frequency band information to generate DSWI, which describes acoustic emission signal in time-frequency-domain, reduces the nonstationary effect in the signal, shows discriminate pattern visualization for different types of faults, and associates with the defect signature of bearing faults. Using the resultant DSWI, CNN architecture is designed to identify the fault in the bearing. To evaluate the proposed algorithm, the performance of this technique is scrutinized by a series of experimental tests acquired from a self-designed testbed and corresponding to different bearing conditions. Experimental results demonstrate that the proposed methodology outperforms conventional approaches in terms of classification accuracy. The result of combining CNN with DSWI input yields an accuracy of 98.79% for classifying multiple bearing defects. Although deep learning-based fault diagnosis methods show excellent performance, machine learning-based fault diagnosis methods are still being studied. Since machine learning-based fault diagnosis methods use features designed by experts, performance is excellent even when a small amount of data is learned compared to deep learning-based fault diagnosis, and features can also be analyzed. This thesis proposes a technique for diagnosing incipient bearing defects under variable speed conditions, by extracting features from different sub-bands of the inherently non-stationary AE signal, and then classifying bearing defects using a weighted committee machine, which is an ensemble of support vector machines and artificial neural networks. The proposed method also improves the generalization performance of neural networks to enhance their classification accuracy, particularly with limited training data. In addition to bearings, induction motors are also very important parts that transmit power to drive equipment. Therefore, not only bearing fault diagnosis but also fault diagnosis of induction motors is vital in terms of maintenance of the equipment. In this thesis, a Mahalanobis distance-based classifier is proposed, which can diagnose various defects of induction motors including bearing failures, rotor unbalance, broken rotor bar, bowed rotor shaft, and rotor misalignment. The proposed method extracts an effective feature vector using the difference in harmonic components related to the failure of the vibration signal. After generating features from harmonic components for defects, diagnostic performance is improved with a classifier using Mahalnobis distance. Experimental results demonstrates that the proposed method has higher classification performance than the conventional method in both noiseless and white Gaussian noise environments. Finally, a technique for estimating the tube leakage location of a CFBC boiler for thermal power generation is proposed. Since the fluid medium, which is the fuel of the CFBC boiler, is a small but hard solid like pulverized coal, it may cause abrasion of the waterwall tube as well as leakage due to the blow of the fluid medium. A method of estimating the leakage location of the CFBC boiler tube for thermal power generation is proposed using an acoustic emission sensor. The proposed method uses an acoustic emission sensor that can effectively detect the acoustic waves generated by the movement of a medium molecular unit, and uses a sensor sensitivity estimation algorithm for each location that considers the attenuation rates of the membrane welds and non-welded parts of the boiler water wall tube.