Journal
IEEE Transactions on Instrumentation and Measurement
Keywords:
Vibrations;
Induction motors;
Fault diagnosis;
Transforms;
Time-frequency analysis;
Feature extraction;
Convolutional neural networks
Effective Fault Diagnosis Based on Wavelet and Convolutional Attention Neural Network for Induction Motors
Dr. Tran Minh Quang
Industry 4.0 Center, National Taiwan University of Science and Technology, Taipei, Taiwan
Email: minhquang.tran@mail.ntust.edu.tw
Summary of scientific publications
The development of industries in general and the operation of factories in particular has an important contribution from electric motors. Electric motors have been widely used in many industrial fields because of their cost effectiveness, simple structure, and ease of maintenance during operation. However, like most other industrial equipment, electric motors can develop many types of damage after a long period of operation, which will cause serious problems if not detected. and early diagnosis. Common failures on electric motors are usually bearing damage, stator failure and rotor failure. This will generate more heat, reduce output torque, consume more power and can also cause serious engine damage. Therefore, early detection and diagnosis of damage to electric motors can help prevent damage from becoming more serious, while maintaining machine operation in normal conditions and reducing dependence on electric motors. subject to periodic maintenance and replacement.
This research focuses on developing a model to accurately diagnose common types of damage on electric motors.. In particular, vibration signals corresponding to different operating conditions of the engine are measured by an acceleration sensor with a sampling frequency of 51.2 kHz. The collected vibration signals are then downsampled to 12.8 kHz and separated into many small signal segments with a length of 512 samples. The proposed model uses Wavelet transform to analyze these vibration signals on scalograms, described as above. Figure 1.
This obtained spectral image data set is then randomly divided according to the ratio: 80% is used for the training process and the remaining 20% is used to evaluate the damage classification model. These two data sets are then fed into the convolutional neural network model with Attention mechanism (CANN) for the training and evaluation process of the model. The CANN model is developed based on a convolutional neural network using the attention mechanism to extract information including spatial attention and channel attention from input data. The design of the model and the operating mechanism of the CANN model are presented in detail above Figure 2. Finally, the classification model is used to predict the errors of the collected data set during engine operation.
The results show that the proposed CANN model is capable of generating similar characteristics (attentions) for the same type of damage and at the same time generating different features (attentions) for each type of damage. This is clearly shown in Figure 3.
The experimental results indicate that the CANN model can achieve high classification accuracy up to 99.43%. This diagnostic result is much better than conventional CNN models and current deep learning models in diagnosing engine errors.
More information about the author:
Dr. Tran Minh Quang
EN