Session: 01-08: Nuclear Plant Operation, Modification, Life Extension, Maintenance and Life Cycle - VIII

Paper Number: 135724

135724 - Intelligent Fault Diagnosis of Rotating Machinery Based on Deep Neural Network 

Abstract: 

In order to improve the fault diagnosis for rotating machinery in nuclear power plants, this paper proposes a fault diagnosis model based on short-time Fourier transform (STFT) and Residual Network (ResNet), using information from multiple sensors. Firstly, acceleration sensors are installed on the fan end, the drive end and the base of the motor respectively to measure vibration signals, and the normal condition and eight typical faults (rotor unbalance, rotor eccentricity, rotor bending, bearing fault, rotor broken bar, winding short circuit, phase missing, voltage unbalance) of the motor are simulated by the mechanical fault simulation test platform. Then, sensor signals from different positions of the motor are obtained as inputs to the fault diagnosis model to reduce the impact of the fault feature attenuation during mechanical propagation. In data preprocessing, STFT is used to convert the multi-source signals of the motor from time domain to time-frequency domain. After preprocessing, the number of is expanded, and the fault characteristics of the vibration signals are strengthened in the time-frequency domain. Finally, ResNet is used to further extract advanced features from time-frequency feature signals and identify fault patterns. Results obtained show that the proposed model based on STFT, ResNet and multi-sensor information strategy can accurately classify the fault modes, and comparisons with the traditional convolutional network model and single sensor strategy model highlight the advantages of the proposed method in rate of convergence, accuracy and noise robustness.

The parameter situation is as follows: The ResNet used consists of 1 convolutional layer, 2 pooling layers, 8 residual blocks, and a Softmax classification layer architecture. The Hamming window is selected as the window function of STFT, with a window length of 512 and an overlap number of 384.Local frequency domain features extracted from 8 consecutive window functions are selected to form a 257*8 time-frequency spectrum. Perform STFT processing on the signals of 3 sensors and stack them to obtain a size of 257*8*3 ResNet input samples. After STFT, the number of samples in each fault state is expanded to 900. After standardizing all samples, randomly divide the training and testing sets in a 7:3 ratio, and train and test the ResNet model.

 

The preliminary results are as follows: After the 15th iteration, the accuracy of ResNet has remained stable at 99.99%, while the accuracy of CNN has fluctuated slightly around 99.90%. At the end of training, the accuracy of ResNet is 99.99%, and the magnitude of the loss function value reached 10^-6, both of which were better than the training results of CNN.

Presenting Author: Xiuchun Zhang Harbin Engineering University

Presenting Author Biography: Zhang Xiuchun (1979-), female, on duty doctoral student, mainly engaged in research on
intelligent operation technology of nuclear power.

Authors:

Xiuchun Zhang Harbin Engineering University
Hong Xia Harbin Engineering University
Yongkang Liu China Nuclear Power Technology Research Institute Co., LTD.
Shaomin Zhu Harbin Engineering University
Yingying Jiang Harbin Engineering University
Jiyu Zhang Harbin Engineering University
Jie Liu Suzhou Nuclear Power Research Institute Co.,Ltd.
Wenzhe Yin Harbin Engineering University