Rolling Bearing Fault Diagnosis Method Based on Parallel QPSO-BPNN Under Spark-GPU Platform

Abstract

Facing the massive rolling bearing vibration data, how to improve the training efficiency, diagnosis efficiency, and diagnosis accuracy of the rolling bearing fault diagnosis model is a challenge. Considering that the Spark-GPU platform provides powerful distributed parallel computing capabilities and back propagation neural network (BPNN) optimized by quantum particle swarm optimization (QPSO) algorithm has the characteristics of low computational complexity and high diagnosis accuracy, a rolling bearing fault diagnosis method based on parallel QPSO-BPNN under Spark-GPU platform is proposed. First, the distributed parallelization of QPSO-BPNN model based on Spark-GPU platform is realized, which can improve the training efficiency and diagnosis efficiency of rolling bearing fault diagnosis model in the big data environment. Second, in order to improve the convergence speed of fault diagnosis model, a parameter update strategy suitable for the distributed parallel training of QPSO-BPNN model is designed. At each iteration during training, the local parameters of each worker node are collected to the master node, and the global parameters are updated according to the weights and synchronized to each worker node. Third, a combination strategy of multiple QPSO-BPNN models based on ensemble learning is proposed. The weighted voting method is adopted to combine the output results of different QPSO-BPNN models to obtain the best fault diagnosis result of a sample, which can improve the fault diagnosis accuracy to a certain extent. Experimental results show that the proposed method can quickly perform model training and fault diagnosis for large-scale rolling bearing vibration data, and the fault diagnosis accuracy reaches 98.73%.

Research on a Small-Sample Fault Diagnosis Method for UAV Engines Based on an MSSST and ACS-BPNN Optimized Deep Convolutional Network

“…Replacing the actual training process of the convolutional network with the network prediction process, the BPNN network should fulfill the following two requirements:(1) Simple structure and fast training speed; (2) Fast prediction speed and high accuracy. As a typical three-layer neural network, BPNN consists of only input, hidden and output layers, with a relatively simple structure [26,27], and has a strong prediction ability for nonlinear complex situations.…”

Intelligent fault diagnosis of hydroelectric units based on radar maps and improved GoogleNet by depthwise separate convolution

“…In this paper, the 'Adam' optimization algorithm is chosen to update the network parameters. The parameter update formula is shown in equations (7) and (8):…”

“…The extracted features are then inputted into the diagnostic model to accomplish accurate fault classification [3][4][5]. Machine learning algorithms commonly used for fault diagnosis in the above process include Bayesian networks, extreme learning machines, backpropagation neural network (BPNN), support vector machines (SVMs), random forest algorithms, and K-nearest neighbor algorithms [6][7][8][9][10][11]. However, due to the shallow network structure of the above algorithms, the feature extraction capability of the relevant algorithms needs to be improved to mine and extract the deeper microscopic features contained in the fault data [12].…”

GIS Partial Discharge Pattern Recognition Based on IWOA-SVM Algorithm

“…Traditional methods of PD recognition are based on years of expert experience, practice, and other relevant knowledge, which are too subjective and have significant limitations. Currently, researchers have proposed some artificial intelligence algorithms to improve the precision of PD recognition such as artificial neural network (ANN) [4][5], back-propagation neural network (BPNN) [6][7], support vector machine (SVM) [8] and deep learning algorithm [9][10], etc. However, due to GIS PD online detection technology still needing to be optimized, and measured data collection is relatively difficult, resulting in the lack of sample data available for PD recognition.…”