Transfer-Based Deep Neural Network for Fault Diagnosis of New Energy Vehicles

Yuping Wang,Weidong Li

doi:10.3389/fenrg.2021.796528

Abstract

New energy vehicles are crucial for low carbon applications of renewable energy and energy storage, while effective fault diagnostics of their rolling bearings is vital to ensure the vehicle’s safe and effective operations. To achieve satisfactory rolling bearing fault diagnosis of the new energy vehicle, a transfer-based deep neural network (DNN-TL) is proposed in this study by combining the benefits of both deep learning (DL) and transfer learning (TL). Specifically, by first constructing the convolutional neural networks (CNNs) and long short-term memory (LSTM) to preprocess vibration signals of new energy vehicles, the fault-related preliminary features could be extracted efficiently. Then, a grid search method called step heapsort is designed to optimize the hyperparameters of the constructed model. Afterward, both feature-based and model-based TLs are developed for the fault condition classifications transfer. Illustrative results show that the proposed DNN-TL method is able to recognize different faults accurately and robustly. Besides, the training time is significantly reduced to only 18s, while the accuracy is still over 95%. Due to the data-driven nature, the proposed DNN-TL could be applied to diagnose faults of new energy vehicles, further benefitting low carbon energy applications.

Highlights

New energy vehicles such as the electrical vehicle and hybrid electrical vehicle play a vital role in achieving low carbon industrial and energy economy, where the rolling bearing is a key component within new energy vehicles
To observe a better pre-training model in rolling bearing fault diagnosis of new energy vehicles, this study proposes DCNNL by combining convolutional neural networks (CNNs) and long short-term memory (LSTM) for pre-training, as illustrated in first, after adding batch normalization (Szegedy et al, 2017b) between the convolutional layer and the pooling layer, the input would be pulled into the convolutional layer back forcibly to the standard normal distribution
To evaluate the effectiveness of the Deep neural networks (DNNs)-transfer learning (TL) method in the diagnosis of new energy vehicles faults, the data set of Case Western Reserve University (CWRU) and the rolling bearing data set of the laboratory are utilized

Summary

INTRODUCTION

New energy vehicles such as the electrical vehicle and hybrid electrical vehicle play a vital role in achieving low carbon industrial and energy economy, where the rolling bearing is a key component within new energy vehicles. According to the abovementioned discussion, TL methods could well benefit the computational efficiency and diagnostic accuracy, which is promising to be used in the rolling bearing fault diagnosis of new energy vehicles. A novel data-driven method named the transfer-based deep neural network (DNN-TL) through integrating CNN, LSTM, and transfer learning is designed in this study. The logic of designing the DNN-TL method is detailed below: First, CNNs and LSTM are designed to extract fault-related features from the signals on a rolling bearing of new energy vehicles. The number of neurons in the layers, the size of convolution kernel, and the fully connected layer are obtained through the step heapsort method. Where Fcnn is the comprehensive evaluation index for single CNN model training, while Flstm is the comprehensive evaluation index for single LSTM model training

Methods name

Method name

EXPERIMENTS AND RESULTS

Experimental Results

CONCLUSION