An ensemble deep learning network based on 2D convolutional neural network and 1D LSTM with self-attention for bearing fault diagnosis

Intelligent classification methods based on deep learning (DL) have become widely adopted for bearing fault diagnosis (BFD). However, it is acknowledged that relying on single feature extraction methods may not yield comprehensive representations of the information features. Additionally, DL-based approaches for extracting features from vibration signals typically utilize either one-dimensional (1D) or two-dimensional (2D) networks, which can restrict the network's ability to extract features effectively. In this paper, a time series data representation method called the relative angle matrix (RAM) method is firstly proposed. This method converts 1D time series into 2D images by calculating the angle differences between multiple vectors and a central vector, thereby extracting the hidden spatial features present in the original data. Then, this paper introduces an ensemble deep learning network called 1D2D-EDL, which integrates 1D-based and 2D-based DL mechanisms for feature extraction and classification, leveraging the strengths of each approach. The 1D2D-EDL comprises two channels: the 1D channel combines long short-term memory (LSTM) and multi-head self-attention (MSA) to process raw 1D time series data, facilitating feature extraction in both the time and frequency domains. Meanwhile, the 2D channel employs convolutional neural network (CNN) components to process 2D images for spatial feature extraction, which are derived from the original time series data using the RAM method. Finally, the feature information from these two channels is fused using a feature fusion method. To preliminarily validate the effectiveness of the RAM method, three competitive 2D conversion methods are employed, including Gramian angular difference field (GADF), Gramian angular sum field (GASF), and Markov transition field (MTF). These methods are applied alongside the proposed RAM method within the same CNN network for fault diagnosis testing. The results indicate that the RAM method significantly enhances the diagnostic accuracy of the CNN compared to the other 2D conversion methods. Furthermore, the bearing fault dataset from the University of Ottawa is utilized to validate the performance of the 1D2D-EDL. A comparative analysis with other DL methods using multiple statistical metrics demonstrates the superiority of the 1D2D-EDL. Specifically, when diagnosing faults under four different speed conditions, the 1D2D-EDL attains accuracy rates of 100 %, 99.33 %, 100 %, and 100 %, respectively. This study proposes the incorporation of a novel perspective classifier to enhance DL models for bearing fault diagnosis. The source code of RAM is available at https://ww2.mathworks.cn/matlabcentral/fileexchange/180197-relative-angle-matrix-ram.