Signals hierarchical feature enhancement method for CNN-based fault diagnosis

Abstract

The high noise and low energy characteristics of the raw signals collected by sensors make the signal features weak and difficult to train. The purpose of this paper is to enhance the fault features of abnormal signals using the hierarchical feature enhancement method (HFE) which contains three layers. In the first layer, the signals are decomposed into multiple modals estimated by a variational optimization problem. The modals we choose are used to reconstruct the signals to form a complex matrix used to extract features in the second layer. In the third layer, the feature signals are converted into two-dimensional space and then are input into the convolutional neural network (CNN) for fault diagnosis after HFE since CNN helps to mine deeper features and compute in parallel on a large scale. The experimental results effectively verify the performance of the HFE for enhancing the weak fault features and preventing noise interference. The signals analyzed by HFE used as input greatly improve the diagnosis ability of CNN. In addition, the ablation and comparison experiments are conducted which still show superiority.

Research on the mechanical fault diagnosis method based on sound signal and IEMD-DDCNN

“…Single-input dual-channel CNN (DCNN) (Zhang et al, 2021), as shown in Figure 3 below, the model has two channels with convolution and pooling layers, but only one input. When data is input, the model is equivalent to copying the input data into two parts and transmitting them to each channel for convolution and pooling.…”