Thresholding Convolutional Neural Network for Fault Diagnosis

ABSTRACT Mechanical fault diagnosis plays an important role in ensuring the safety and reliability of industrial equipment, especially under noisy operating conditions. Convolutional neural network (CNN) has been applied to mechanical fault diagnosis in recent years. Most existing approaches focus on extracting features under normal operating conditions; however, in real‐world industrial environments, sensor signals are often contaminated by noise, which may hinder the feature extraction capability of CNN models. To solve the above problem, this paper proposes a novel thresholding CNN model (TH‐CNN), which is composed of an optimized soft thresholding function (gated thresholding function, G‐T) and an improved model based on traditional CNN. The improved CNN, whose structure is very simple and does not involve overly complicated operations, is used to extract data features; G‐T serves as a nonlinear mapping layer in the model, which retains important feature information through a scale factor based on its importance, thereby achieving data denoising. In experiments, various types of noise are introduced to the Case Western Reserve University and Paderborn University bearing data sets to simulate signal disturbances commonly encountered in practical scenarios. Results show that the proposed TH‐CNN achieves higher test accuracy and good ability of generalization compared to other models.