Deep Learning-Based Methods for Rolling Bearing Fault Diagnosis: Optimization and Comparison

Abstract

Rolling bearings are the vital support elements of industrial rotating machinery. Operational conditions have a direct impact on the safety, dependability and economical operation and maintenance of the production system. An effective and accurate fault diagnosis technology is very important for predictive maintenance and reducing the probability of downtime. With the development of deep learning and signal processing technologies, intelligent fault diagnosis based on vibration signals has become a research hotspot, and various technical methods and implementation schemes have been proposed. In this paper, three traditional deep-learning-based approaches for the diagnosis of rolling bearing faults are systematically studied, including the time-frequency image diagnosis method based on the ICEEMDAN decomposition integrated with the Hilbert transform and CBAM - ResNet, the FSCL diagnosis model combining FFT + SVD with CNN - LSTM, and the classification framework fusing the continuous wavelet transform signal-to-image conversion with SE-ResNet. This paper also discusses the technical frameworks and effectiveness of each approach. In this study, the inherent limitations of the three approaches were systematically studied from pre-processing efficiency, model complexity, end-user adaptability and generalization ability of scenarios, and relevant optimization strategies were put forward. The research results can be used as reference to promote the technology progress and industrial application of intelligent fault diagnosis for rolling bearings.