Bearing fault detection using machine learning on vibration and sound signals

Abstract

A bearing is a machine element that constrains relative motion to only the desired motion and reduces friction between moving parts, especially in high-speed trains, wind turbines, and railway wheelset applications. As the complexity of equipment and the harshness of operational environments in these industries increase, bearings, as critical components, are prone to faults that can lead to significant damage and costly maintenance. Therefore, accurately and in real-time detecting these faults is crucial for ensuring the safe and efficient operation of equipment. This research explores modern signal processing techniques and machine learning algorithms, integrating multimodal data such as vibration and sound, to propose efficient fault diagnosis models, providing valuable insights for future industrial intelligent maintenance systems. Firstly, for fault diagnosis in high-speed train rolling bearings, this thesis introduces a model based on frequency domain feature extraction and a Bidirectional Long Short-Term Memory (Bi-LSTM) network. High-speed trains, operating under long periods of dynamic load and high-frequency vibrations, are prone to bearing fatigue damage, making early fault detection essential for safety. Traditional fault detection methods mainly focus on time-domain signal analysis, which often fails to capture key frequency domain features. To address this, the study first transforms time-domain signals into the frequency domain using Fast Fourier Transform (FFT), followed by fault classification using the Bi-LSTM model to identify fault types from both directions of sequential data. Experimental results demonstrate that the Bi-LSTM model based on frequency domain signals significantly improves fault classification accuracy compared to traditional methods, validating its effectiveness and superiority. Secondly, a real-time monitoring system for wind turbine bearing fault detection is designed and implemented, utilizing a simplified neural network model running on a Raspberry Pi. Wind turbine bearings are subjected to high loads and harsh environmental conditions, and bearing faults can lead to turbine shutdowns, affecting energy production and increasing maintenance costs. To achieve real-time monitoring and fault prediction, a simplified neural network algorithm is designed, efficiently implemented on a Raspberry Pi. By segmenting sensor data, the system quickly analyzes each data segment and delivers predictions within milliseconds. This model demonstrates high fault detection efficiency and response speed, significantly reducing maintenance costs and improving operational efficiency. Finally, for railway wheelset bearings, this thesis proposes a new fault detection method based on GreyWolf Optimizer (GWO) and Support Vector Machine (SVM). Wheelset bearings in railway vehicles endure complex loads and environmental pressures, making fault detection critical. Traditional single-source vibration signal analysis often falls short of capturing comprehensive fault characteristics. The study innovatively integrates vibration and sound data, providing a richer set of diagnostic information. The GWO is employed to optimize the SVM model’s hyperparameters, resulting in a highly efficient fault detection method. The experimental results show that the proposed approach significantly improves fault detection performance compared to traditional models, achieving rapid and accurate predictions, thereby minimizing downtime and maintenance costs. Overall, this research demonstrates the potential of various machine learning and optimization algorithms in bearing fault detection across different industrial applications. By integrating frequency domain signal processing, neural networks, and optimization algorithms, the thesis overcomes the limitations of traditional time-domain methods, significantly enhancing the accuracy and real-time capabilities of fault diagnosis. Specifically, the high-speed train bearing fault diagnosis method, based on FFT and Bi-LSTM, showcases superior performance in handling complex frequency domain signals. The real-time monitoring system for wind turbine bearings leverages lightweight neural networks on embedded devices, offering high efficiency and low latency. The vibration-sound fusion method for wheelset bearings, optimized by GWO, achieves better detection precision and speed. Future research could explore the integration of additional sensor modalities and combine deep learning algorithms with optimization techniques to further improve diagnostic accuracy and application breadth. Overall, the three methods proposed in this thesis offer effective solutions for intelligent maintenance systems and provide essential technological support for the safe operation and maintenance of industrial equipment in the future.