SeqAttention-Net: Design of a Deep Neural Network for Bearing Fault Detection Based on Small Sample Datasets

Abstract

As critical components of mechanical equipment, bearings play a vital role in ensuring the stability and safety of equipment operation. However, traditional fault diagnosis methods face challenges such as reliance on manual experience, difficulties in standardisation, and deficiencies in real-time performance and accuracy. In recent years, fault diagnosis methods that combine vibration analysis and artificial intelligence technology have gained increasing attention. In particular, deep learning methods have become a research focus in this area due to their excellent feature extraction capabilities. This paper presents a deep learning-based bearing fault diagnosis model, SeqAttention-Net, which aims to address the problem of bearing fault detection in small sample data sets. The SeqAttention-Net model overcomes the challenges of small sample sizes by combining sequence data transformation and attention mechanisms. The model pre-processes bearing vibration signals using Fast Fourier Transform (FFT) to extract key frequency features, effectively capturing the periodic changes in fault characteristics. In addition, the model incorporates white noise into the training set to simulate the complex noise environment in industrial production, which enhances the model’s generalisation ability and accuracy in detecting unknown samples. Experimental results show that SeqAttention-Net outperforms recent work in terms of accuracy, recall and F1 score. By integrating multi-head attention mechanisms and transformer encoder layers, the model effectively processes long-range dependencies and complex temporal relationships in sequence data, achieving accurate classification of bearing fault types.