Research on Fault Feature Extraction and Early Warning Method Based on MLP and Attention Mechanism CNN Fusion

Abstract

In modern industrial automation systems, fault feature extraction and early warning are the key technologies to ensure the stable operation of equipment. Traditional machine learning methods such as multi-layer perceptron often face the limitations of feature representation ability when dealing with such problems. In recent years, the attention mechanism combined with convolutional neural networks has become an effective way to improve the effect of feature extraction. CNN can effectively capture the spatial correlation in the image or signal by its local connection and weight sharing characteristics, and the attention mechanism can automatically focus on the most discriminative part among many features. The MLP is fused with CNN of attention matrix. Firstly, the original fault data is extracted by using CNN, and then the extracted features are weighted by attention module, emphasizing the most critical information for fault diagnosis. This fusion model not only inherits the nonlinear mapping ability of MLP, but also enhances the feature selection and representation ability of CNN in complex signal processing. Experiments show that the method can significantly improve the accuracy and robustness of fault feature extraction. Among 300 fault sample data, the S-network can correctly distinguish 295 fault types, and the early warning accuracy is more than 98%, which proves the effectiveness of the method. This study can achieve more effective early warning, reduce the cost of equipment maintenance, and improve the reliability of the system.