关键词: 电机振动信号, 特征提取, 深度自编码器, 多尺度结构, 机械故障识别

Abstract: To address the challenge of extracting weak fault features and achieving high recognition accuracy in motor vibration signals,a multi-scale deep autoencoder model for adaptive feature compression and robust discriminative learning is proposed. The encoder incorporates parallel convolutional channels with varying receptive fields to capture local dynamics across different time scales. Meanwhile,sparsity and discriminative constraints are embedded in the latent space to enhance fault sensitivity. The decoder integrates an attention mechanism to improve reconstruction accuracy and focus on subtle anomalies. Simulation experiments show that the proposed model achieves an average accuracy of 97. 86% and an F1-score
of 97. 42%,representing an improvement of 3. 47% ~ 14. 55% over traditional time-frequency methods and mainstream deep models. It also maintains superior robustness and stability under varying noise levels and sampling strategies,validating its practical potential and engineering applicability in intelligent fault diagnosis of industrial vibration systems.