SmartFault-M: A Hybrid CNN–LSTM with Attention for Multi-Domain Fault Diagnosis in Industrial Mechatronic Systems

Abstract

Diagnosing faults in industrial mechatronics can be difficult because the systems involve intricate interactions among mechanical, electrical, and control sub-systems which produce multidimensional, non-stationary data from various sensors. Traditional fault diagnostics models based on model-based or data-driven approaches in single domains cannot provide adequate performance because they cannot effectively capture the multi-scale and temporal nature of faults. Therefore, in this paper, a novel approach called Smart Fault Monitoring (SmartFault-M), a hybrid model based on a dual-stream CNN–LSTM framework for smart fault diagnosis, is proposed, in which two parallel streams with time-domain and frequency-domain feature maps extract informative features and attention-based learning captures important patterns. In SmartFault-M, CNN and LSTM are used to learn spatial and temporal information from the data, respectively, while attention learning emphasizes the informative features. For validating the effectiveness of our proposed approach, we perform extensive experiments using the bearing dataset of Case Western Reserve University (CWRU) under different fault states. As experimental results show, the accuracy of the proposed SmartFault-M reaches 98.25% and outperforms baseline methods like CNN (91.42%), LSTM (92.36%), and CNN–LSTM (95.18%) models. In addition, the model shows consistently high precision, recall, and F1-score for all classes of faults. From these experiments, it is obvious that our proposed model with multi-domain feature extraction and attention-based learning achieves superior performance for smart fault diagnosis tasks.