LightM-ConvKNet: An Efficient Deep Learning Framework for Multisensor Industrial Fault Diagnosis Under Cross-Operating Conditions

In practical industry, accurate fault diagnosis of industrial equipment is crucial. Convolutional Neural Networks (CNN) have shown strong performance in local feature representation as an important tool for industrial fault diagnosis. However, the local perception characteristics of convolutional kernels limit their ability in capturing the global contextual information of multiple sensor data and fail to fully reveal the local dependencies among sensor data with distant topological structures. To address these challenges, this paper proposes a lightweight multi-sensor industrial fault diagnosis framework named as LightM-ConvKNet. By constructing a 3-D spatiotemporal synergistic tensor, the method captures the temporal nature of signal data while enriching the topological structure of multi-sensor data. Besides, a Tanh-activated Kolmogorov-Arnold Network with Drop mechanism (KAN-TD) is designed as the classification layer. Hence, the proposed model can effectively extract multi-level fault features and achieves high-accuracy diagnosis while maintaining lightweight deployment. The effectiveness and superiority of the proposed method are validated through two experiments.