A lightweight end-to-end fault diagnosis model for IoT sensors

Abstract Against the broader backdrop of Industry 4.0 and the rapid development of the Internet of Things (IoT), sensor faults can easily trigger safety risks. Therefore, efficient and intelligent sensor fault diagnosis methods have become an urgent need. Nevertheless, existing sensor fault diagnosis approaches still have shortcomings: deep learning models often involve a large number of parameters, adapt poorly to imbalanced samples, and lack interpretability. To address these issues, this paper proposes a lightweight end-to-end LaplaceConv-LSTM sensor fault diagnosis model. The proposed model directly maps sensor signals to the fault-category space. It adopts a three-level hierarchical feature-extraction architecture that jointly achieves robustness to data imbalance, lightweight design, and high diagnostic accuracy. The LaplaceConv layer performs multi-scale feature extraction of fault signals; its physically informed kernel generation improves robustness to data imbalance and enhances interpretability. Depthwise separable convolution is used for lightweight deep feature extraction. An LSTM layer is adopted for deep time-series modeling to capture global features, and a fully connected layer finally completes the classification. Compared with mainstream lightweight convolutional neural network architectures MobileNetV3 and ShuffleNetV2, the proposed model reduces the model size by 70.17% and 79.46%, respectively, and decreases FLOPs by 34.02% and 39.09%, further improving lightweight performance and diagnostic efficiency. Experimental results on the mud tank radar level and humidity sensor datasets show that under imbalanced data conditions, the model achieves F1 scores of 0.9343 and 0.9476 (improving by 7.83% and 15.08% over the baseline CNN, respectively), and mAP values of 0.9693 and 0.9775 (improving by 3.3% and 10.11% over the baseline CNN, respectively). In addition, the proposed model also outperforms MobileNetV3- and ShuffleNetV2-based models, validating its robustness to data imbalance and its potential for edge deployment. Therefore, the model provides an efficient and high-accuracy solution for IoT sensor fault diagnosis.