Lightweight Semi-supervised Domain Adaptation for Bearing Fault Diagnosis on Edge Devices

Abstract

Rolling bearings serve as critical components in rotating machinery, but practical fault diagnosis is affected by operating-condition changes, limited fault samples, and the restricted computing resources of edge devices. This paper proposes LiteDANN, a lightweight semi-supervised domain-adaptation framework for cross-condition bearing fault diagnosis. Drive-end vibration signals from the Case Western Reserve University bearing dataset are segmented and converted into 64 × 64 short-time Fourier transform spectrograms. A MobileNetV2-style backbone extracts compact features, while a gradient reversal layer and a domain classifier reduce the discrepancy between different operating conditions. Focal loss and weighted random sampling are used to improve learning from scarce and imbalanced target-domain fault samples. The 0 hp condition is used as the source domain, the noisy 3 hp condition is used as the target domain, and the unseen 2 hp condition is reserved for testing. The recorded experiment achieves 99.69% accuracy and 99.70% macro-F1 on the unseen test condition. The model contains 37,478 trainable parameters. After ONNX conversion and INT8 quantization, the model size is reduced to 0.10 MB, enabling fast real-time inference on resource-limited edge hardware. These results indicate that LiteDANN provides a practical basis for real-time bearing diagnosis on resource-constrained edge platforms.