Multi-Scale Fusion Adversarial Domain Adaptation: A Cross-Domain Fault Diagnosis Method Based on Lightweight Model

In recent years, deep learning has achieved significant progress in cross-domain fault diagnosis. Currently, most existing models still rely on large-scale labeled data and complex architectures, limiting their practical deployment in industrial applications. To overcome these limitations, this study aims to develop a lightweight method capable of effectively mitigating the impact of cross-domain feature distribution discrepancies, thereby enhancing model generalization and applicability under complex working conditions. Firstly, a Multi-scale Fusion Adversarial Domain Adaptation (MFADA) approach is proposed. It improves inter-domain feature alignment by integrating local structural features and high-level semantic information from both the source and target domains. Secondly, a lightweight feature extraction module is designed by combining pointwise convolution (PW), depthwise convolution (DW), and max pooling (MaxPool). This structure achieves strong feature extraction capability with minimal parameters, significantly improving computational efficiency. The proposed method is validated on current signals from industrial robots and vibration signals from bearings under various complex operating conditions. Experimental results demonstrate that MFADA achieves excellent performance on multiple transfer tasks. It reaches a maximum diagnostic accuracy of 99.17%, while maintaining a compact model size and low computational cost, demonstrating excellent performance.