To address the challenges of high structural noise, unstable operating conditions, and susceptibility to single-channel failure in multi-sensor monitoring of industrial robot transmission components, this paper proposes a cross-machines domain generalized fault diagnosis method based on fault intrinsic representation and channel self-healing. This method aims to extract essential fault characteristics from distinctive public datasets and transfer them to complex industrial robot target scenarios. First, considering the characteristics of single-point multi-directional monitoring of robot joints, a dual-channel vibration feature extraction framework is constructed. Information enhancement and splicing are used to generate intra-domain and cross-domain joint features. Combined with a set-level class-prototype regularized mechanism, the distribution differences specific to the operating conditions are decoupled and eliminated from multi-source domain data, extracting the common fault intrinsic representations across datasets. Second, to solve the monitoring blind spot problem caused by single sensor failure, a semantically supervised channel self-heal module is designed. This module uses the feature distribution of intact channels to deduce the semantic information of missing channels, achieving signal self-healing and complementation during the testing phase. Finally, rigorous cross-machines transfer experiments are designed using three public datasets containing basic fault characteristics as source domains. Zero-shot tests are conducted on the gearbox dataset characterized by high structural noise and the bearing dataset featuring extreme non-stationary start-stop conditions. Experimental results demonstrate that the proposed method effectively overcomes cross-machines distribution shifts and maintains high-precision fault diagnosis, even under extreme conditions of complete single-channel sensor failure, verifying the feasibility and robustness of transferring laboratory data to complex robot application scenarios.
Yao et al. (Fri,) studied this question.