Federated fault diagnosis method for collaborative self-diagnosis and cross-robot peer diagnosis

In multi-robot collaboration, individual failures can propagate to other robots due to the topological coupling between them. Existing fault diagnosis models are designed for single robots and fail to meet the practical requirements of multi-robot scenarios. To address this, this study develops a federated learning-based fault self-diagnosis model for individual robots and a multi-robot mutual diagnosis model that accounts for group behavior consistency. This approach effectively isolates faulty robots in multi-robot systems. Initially, each robot's local data is encoded using the Gramian Angular Field (GAF) to generate two-dimensional time-frequency plots, creating local fault datasets. Next, a federated learning framework is established, where fault models for different robots are pre-trained using the local fault datasets. The local model parameters from multiple robots are then aggregated for shared learning, mitigating the potential knowledge shift during individual robot training. Finally, a multi-robot mutual diagnosis model is developed, incorporating group speed and direction consistency to ensure fault diagnosis based on behavioral coherence. Experimental results demonstrate that the proposed self-diagnosis model accurately identifies faults in individual robot components, while the mutual diagnosis model effectively recognizes system-wide faults.