Di-fedDIL: Distributed fault diagnosis of idler bearings of long-distance belt conveyor under cloud-edge resource constraints

In long-distance belt conveyor systems, idler bearings are widely distributed along the transportation route, making distributed fault diagnosis necessary. A cloud-edge architecture provides a feasible solution by enabling local inference at edge devices and collaborative model training through the cloud. However, distributed idler-bearing diagnosis still faces two major challenges: (1) heterogeneous sensor data across different idlers and time-varying fault features under evolving operating conditions and (2) limited edge computing and communication resources. To address these challenges, this study proposes Di-FedDIL, a distributed fault diagnosis framework for idler bearings under cloud-edge collaboration. The framework consists of the dilated separable attention convolution network (DiSAC-Net), federated learning via contrastive alignment and progressive Euclidean regularization (FedCAP), and dual-anchor incremental learning (DIL). For edge diagnosis, DiSAC-Net performs lightweight feature extraction and online fault inference at edge devices. For federated collaborative learning, FedCAP enables distributed clients to collaboratively learn shared diagnostic knowledge while adapting to heterogeneous local data. For incremental updating, DIL continuously adapts personalized models to evolving operating conditions while balancing new knowledge acquisition and historical knowledge retention. In addition, a hierarchical greedy optimization strategy jointly performs task offloading and bandwidth allocation to reduce diagnostic response time under cloud-edge resource constraints. Experimental results show that Di-FedDIL outperforms the compared federated learning methods. The proposed resource-allocation strategy reduces task completion time by approximately 25% compared with fully local computing. These results demonstrate that Di-FedDIL provides a practical solution for distributed idler-bearing fault diagnosis while improving execution efficiency under limited cloud-edge computing and communication resources.