Jointly optimized semi-supervised graph embedding stochastic configuration networks for industrial fault diagnosis

Abstract We address industrial fault diagnosis under scarce labels and strict training-time constraints, a setting where many existing methods implicitly assume abundant annotations. We propose the Jointly Optimized Semi-Supervised Graph-Embedded Stochastic Configuration Network (JOSGESCN). The method constructs geometric similarity between labeled and unlabeled samples and injects this structure into the model via manifold regularization, enabling effective use of unlabeled data. We further introduce a supervisory mechanism that adaptively configures hidden nodes to strengthen representation capacity. A joint optimization strategy reduces error propagation from one-shot pseudo-labeling and encourages low-density separation of the decision boundary. We evaluate JOSGESCN on three industrial fault-diagnosis datasets against SVM, RVFL, SCN, SSRVFL, JOSRVFL, and LPSCN. Across all datasets, JOSGESCN delivers superior accuracy, demonstrating consistent gains over strong baselines while maintaining practical training efficiency. These results indicate that coupling graph-based manifold priors with adaptive stochastic configuration is a promising direction for label-efficient, high-performance industrial health monitoring.