DeepSMOTE with Laplacian matrix decomposition for imbalance instance fault diagnosis

In industrial environments, the unpredictability and irreproducibility of faults often result in insufficient sample sizes and atypical data features, significantly increasing the challenges faced by traditional fault diagnosis methods. To address these issues, this paper proposes a novel fault diagnosis approach that integrates the Borderline embedded deep synthetic minority oversampling technique (BE-DeepSMOTE) with Laplacian matrix decomposition, with the aim of tackling fault identification problems in imbalanced data scenarios. BE-DeepSMOTE employs a deep encoder–decoder framework to enable end-to-end learning and reconstruction of multi-dimensional features. It further incorporates the Borderline SMOTE technique to oversample minority class instances in the feature space, thereby enhancing their representation while ensuring statistical consistency with the original dataset to mitigate data imbalance. Furthermore, we introduce an ensemble classifier that combines Adaboost with Laplacian matrix decomposition. This ensemble classifier leverages the synergy of multiple weak classifiers to extract geometric properties and graph structure similarities from the data, while employing an adaptive weighting mechanism to improve the diagnostic accuracy. Experimental results from two industrial processes demonstrate that the proposed approach significantly enhances the diagnostic accuracy and stability in imbalanced instance environments.