Online sensor fault diagnosis method based on time series prediction and independent component naive Bayesian classifier

Abstract The operational status of sensors is critical for effective system monitoring and control. To address the challenge of online diagnosis of sensor faults—including stuck, drift, bias, and their early weak faults—this paper proposes a hybrid method combining time series prediction and an Independent Component-based Naive Bayesian Classifier (Crossformer-ICNBC). First, the Crossformer separately establishes dependencies in the temporal and spatial dimensions using its unique two-stage attention mechanism, extracts multi-scale temporal features through a hierarchical encoding-decoding structure, enhancing the accuracy of sensor data prediction. The current sensor output is predicted by learning the sensor time series. Then, the predicted and fault values are subtracted to obtain the residual data. Finally, independent component analysis is applied to extract features from the residual data, and ICNBC is used to detect and classify the faults. To validate the effectiveness of the proposed method, it is evaluated against several prediction and classification models on two benchmark datasets: Intel Labs and Home Monitoring System. Experimental results demonstrate that ICNBC achieves high diagnostic accuracy while fully leveraging the predictive capabilities of Crossformer. The proposed method achieves diagnostic accuracies exceeding 95% and average F1 scores above 96%. Moreover, the diagnostic accuracy for early-stage weak faults related to drift and bias also exceeds 90%, indicating outstanding diagnostic performance.