An Explainable Spatio-Temporal Deep Learning Framework (ST- XFNet) for IoT Sensor Fault Detection and Classification in Electrical Systems

Abstract
Reliable fault detection and classification (FDC) is critical for the safe and uninterrupted
operation of modern electrical and Internet-of-Things (IoT) enabled industrial systems.
Traditional rule-based and signal-processing techniques struggle with the high-dimensional,
nonlinear and noisy data streams produced by distributed sensor networks, while
contemporary deep learning models, although accurate, suffer from limited interpretability,
sensitivity to noise and class imbalance. This paper proposes ST-XFNet, an explainable
spatio-temporal deep learning framework that integrates a one-dimensional Convolutional
Neural Network (1D-CNN) for spatial feature extraction, a Bidirectional Long Short-Term
Memory (BiLSTM) network for temporal dependency learning, an attention mechanism for
adaptive feature weighting and a SHapley Additive exPlanations (SHAP) module for post-hoc
interpretation. Synthetic bias, drift, random and polynomial-drift faults are injected into the
Intel Berkeley Research Lab IoT sensor dataset to construct a balanced multi-class
benchmark. Extensive experiments demonstrate that the proposed model attains a
classification accuracy of approximately 99%, an ROC-AUC of 0.996 and maintains over
97% accuracy under additive Gaussian noise, outperforming standalone CNN, LSTM and
CNN-LSTM baselines. SHAP analysis further reveals that temperature and voltage are the
most influential features for fault discrimination, providing actionable insight for
maintenance engineers. The results confirm that ST-XFNet offers an accurate, robust and
interpretable solution for proactive fault diagnosis in smart electrical infrastructures.