Smart grid reliability hinges on catching equipment degradation before a catastrophic failure occurs. While traditional predictive maintenance often relies on a single data stream—such as vibration analysis for mechanical dynamics or thermal imaging for heat dissipation—single-modality systems have clear blind spots. Vibration sensors miss thermal degradation, while non-contact thermal monitoring cannot capture mechanical resonance. To overcome these limitations, a research team developed a unified deep learning framework that fuses heterogeneous vibration and thermal data using an attention-based sensor fusion architecture.
Key Architectural Highlights
Vibration Feature Extraction: A 1D Convolutional Neural Network (CNN) extracts local temporal patterns, which are then passed through a Bidirectional Long Short-Term Memory (BiLSTM) network to capture long-range dependencies.
Thermal Feature Extraction: A Temporal Convolutional Network (TCN) processes multivariate thermal and environmental time-series data to capture broad temporal contexts efficiently via dilated convolutions, followed by Global Average Pooling.
Attention-Based Multimodal Fusion: Rather than simply concatenating the inputs, an attention layer dynamically weights the vibration and thermal embeddings based on which modality provides the most critical fault signatures for a given operational state. Experimental Results & Performance Evaluated across combined industrial time-series datasets, the multimodal fusion model significantly outperformed traditional machine learning baselines and single-modality deep learning models.
Overall Classification Performance Accuracy: 96.74% (vs. 91.85% for vibration-only LSTM and 92.11% for thermal-only TCN) Precision: 95.89% Recall: 95.34% F1-Score: 95.61%
reference:
M. R. Al Amin Khan, M. O. Faruq, M. N. Suhaimee, F. Shakil, S. R. Sultana and S. Firdaus, "AI-Driven Multimodal Sensor Fusion for Predictive Maintenance of Smart Grid Equipment Using Vibration and Thermal Data," 2025 IEEE 2nd International Conference on Computing, Applications and Systems (COMPAS), Kushtia, Bangladesh, 2025, pp. 1-6, doi: 10.1109/COMPAS67506.2025.11381618.
keywords: {Vibrations;Deep learning;Accuracy;Torque;Multimodal sensors;Fault detection;Smart grids;Convolutional neural networks;Monitoring;Predictive maintenance;Predictive maintenance;smart grid equipment;multimodal sensor fusion;vibration analysis;thermal data;deep learning;temporal convolutional networks;attention mechanism;equipment health monitoring;industrial IoT},
Top comments (0)