Adaptive SNN Framework on FPGA for Multi-Modal Signal Processing in Industrial loT: Addressing Data Drift for Predictive Maintenance

The Industrial Internet of Things (IoT) drives demand for predictive maintenance (PDM) systems to process multi-modal signals-vibration, temperature, and acoustic-in real-time, preventing equipment failures. These modalities were chosen as they are prevalent in industrial settings, with vibration capturing mechanical dynamics, temperature revealing thermal trends, and acoustic signals detecting faults like grinding. Conventional deep learning, like CNNs and RNNs, faces challenges: high energy use (>۱۰W), excessive latency (>۵۰ ms), and data drift, where signal distributions shift due to sensor wear or environmental changes, degrading performance. We address this with an adaptive Spiking Neural Network (SNN) framework on Field-Programmable Gate Arrays (FPGAs). It integrates a multi-modal fusion layer to unify signals, a drift detection module using spike pattern divergence, and parallel neuron cores optimized for low power, leveraging Spike-Timing-Dependent Plasticity (STDP) for unsupervised adaptation. STDP was selected for its local, label-free weight updates, ideal for IoT's resource constraints, mimicking biological plasticity to adjust to drifts efficiently. Evaluated on bearing datasets with synthetic drift (e.g., Gaussian noise shifts), it achieves ۷۷% energy savings versus CNNs, ۴ ms latency, and ۰.۹۱ F۱-score under ۲۵% drift, outperforming static models. This advances neuromorphic computing for IoT PdM, with applications in biomedicine (e.g., ECG monitoring) and wearables (e.g., adaptive activity tracking). The framework's on-device processing enhances privacy, critical for sensitive industrial data. By detailing modality choices, STDP benefits, and performance metrics, this abstract clarifies the system's design and impact, ensuring robust, efficient PdM solutions for dynamic environments.

کلیدواژه ها:

Spiking Neural Networks ، Predictive Maintenance ، Multi-Modal Fusion ، Data Drift ، FPGA Optimization