Real-time anomaly detection using hardware-based unsupervised Spiking Neural Network (TinySNN)

We present TinySNN, a novel unsupervised spiking neural network hardware designed for real-time anomaly detection. TinySNN provides an energy-efficient edge computing solution for detecting anomalies in industrial settings. TinySNN can model and extract spatio-temporal features in sensory data, thereby enabling it to identify anomalous inputs with high accuracy. During inference, the spike rate of the TinySNN model acts as an indicator of anomalous patterns within these features. We demonstrate TinySNN's potential using publicly available vibration datasets, achieving impressive anomaly detection results. TinySNN demonstrates exceptional sensitivity to subtle deviations from normal operation, and can dynamically adapt during online, unsupervised training. We provide a digital implementation of TinySNN on an FPGA for hardware efficiency. The TinySNN hardware can be trained online on real industrial data without requiring power hungry computing architectures like GPUs. This approach presents a promising and practical solution that allows for the dynamic learning of normal industrial operations and assists in mitigating risks through continuous monitoring. TinySNN can help ensure the safe and reliable operation of critical industrial systems through neuromorphic processing.