Spiking neural network cuts inverter fault diagnosis energy 382x
11 microjoules per diagnosis with 100% accuracy – a new neuromorphic approach.
Fault diagnosis in three-phase inverters must fit within a sub-watt power budget, but conventional CNN-based methods demand dense multiply-accumulate operations, consuming too much energy. Researchers from Xiaoyi Lei's team propose converting a CNN trained on current-vector trajectory matrices into a spiking neural network (SNN) evaluated using the NengoLoihi framework. The SNN exploits the sparse structure of trajectory matrices, activating computation only in informative regions instead of processing the full feature map densely. This event-driven approach minimizes inference energy.
Experiments on a three-phase inverter platform show the SNN achieves 11 microjoules per diagnosis – a 382× reduction compared to a GPU-based CNN – while delivering 100% diagnostic accuracy. Robustness is validated under unbalanced loading, current amplitude step changes, and injected measurement noise. The work demonstrates that neuromorphic hardware can enable ultra-low-power, real-time fault detection without sacrificing accuracy, promising practical deployment in energy-constrained converter control systems.
- CNN converted to spiking neural network (SNN) for event-driven computation on Loihi hardware.
- 11 microjoules per diagnosis – 382× energy reduction vs. GPU-based CNN.
- 100% diagnostic accuracy maintained under unbalanced loads, amplitude steps, and noise.
Why It Matters
Enables sub-watt, real-time fault detection for industrial three-phase inverters, slashing energy costs while maintaining perfect accuracy.