Neuromorphic Parameter Estimation for Power Converter Health Monitoring Using Spiking Neural Networks

Power electronic systems play a crucial role in modern industrial and consumer electronics, particularly in power converter applications. With the advancement of digital twin technology, physics-informed neural networks (PINNs) have been widely used for online parameter identification and predictive maintenance of power converters. However, PINNs typically rely on GPU or cloud accelerators for floating-point multiply-accumulate operations, which limits their application on edge devices in terms of energy consumption and real-time performance. Recently, spiking neural networks (SNNs), as a biologically inspired computational model, have gained attention for their low-energy advantages on neuromorphic hardware.

Health monitoring of power converters requires always-on inference on energy-constrained edge devices, which is challenging for traditional GPU-based PINNs. Specifically, PINN inference requires floating-point multiply-accumulate operations on a GPU or cloud accelerator, making always-on, edge-located condition monitoring impractical for cost- and power-constrained converter systems.

The core innovation of this paper lies in proposing an architecture combining spiking neural networks (SNN) and a differentiable ordinary differential equation (ODE) solver for parameter estimation in power converters. • The method separates spiking temporal processing from physics enforcement, using a three-layer leaky integrate-and-fire (LIF) neuron to estimate passive component parameters. • The ODE solver provides physics-consistent training by decoupling the ODE physics loss from the unrolled spiking loop, addressing computational complexity and numerical stability.

• �� SNN Estimator: Processes the noisy waveform temporally and outputs three scalar parameters (inductance, capacitance, and lumped resistance). • Differentiable ODE Solver: Takes the estimated parameters and integrates the ODEs to produce predicted waveforms. • Reconstruction Loss: Computes the mean squared error (MSE) between predicted and measured waveforms. • Gradients flow from the reconstruction loss through the ODE solver into the SNN weights via surrogate gradients, enabling standard backpropagation.

Experiments are conducted on a synchronous buck converter benchmark using true parameters for ODE simulation, with structured EMI noise added. The SNN estimator uses three LIF hidden layers and is trained using the Adam optimizer. Results show that the SNN outperforms traditional feedforward networks in estimating lumped resistance error and offers significant energy advantages.

Experimental results show that the SNN reduces lumped resistance error from 25.8% to 10.2% on an EMI-corrupted synchronous buck converter benchmark, within the ±10% manufacturing tolerance of passive components. Additionally, the SNN achieves 93% spike sparsity on neuromorphic hardware, with an estimated ~270x energy reduction. Persistent membrane states enable degradation tracking and event-driven fault detection via a +5.5 percentage-point spike-rate jump at abrupt faults.

This method is suitable for power converter health monitoring on resource-constrained edge devices, particularly in scenarios requiring always-on monitoring. By achieving low-energy inference on neuromorphic hardware, the architecture offers new possibilities for digital twins and predictive maintenance in power electronics.

Despite significant advantages in energy consumption and real-time performance, surrogate gradient noise in spiking neural networks during training may lead to oscillations in parameter estimation. Further research is needed for statistical validation under diverse operating conditions and noise realizations. Future directions include multi-trial statistical validation and hardware deployment on BrainChip Akida or Intel Loihi 2 for measured power and latency characterization.