Research on Health Assessment Technology for TwoStage SMPS Aging Testbed Based on KolmogorovArnold Networks

As a critical energy conversion device, switch-mode power supplies (SMPS) are prone to component aging due to prolonged high-load operation, which can lead to failures or shutdowns. Health assessment, through real-time monitoring and failure prediction, can reduce maintenance costs and downtime risks, ensure stable operation, and facilitate the transition of maintenance practices from traditional reactive maintenance to 'condition-based maintenance.' However, the extended operation time of SMPS combined with the high sampling rates of data acquisition equipment results in accelerated expansion of monitoring data, making existing health assessment algorithms designed for low-frequency data difficult to implement in practical SMPS applications. To address this challenge, this study selects a second-stage SMPS from an aging test bench as the research object and constructs a simulation circuit-level health assessment model based on circuit state equations. The circuit model acquires voltage signals at a high sampling rate of 450 kHz. This study employs the Kolmogorov-Arnold Networks (KAN) model, which features lightweight linear computation characteristics. This not only reduces the computational complexity of the model but also alleviates the data processing burden, enabling the construction of an efficient health baseline. On this basis, the Mahalanobis distance between the sample to be assessed and the health baseline is calculated and normalized into a health index, thereby achieving continuous characterization of the health status of the sample and identification of anomaly sensitivity. This enhances the resolution and accuracy of the assessment. Joint experiments based on NASA's public dataset and a physical simulation model demonstrate that the proposed method effectively improves the assessment capability compared with traditional methods.