Hierarchical temporal memory-based predictive maintenance of DC-link capacitors in power electronic converters

Predictive maintenance of capacitors in the DC-link of the power electronic converters through a condition monitoring system is essential because it improves the converters' availability and significantly cuts down the converter maintenance time. Capacitors are an essential component to monitor, as they account for 30% of failures in a converter. This paper addresses the problem of early and online detection of DC-link capacitor aging under varying operating conditions using unlabeled data, without requiring additional sensors or offline retraining. The aging behaviour of the DC-link capacitor is determined using a degradation data model that considers the dissipation factor, capacitance, and equivalent series resistance. The proposed HTM method continuously learns temporal and spatial patterns in ripple-voltage data and distinguishes between normal operational variations and aging-induced deviations in real time. The effectiveness of the approach is evaluated using simulation data under multiple degradation levels and operating conditions, and its performance is compared with several state-of-the-art unsupervised anomaly detection techniques. Quantitative evaluation demonstrates that the proposed HTM-based approach achieves detection accuracy above 95% while maintaining false alarm rates below 4%, consistently outperforming other unsupervised baseline methods under varying operating and degradation conditions. These findings demonstrate that the proposed HTM-based method is more efficient in aging detection. • Most existing condition monitoring techniques require external measuring devices. • Neuro-inspired HTM-based machine learning method is used for ageing detection. • Proposed approach is compared with five state-of-the-art age detection algorithms • Capacitance of a capacitor is a crucial degradation indicator along with the ESR. • The proposed method is effective in aging detection than the conventional methods.