Wavelet Transform Data-Driven Machine Learning-Based Real-Time Fault Detection for Naval DC Pulsating Loads

All-electric warships require high power-dense distribution system to power-advanced weapon loads. Medium-voltage dc (MVdc) power distribution is well suited to fulfill this requirement if certain risks are addressed. A particularly emerging problem is that the advanced pulsating loads draw large currents in extremely short periods of time and behave similar to the shunt fault. The nature of the load and the operating cycle determines the unique structure of the pulse in time and frequency domains. If the load operating cycle consists of a finite number of transitions, then the corresponding frequency content of the current profile can be used to identify these transients. The wavelet transform is used to extract this useful frequency domain information from the sampled current data. A proposed computationally light data-driven machine learning-based fault detection and load monitoring solution extracts the frequency-domain features of the observed transient and compares that to a database of stored features to identify the observed transient and then to further identify faults that may create an abnormal disturbance in the load current profile such as arcing and shunt faults. It can be further applied to any load profile with prerequisite of a finite number of repetitive transients during normal condition. This article focuses on the fault detection only and not for fault isolation while it can achieve isolation capability once the fault was diagnosed. In final real-time implementation, the recursive Haar stationary wavelet transform (SWT)-fed computationally light machine learning is employed to validate the proposed scheme in a single-core Texas Instrument (TI) digital signal processer (DSP) TMS320F28335.