Dual-Mode Hidden Markov Models For Smart Detection Of Clogging In Variable Frequency Drives

Efficient detection of anomalies/failures in the cooling system of Variable Frequency Drive (VFD) is important to reduce costs linked to downtime periods caused by overheating of electronic circuits. Smart condition monitoring tools, including those utilizing supervised machine learning, are recently seen particularly efficient for failure detection due to their enormous flexibility and adaptability. Recently, attention has been paid to monitoring the impact of environmental factors on VFD's operation involving pollution and humidity. Clogging is seen as a particularly harmful failure as it can deteriorate power electronics, thus leading to prolonged machine downtimes. In this study, the use of Dual-Mode Hidden Markov Models (HMM) was proposed for clogging detection via anomaly/novelty detection, so that both pointwise (parcel) and contextual (temporal) anomalies could be identified in time series data from a VFD. This was achieved thanks to simultaneously using likelihoods and signal reconstruction errors as anomaly scores. Detection of cooling issues due to blockages in the VFD's cooling system: inlets, outlets, and heatsink was of interest. We have also automatized model initialization and so was achieved thanks to the unsupervised Gaussian Mixture Models algorithm. This study shows that HMMs' sensitivity to clogging anomalies is improved for models with higher number of hidden states (complexity). The results show that HMMs are feasible for clogging detection in VFD's inlet and heatsink with 80% accuracy. The method, in addition to failure detection, is also shown feasible in preventive maintenance.