Mamba-Based Fault Classification of Electrical Generators under Fixed- and Variable-Speed Control Conditions

This study proposes a Mamba-based deep learning framework for intelligent fault classification in synchronous generators operating under different electrical and control conditions. Two experimentally acquired datasets were considered: a fixed-speed 3 kVA four-pole smooth-pole synchronous generator and a variable-speed 2 kVA four-pole salient-pole synchronous generator. The datasets include healthy operation and multiple internal electrical fault conditions represented through three-phase current and voltage measurements, with the variable-speed system additionally reflecting converter-based control behavior. The measured signals were normalized, segmented into fixed-length sequences, and processed using a Mamba architecture designed to capture long-range temporal dependencies and fault-related waveform variations. An 80/20 training–testing split was adopted for model development and evaluation. For Generator 1, the proposed model achieved 98.84% training accuracy and 97.36% testing accuracy, with testing AUC and specificity values of 99.12% and 99.34%, respectively. For Generator 2, the corresponding training and testing accuracies were 98.21% and 96.88%, while testing AUC and specificity reached 98.76% and 99.18%. Confusion-matrix and ROC analyses further demonstrated consistent discrimination among healthy and fault classes with limited misclassification. The results indicate that Mamba can effectively extract temporal fault signatures under both fixed-speed and variable-speed control environments, supporting its potential application in intelligent condition monitoring and supervisory fault-diagnosis systems for rotating electrical machines.