Lightweight CNN Architectures for Fault Diagnosis of Power Generator sets: A Comparative Study of MobileNet and AlexNet

Abstract

The use of power generator sets (3.5kVA – 5.5kVA) for domestic and commercial backup supply has become a mainstay in Nigeria due to the unstable electricity from the grid. To keep these backup supplies running, traditional diagnostic approaches that are reliant on manual inspections and physical measurements have been adopted. These are often time-consuming, reactive, and unsuitable for real-time monitoring. To address these challenges, a machine learning approach is presented by performing a comparative analysis of MobileNet and AlexNet convolutional neural networks for automated audio-based fault diagnosis in 5kVA generators. Fault signatures are obtained from acoustic data recorded from 25 generator units under five operational states—Normal, Caburetter, Exhaust, Valve, and Plug faults. Mel-Frequency Cepstral Coefficients (MFCC), Continuous Wavelet Transform (CWT), and Short-Time Fourier Transform (STFT) were employed to transform the raw audio signals into two-dimensional spectrograms that contain both temporal and spectral fault signatures. Using transfer learning, these spectrograms were utilized as input features to train versions of MobileNet and AlexNet, which were pre-trained on ImageNet weights. Their performance was evaluated using accuracy, precision, recall, F1-score, and ROC-AUC metrics. Results obtained from the evaluation metrics show that MobileNet significantly outperformed AlexNet across all feature transformations (MFCC, CWT, and SSTFT). It achieved a peak accuracy of 92% and an AUC of 0.99 with STFT spectrograms. In contrast, AlexNet achieved lower accuracies (54–59%), indicating lower discriminative power. The class-wise ROC-AUC analyses confirmed that MobileNet achieved near-perfect classification, particularly in distinguishing between Normal and any of the fault conditions, while AlexNet struggled with subtle classes, such as Plug and Valve faults. These findings indicate that STFT is the most discriminative spectrogram and MobileNet is the best-performing diagnostic framework. This makes it suitable for deployment in resource-constrained environments and edge devices. This research contributes to the advancement of intelligent, real-time condition monitoring of domestic generator sets, thereby reducing downtime and enhancing energy reliability in off-grid contexts.