Deep Learning for Multi-Component Gearbox Fault Diagnosis on the PHM 2009 Benchmark: A Critical Review and Comparative Evaluation of Time–Frequency Representations and Network Architectures

Concurrent faults in gears, bearings and shafts remain among the hardest cases for vibration-based condition monitoring of industrial gearboxes, because their signatures overlap, modulate one another and shift with speed and load. The 2009 PHM Society gearbox has become the standard public benchmark for this multi-component problem, yet published results on it are scattered across input types, network designs and validation schemes, which makes headline accuracies hard to compare. This article presents a critical review of deep-learning diagnosis on this benchmark and the wider gearbox literature, organised around three design choices: signal representation, model family and evaluation protocol. To ground the discussion, we re-analyse a controlled experiment in which three convolutional architectures were trained on 10,240 images per representation (continuous wavelet transform scalograms, short-time Fourier transform spectrograms and wavelet synchrosqueezed images) covering eight spur-gearbox health states. A compact five-stage batch-normalised network reached 99.90%, 99.85% and 99.66% test accuracy with wavelet, Synchrosqueezed and spectrogram inputs, respectively, but the margins among the best models correspond to only a few of the 2,048 test images. On a separate held-out validation subset, accuracy fell to 94.53%, 92.03% and 77.34%, making the representation effect far more visible. Comparison with published capsule-attention, architecture-search and transformer studies shows that reported accuracies on the same benchmark range from below 80% to above 99.5%, mainly because of differences in splitting protocol. We conclude with a reporting checklist and research agenda for leakage-aware, physics-informed and deployable gearbox diagnosis.