<p>Early detection of faults in hydraulic turbine generator units is required to assess reliability and prevent failures and maintenance in hydropower. Existing monitoring approaches, however, mainly rely on either vibration or threshold analysis, which often fail to capture early fault signatures and generalise poorly across different machines and conditions. Given these limitations, this paper proposes an integrated acoustic-driven early warning framework based on Generalised Machine Acoustic Signatures (GMAS), an Auxiliary Classifier Generative Adversarial Network (ACGAN), and Auto-Associative Kernel Regression (AAKR). The GMAS framework extracts robust time–frequency acoustic features, while ACGAN enhances fault-feature representation and addresses the limited number of fault samples. The AAKR model performs residual-based analysis to enable high-sensitivity early fault detection. The proposed framework was validated using multiple industrial acoustic datasets, including MIMII, ToyADMOS, and DCASE, and further evaluated on real hydropower Kaplan turbine operational data. The experimental results show that the proposed algorithm has better detection performance than the baseline algorithms. It achieves 96.8% accuracy, an AUC of 0.98, and a significantly longer early warning lead time. The findings indicate that the proposed GMAS-ACGAN-AAKR framework is effective for early-fault detection in predictive maintenance of hydraulic turbine generator units (HTGUs).</p>