Internal leakage in the early stages within valves of hydropower-unit auxiliary systems (e.g., compressed-air subsystems) is typically minor, concealed, and heavily influenced by operating-condition variability, making it difficult to detect using fixed empirical thresholds. Supervised learning approaches are similarly constrained by the scarcity of on-site fault samples. This study establishes a practical acoustic anomaly-detection framework that (ⅰ) learns healthy valve acoustic signatures primarily from normal data, (ⅱ) remains applicable across multi-pressure operating conditions, and (ⅲ) supports robust alarm decision-making through condition-aware and online-adaptive thresholding.

A controllable experimental platform was constructed by replicating the layout of a hydropower-station auxiliary compressed-air system and incorporating a real in-service DN200 hemispherical valve as the test object. Valve sounds were recorded using two microphones placed symmetrically around the valve body at approximately 1.0–1.1 m, with synchronized timestamps to ensure consistent multi-channel acquisition. Data were collected under multiple pressure levels ranging from 0.2 to 0.7 MPa. Two representative states were examined: a fully closed valve representing the normal condition and a 10% opening used to emulate internal leakage. To improve the stability of acoustic inputs, multi-channel waveforms were aligned via cross-correlation, DC offsets were removed, and a band-pass filter was applied to retain diagnostically relevant frequency content. Recordings were then resampled to a unified sampling rate, and segments exhibiting clipping or prolonged saturation were discarded. The processed signals were segmented using a fixed-length sliding window (approximately one second per segment) with different strides for training and evaluation, allowing normal data to provide sufficient training diversity while preserving high temporal coverage during testing. Log-Mel spectrograms were extracted as compact time-frequency representations through short-time Fourier transform, Mel filter-bank projection, logarithmic compression, and per-channel standardization based solely on normal data statistics. On the modeling side, a self-supervised dual-path Transformer (SSDPT) was employed to alternately capture dependencies along the time and frequency dimensions, enabling fine-grained characterization of leakage-induced spectral structures and their temporal evolution. Training combined a discriminative identification objective (learning to recognize normal operating signatures across groups and conditions) with a reconstruction objective under random patch masking, encouraging robust representation learning without requiring extensive labeled fault samples. During inference, anomaly scores were computed primarily from classification-based confidence decay (i.e., reduced confidence in the learned healthy identity implies greater abnormality), and this scoring strategy was compared against reconstruction-inclusive alternatives.

The classification-based score provided the most reliable separation between normal and leakage segments. Across the complete dataset, the overall area under the receiver operating characteristic (ROC) curve reached 0.707, while the partial area under the curve (AUC) in the low-false-alarm region (false positive rate ≤ 0.10) reached 0.417, indicating meaningful discrimination capability under practical low-false-alarm constraints. Performance was, however, strongly pressure-dependent: medium and high pressures exhibited clearer separability and more stable high-score tails associated with leakage, whereas certain low- and mid-pressure conditions showed substantial score overlap between normal and leakage segments, limiting the effectiveness of a single global threshold. Pressure-wise analysis revealed near-complete separability at the highest pressure level and useful separability at some medium pressures, while other lower-pressure settings approached chance-level ordering except for a small subset of strongly abnormal segments detectable at very low false-positive rates. To translate scores into actionable alarms, two complementary thresholding mechanisms were developed. First, pressure-level-specific thresholds were designed to compensate for systematic distribution shifts across pressures and to reduce mismatches caused by mixed-condition score scaling. Second, an online adaptive threshold scheme was formulated to update alarm boundaries during long-term operation by tracking a rolling high quantile of recent scores and calibrating robustness via median absolute deviation, thereby improving stability against gradual background drift and intermittent disturbances.

This study demonstrates that SSDPT-based self-supervised acoustic anomaly detection can serve as a feasible, engineering-oriented approach for internal leakage monitoring in hydropower auxiliary valves when fault labels are limited. Multi-pressure experiments confirm that operating conditions markedly affect score distributions and detection separability, making condition-aware thresholding essential for reliable deployment. The proposed pressure-level and online adaptive threshold strategies enhance decision robustness across operating regimes and over time. Remaining challenges are concentrated in low-pressure scenarios, where leakage signatures may be weak or masked by background noise. Future work may address these regimes through richer sensing configurations, acoustic-vibration multimodal fusion, and validation within longer-term field monitoring pipelines.