Risk-Controlled Adaptive-Resolution Fault Diagnosis of Power Transformers Using Dissolved Gas Analysis: Lightweight Hierarchical Learning, Credibility, and Deployment Support

Dissolved gas analysis (DGA) is widely used for transformer condition assessment, but fixed-resolution classifiers may release overly specific fault labels when the available evidence is weak. This study develops a risk-controlled adaptive-resolution framework in which a lightweight hierarchical residual neural encoder generates fine-grained and electrical/thermal parent-level candidates, and separate credibility estimates determine whether the released result should remain fine-grained, fall back to the parent mechanism, or abstain. The 25-dimensional DGA representation is encoded by three residual fully connected blocks (64-32-24), yielding 9492 trainable parameters for the hierarchical B2 model. Credibility combines calibrated uncertainty and DGA physics-consistency evidence, while data quality and distributional support are treated separately as deployment admissibility safeguards. Under 10 × 5 repeated evaluation, the B2 backbone achieved a Macro-F1 of 0.723 ± 0.026 in ordinary cross-validation and 0.664 ± 0.086 under reference-grouped evaluation. In the adaptive-resolution analysis, hierarchical risk decreased from 0.212 to 0.094 and cross-parent risk from 0.062 to 0.026 while retaining approximately 91.5% answer coverage in the ordinary protocol. Physics augmentation was context-dependent: it was inferior to the calibrated margin under ordinary resampling but substantially improved AURC under reference-grouped evaluation. CPU-only profiling gave a median end-to-end latency of 0.713 ms per observation (P95 1.028 ms), far below the sampling cadence of the utility online DGA data. The study therefore emphasizes risk-controlled information release rather than universal superiority of the underlying classifier.Exploratory temporal results are presented solely as a transparency analysis and are not interpreted as evidence of validated early-warning capability.