Edge-Optimized Structural Health Indexing of Pole-Mounted Transformers via Multi-Task Deep Regression and Visual Saliency

Manual visual inspection remains the default condition-assessment method for pole-mounted distribution transformers (PMTs) in dense urban grids, producing inconsistent, technician-dependent health scores that do not scale across thousands of field assets. This paper presents an edge-deployable, two-stage deep learning pipeline that automates PMT structural health indexing from a single RGB photograph. A binary EfficientNet-B0 classifier gates non-transformer inputs before a second EfficientNet-B0 network, extended with 13 independent regression heads, scores each transformer against 13 standardized structural defect parameters (oil leakage, corrosion, rust, bushing cracks, and others) on a 0-6 severity scale. Gradient-weighted Class Activation Mapping (Grad-CAM) grounds every prediction in a pixel-level saliency map, and a technician-feedback-driven adaptive layer refines global and case-specific offsets post-deployment without full retraining. On a held-out test set of 779 field images, the 13-head regression model achieves an overall Mean Absolute Error (MAE) of 0.091 and a mean coefficient of determination (R²) of 0.989 across all 13 parameters, reflecting in-distribution performance over the augmented corpus. However, strict group-disjoint evaluation on truly unseen field assets reveals significant performance degradation, identifying the limited 37-image seed dataset as the primary bottleneck for real-world generalization. Independent of regression accuracy, the complete two-stage inference pipeline—PMT gating, 13-head regression, and Grad-CAM saliency generation—executes in 115 ms per image on commodity CPU hardware (AMD Ryzen 5 3600, 6-core/12-thread) with the GPU explicitly disabled, confirming that the software architecture itself is field-ready for low-cost, edge-side deployment.