Edge-AI Severity Regression and Asymmetric Loss for Controlled PV Electroluminescence Benchmarking and SCADA Alerting

Abstract

This manuscript reports a controlled benchmark study that combines continuous EL severity regression, a Safety-Aware Asymmetric Huber Loss (SAHL) objective, Raspberry Pi ONNX deployment, and MQTT/Node-RED/PostgreSQL alerting on laboratory electroluminescence (EL) data. The work integrates known elements—cost-sensitive regression, edge inference, and auditable database-layer alert logic—into a bounded PV maintenance workflow rather than demonstrating field-validated autonomous inspection. Relative to an MSE baseline, SAHL at λ = 1.5 reflects an operational policy trade-off: critical recall rises from 0.784 ± 0.048 to 0.806 ± 0.046 and MAE falls from 0.1959 ± 0.0037 to 0.1700 ± 0.0047, accompanied by a moderate precision drop (0.856 ± 0.021 to 0.828 ± 0.035). A multi-seed weight ablation across λ ∈ {1.0, 1.5, 2.5, 5.0} shows no single weight dominates: 1.0× yields peak F1 (0.774) and precision (0.854), 1.5× minimizes MAE, 2.5× boosts recall at precision cost, and 5.0× is unstable. Weight selection therefore depends on operational objectives. A sustained Raspberry Pi 4 benchmark over 1,588 measured inferences reports ONNX latency of 190.32 ± 27.93 ms (median 177.63 ms, P95 238.77 ms), throughput of 5.25 FPS, and peak RSS of 128.98 MB. Full-module outdoor validation remains future work.