EdgePHM: An Edge-Native AI Ensemble for Industrial Battery Prognostics Fusing Quantization, Explainability, and Cost-Sensitivity
Abstract
Current predictive maintenance architectures force a critical compromise between diagnostic precision and physical hardware constraints, routinely failing to deploy complex learning models onto resource-limited microcontrollers. The proposed framework dismantles the limitation through a decentralized, post-training quantized heterogeneous architecture. Integration of a Mahalanobis distance applicability domain intercepts out-of-distribution sensor anomalies prior to algorithmic execution. Empirical validation against multi-cell aerospace telemetry confirms a macro F1-score of 0.964 and a verified macro ROC-AUC of 0.999. Pareto optimization strictly identifies the 16-bit floating-point configuration as the engineering sweet spot, compressing the memory footprint down to 164.49 kilobytes. Single instruction multiple data acceleration drives inference latency down to 4.234 milliseconds per cycle. Explainable feature importance protocols autonomously isolate thermodynamic degradation signatures, securing a 63.333 percent prognostic lead time. Eradication of critical machine failures transforms classification precision into tangible financial security, generating a confirmed 1362.950 percent return on investment. Decentralized processing networks fundamentally neutralize reliance on vulnerable cloud infrastructures. Ubiquitous deployment of localized, physics-aware microprocessors establishes a definitive baseline for sustainable manufacturing, directly advancing global circular economy mandates and climate action directives.
