SQ-CAP: Sensor-quality-guided Conflict-aware Pooling for Auditable Edge-AI Fault Screening

Abstract

Factory monitoring combines complementary sensors, but disagreement can arise from either process faults or sensor degradation. We present Sensor-Quality-Guided Conflict-Aware Pooling (SQ-CAP), a constrained decision layer for auditable edge artificial intelligence fault screening. It combines temperature-scaled detector logits, separately estimated quality weights, and absolute logit disagreement. On a reproducible twenty-seed synthetic benchmark, the 8-bit integer (INT8) scalar-input profile achieves an area under the precision–recall curve (AUPRC) of 0.807, compared with 0.760 for quality-blind logit averaging and 0.813 for unconstrained logistic fusion. Refitted ablations identify conflict as the principal source of the synthetic gain; additional score-level sensitivity tests show that this advantage depends on the conflict–fault relationship and can reverse under distribution shift. A Case Western Reserve University (CWRU) bearing-data experiment uses vibration features with injected feature perturbations, rather than synchronized camera–microphone measurements. Adding the quality covariate improves the area under the receiver operating characteristic curve (AUROC) from 0.828 to 0.951, but an audit finds substantial fault information in the quality proxy itself. This result supports predictive use of an additional covariate, not degradation-specific reliability validation. The contribution is an explicit, reproducible pooling design with inspectable components; resource costs remain analytic estimates and deployment-level fusion performance is untested.