Semi-Supervised Acoustic Anomaly Detection for Artificial Intelligence Stored-Grain Insect Infestation Under Silo Sensor Drift: A Computer Detection-Latency Evaluation

Acoustic sensing of stored-grain insect infestation is vulnerable to gradual drift in silo-mounted microphones, which shifts the spectral baseline and inflates false alarms in supervised detectors trained on labelled infestation episodes. The framework is formalised as a reconstruction-probability model over short-time acoustic frames, with a drift-compensation term estimated from unlabelled silo audio. Detection latency is defined as the number of frames between infestation onset and the first alarm that exceeds a calibrated threshold, and its operating characteristics are derived analytically and checked by a replication-based evaluation across a declared design grid. The evaluation uses 2000 replications per design cell, cluster counts of 8, 16, 24 and 32, and effect sizes of 3, 5 and 8 percentage points, with an intraclass correlation of 0.64, a between-cluster standard deviation of 6.0 and a residual standard deviation of 4.5. Under these parameters, 56 clusters are required to detect a 5-percentage-point effect at 80% power. The contribution is a latency-oriented specification of semi-supervised acoustic detection that separates drift compensation from infestation classification and states the conditions under which the detector remains calibrated.