Deployment-aware acoustic anomaly detection on an embedded microcontroller

Abstract

Industrial acoustic anomaly detection is often optimized in off-device benchmarks, while the consistency of exported thresholded decisions on embedded microcontrollers remains less studied. This paper presents a bounded embedded consistency-evaluation case study for an internet of things (IoT)-class acoustic monitoring pipeline on the public MIMII fan 0 dB subset. The pipeline uses a 40×63×1 log-mel representation, compact convolutional autoencoder (CAE), INT8 TFLite deployment, and latent-space Mahalanobis scoring. The raw Mahalanobis rule defines the on-device decision; a 2-out-of-5 temporal rule is retained only for offline ordered-subset analysis. For fan/id 06/0 dB, repeated stratified evaluation yielded average area under the curve (AUC) 0.96 and F1 0.82. Supporting all-ID evaluation over id 00, id 02, id 04, and id 06 showed the raw Mahalanobis path had the strongest aggregate metrics among tested scoring paths. Stored-sample replay on ESP32-S3 used a fixed 20-sample id 06 batch with generated C audio headers through the canonical DSP path. Both reference-input and DSP-path checks produced 20/20 raw-decision agreement with the off-device reference. The results support bounded reference-device consistency for the evaluated batch and provide a reference-device basis for future embedded, IoTedge, and block-level reconfigurable-system studies, but not cross-ID robustness, streaming, latency, or energy behavior.