Multi-session, multi-device acoustic dataset for progressive tool degradation monitoring.

Audio sensing provides a low-cost, non-contact modality for monitoring mechanical equipment health. Many degradations and faults manifest as gradual changes in spectral and temporal structure (e.g., increased broadband friction noise, harmonic shifts, airflow turbulence changes, and transient impulses), enabling early-warning systems that can support condition-based maintenance and reduce downtime. This article presents a multi-device acoustic dataset designed to study degradation monitoring under realistic cross-device and multi-session variability. The dataset contains labeled recordings from three common motor-driven tools: a shop-vac with (i) discrete fill-level gradations (0%, 30%, 50%, 70%, 100%) and (ii) a mechanically-induced faulty state; a vacuum with (i) discrete clogging gradations (0%, 30%, 50%, 70%, 100% air-filter occlusion) and (ii) power-dial settings (0-8); and an orbital sander with wear-state gradations corresponding to sandpaper lifecycle (New, Moderate, and Worn/Faulty). Recordings were captured with multiple commodity microphones spanning smartphones and external microphones, and were intentionally split into training/testing device groups for cross-device evaluation. Each training condition was captured at two different times and locations, with microphone placement varied during capture to reduce overfitting to environment and geometry. The dataset supports research in robust acoustic condition monitoring, cross-device generalization, domain shift, and data-efficient learning for early fault detection and prognostics.