Submitted:
17 November 2026
Posted:
19 November 2026
You are already at the latest version
Abstract
This data descriptor presents a manufacturing sound dataset collected during actual production on industrial shop floors. It comprises 57 continuous recordings totaling 28.91 hours, with 6,598 annotated target-operation events from 17 machines at nine companies. Seven manufacturing processes are represented: hydroforming, injection molding, pipe bending, plasma cutting, seam welding, stamping, and tube cutting. Structure-transmitted machine sound was captured using Internal Sound Sensors (ISSs), each comprising a USB microphone coupled to a stethoscope whose bell was attached to the machine. The dataset provides continuous WAV recordings, manually reviewed operation annotations, and metadata describing the machines, sensor installations, and recording sessions. Retaining continuous recordings preserves both target-operation sounds and the surrounding production context, allowing users to analyze complete recordings or extract individual events. The recordings capture variation in acoustic patterns and operation durations across machines, production settings, and processes. This combination of industrial diversity and temporal annotations supports research on operation detection, temporal segmentation, and recurring acoustic patterns. It also provides a resource for self-supervised representation learning, transfer learning, and evaluation under limited labeled-data conditions. The dataset is publicly available through Harvard Dataverse.
Keywords:
manufacturing sound dataset
; internal sound sensor
; temporal event detection
; machine monitoring
; self-supervised learning
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.