ConvDTW-ACS: Audio Segmentation for Track Type Detection During Car Manufacturing

• DOI: 10.48550/arxiv.2402.18204
• Publication Date: 2024-02-28
• Authors: Álvaro López-Chilet; Zhaoyi Liu; Jon Ander Gómez; Carlos Álvarez-Bel; Marivi Alonso Ortiz; Andres Orejuela Mesa; David Newton; Friedrich Wolf; Sam Michiels; Danny Hughes

This paper proposes a method for Acoustic Constrained Segmentation (ACS) in audio recordings of vehicles driven through production test track, delimiting the boundaries surface types track. ACS is variant classical acoustic segmentation where sequence labels known, contiguous and invariable, which especially useful this work as track has standard configuration types. The proposed ConvDTW-ACS utilizes Convolutional Neural Network classifying overlapping image chunks extracted from full spectrogram. Then, our custom Dynamic Time Warping algorithm aligns predicted probabilities to timestamps type can be extracted. was evaluated on real-world dataset collected Ford Manufacturing Plant Valencia (Spain), achieving mean error 166 milliseconds when delimiting, within audio, surfaces results demonstrate effectiveness accurately segmenting different types, could enable development more specialized AI systems improve quality inspection process.