Work

Audio anomaly detection in industrial machines using inverse-Mel scale spectrograms

Year: 2025

Type: article

Abstract: Industrial machine sounds are commonly analyzed using Mel frequency scales, which are based on human auditory response and are primarily sensitive to low-frequency sounds. However, sounds from real industrial machines span over a wide range of frequencies. Traditional linear or Mel frequency scales fail to adequately capture the high-frequency components present in machine sounds. In this work, we propose utilizing inverse-Mel scale spectrograms to effectively collect information from the high-frequency regions of industrial machine sounds. Combination of Mel and inverse-Mel scale spectrograms is also pursued to achieve improved resolution for both low and high frequencies within a single spectrogram. Deep Autoencoder, shared as baseline system in the DCASE 2020 (Detection and Classification of Acoustic Scenes and Events) Challenge Task 2, is used as Audio Anomaly Detector (AAD). Experiments with spectrograms as features on the MIMII (Malfunctioning Industrial Machine Investigation and Inspection) dataset and the Air compressor machine sound dataset demonstrate significant improvements over the challenge baseline scores with a maximum improvement of up to 22% in the Slider machine and an average improvement of up to 8% in the partial-Area under the receiver operating characteristic (pAUC) score across all machines. • Proposal to extract the often neglected high-frequency sound energies present in industrial machine sounds and use them as front-end features for Audio Anomaly Detection in industrial machines. • Formulate a mathematical formula for the inverse-Mel scale and use it to construct inverse-Mel scale spectrograms. • Integrate the Mel scale and inverse-Mel scale spectrograms into a single spectrogram. The combined spectrogram offers good resolution for both low and high frequencies. • Testing with the inverse-Mel scale and combined spectrograms on the DCASE 2020 Challenge Task 2 datasets demonstrate an average improvement of up to 8% in the pAUC score.

Source: Systems and Soft Computing

DOI: 10.1016/j.sasc.2025.200398

Authors: Kader B.T. Shaikh, Naresh P. Jawarkar, Vasif Ahmed, Nadir N. Charniya

Institution: Babasaheb Naik College of Engineering, Pusad

Language: English

Topic: Machine Fault Diagnosis Techniques

Subfield: Control and Systems Engineering

Field: Engineering

Domain: Physical Sciences

SDG: Industry, innovation and infrastructure

Open Access status: gold

Estimated APC paid: $1,870