TY - JOUR

T1 - Enhancing quality inspection of highly variant geared motors

AU - Wißbrock, Peter

AU - Koschek, Lukas

AU - Ren, Zhao

AU - Nejdl, Wolfgang

N1 - Publisher Copyright: © 2025 Elsevier Ltd

PY - 2025/5/14

Y1 - 2025/5/14

N2 - Quality inspection is an important step in industrial production to avoid early failures and customer complaints. Custom-built products are often produced in a high number of variants and a low number of instances per variant, which makes automation of quality inspection a difficult task. Today's manual inspection is cost-intensive and error-prune. This paper introduces the unexplored problem setting of quality inspection of highly variant geared motors and provides a full concept solving this industrial sound analytics problem. To enable research on this important topic, we introduce the publicly accessible Lenze quality inspection dataset, named Lenze-QI* and highlight its special challenges and characteristics. In contrast to other existing datasets, it is fully real-world, and it includes the variants configuration that can be used in machine learning algorithms. It is further demonstrated that the problem setting cannot be solved via state-of-the-art methods but using our proposed two-step approach. First, three feature vectors from the domain of advanced signal processing are proposed. The best performing approach is based on psychoacoustics, while also the approaches using the logarithm envelope or deep learning in combination with optimized spectrogram parameters are still useful. Second, we introduce the conditional node type to be used in an isolation forest, taking the variant's configuration as input. The resulting conditional isolation forest is a novel anomaly detection approach taking additional attributes into account. Overall, the best performance for Lenze-QI can be achieved by an ensemble of the three advanced signal processing approaches in combination with the novel conditional isolation forest. The state-of-the-art performance is overruled by an increase of 13% in the area under the receiver operating characteristic. We demonstrate that the quality inspection of geared motors can be automated despite many challenges. *The dataset Lenze-QI can be downloaded following https://doi.org/10.5281/zenodo.13854459.

AB - Quality inspection is an important step in industrial production to avoid early failures and customer complaints. Custom-built products are often produced in a high number of variants and a low number of instances per variant, which makes automation of quality inspection a difficult task. Today's manual inspection is cost-intensive and error-prune. This paper introduces the unexplored problem setting of quality inspection of highly variant geared motors and provides a full concept solving this industrial sound analytics problem. To enable research on this important topic, we introduce the publicly accessible Lenze quality inspection dataset, named Lenze-QI* and highlight its special challenges and characteristics. In contrast to other existing datasets, it is fully real-world, and it includes the variants configuration that can be used in machine learning algorithms. It is further demonstrated that the problem setting cannot be solved via state-of-the-art methods but using our proposed two-step approach. First, three feature vectors from the domain of advanced signal processing are proposed. The best performing approach is based on psychoacoustics, while also the approaches using the logarithm envelope or deep learning in combination with optimized spectrogram parameters are still useful. Second, we introduce the conditional node type to be used in an isolation forest, taking the variant's configuration as input. The resulting conditional isolation forest is a novel anomaly detection approach taking additional attributes into account. Overall, the best performance for Lenze-QI can be achieved by an ensemble of the three advanced signal processing approaches in combination with the novel conditional isolation forest. The state-of-the-art performance is overruled by an increase of 13% in the area under the receiver operating characteristic. We demonstrate that the quality inspection of geared motors can be automated despite many challenges. *The dataset Lenze-QI can be downloaded following https://doi.org/10.5281/zenodo.13854459.

KW - Acoustic Anomaly Detection

KW - Anomalous Sound Event Detection

KW - Domain Generalization

KW - Industrial Sound Analytics

KW - Psychoacoustics

KW - Quality Inspection

KW - Real-World Dataset

U2 - 10.1016/j.apacoust.2025.110687

DO - 10.1016/j.apacoust.2025.110687

M3 - Article

AN - SCOPUS:105000669270

VL - 235

JO - Applied acoustics

JF - Applied acoustics

SN - 0003-682X

M1 - 110687

ER -