Autoencoder with Group-based Decoder and Multi-task Optimization for Anomalous Sound Detection

Yifan Zhou

Abstract: In industry, machine anomalous sound detection (ASD) is in greatdemand. However, collecting enough abnormal samples is diffcultdue to the high cost, which boosts the rapid development of unsu-pervised ASD algorithms. Autoencoder (AE) based methods havebeen widely used for unsupervised ASD, but suffer from problemsincluding 'shortcut’, poor anti-noise ability and sub-optimal qualityof features. To address these challenges, we propose a new AE-basedframework termed AEGM, Specifically, we first insert an auxiliaryclassifier into AE to enhance ASD in a multi-task learning manner.Then, we design a group-based decoder structure, accompanied byan adaptive loss function, to endow the model with domain-specificknowledge. Results on the DCASE 2021 Task 2 development setshow that our methods achieve a relative improvement of 13.11%and 15.20% respectively in average AUC over the offcial AE andMobileNetV2 across test sets of seven machines.