Self-supervised representation learning with artificial fish swarm optimization for machine anomalous sound detection

Machine Anomalous Sound Detection is crucial for artificial intelligence automation in the context of the fourth industrial revolution. Recent approaches employ self-supervised representation learning, which combines representations extracted through training models on normal and pseudo-anomaly data in the upstream phase with anomaly detection algorithms in the downstream phase. Although effective in extracting representations and demonstrating some robustness, this approach overlooks the complexity and intrinsic relevance of the individual tasks in the upstream models. To address these challenges, this paper introduces the Self-Supervised Multi-Task Representation Learning with Artificial Fish Swarm Optimization. This network integrates two task classifiers into the upstream model, specifically designed to distinguish the sound origin of the target machine and discern attribute information related to the machines. We assign weights to individual tasks and optimize these weights using the artificial fish swarm algorithm. Our approach was evaluated on the DCASE2022 Task 2 evaluation dataset, where the method outperformed integrated models by an average of 6.8% in terms of AUC and pAUC.