Anomaly Sound Detection of Industrial Equipment Based on Incremental Learning

The objective of the anomaly sound detection task is to monitor for sounds coming from the target object and analyze whether they are coming from it normally or in an anomalous status. Existing anomaly sound detection models have limitations and are difficult to learn effective knowledge from continuous data streams, which makes them unsuitable for general applications. Incremental learning allows a model to gradually improve its performance and accuracy by utilizing new data and knowledge without requiring to retrain the entire model. This paper proposes an incremental learning-based anomaly sound detection model that enhances the model’s capacity to learn from continuous data streams, reduces knowledge forgetting, and improves the model’s stability in anomaly sound detection task. The amalgamation of knowledge distillation loss and crossentropy loss in the process of model incremental learning causes the model to remember the old knowledge while learning new devices. Experiments using Task 2 data from the DCASE 2020 challenge shows that the proposed method effectively improves the average AUC and average pAUC by 7% to 10% when compared to the fine-tuning strategy.