Blind Anomalous Sound Detection without Specifying Target Type

Unsupervised anomalous sound detection (ASD) is an important and practical task. Recently, ASD methods based on autoencoder reconstruction errors and ambiguity of classifiers have improved ASD performance. The number of target types tends to increase, but it is general to train machine-dependent models for each machine type and switch models at inference time. In addition, this machine-dependent modeling is brittle with the condition difference between training and testing because whole training data containing various machine-type data cannot be used. To address these problems, it is effective to condition models using machine-type labels. However, even in this method, it is necessary to know the type of machine in advance, which is not convenient when different machines work in the same place. We propose a blind ASD model trained using multitask learning, which does not need to know the machine type at inference time. We conduct experiments on task2 of the DCASE 2022 challenge to compare the performance of conditioned models with that of our blind ASD models. The results show that the proposed method improved the ASD performance compared with machine-dependent models and conditioned models, even though the target machine type was not specified and the model size is the same.