Predictive AutoEncoders Are Context-Aware Unsupervised Anomalous Sound Detectors

Abstract

In this paper, we propose a Predictive AutoEncoder (PAE) capable of exploiting context information for unsupervised anomalous sound detection (ASD). The conventional unsupervised ASD approaches mainly employ the straightforward deep neural network (DNN) to detect abnormal sounds. However, this model fails to consider the utilization of the relationship between frames, resulting in limited performance and constrained input length. Recently, context information has been proven to be valid for sequence data processing. In our method, the PAE consisting of transformer blocks is proposed to predict unseen frames by remaining available inputs. Based on the self-attention mechanism, our model captures not only content information within the frame but also context information between frames to improve ASD performance. Moreover, our method extends the input length of AE-based models due to its outstanding capability of long-range sequence modeling. The extensive experiments conducted on the DCASE2020 Task2 development dataset demonstrate that our method outperforms the state-of-the-art AE-based methods and verify the effectiveness and stability of our proposed method for long-range temporal inputs.