基于自编码器的无监督机器异常声检测

Aiming at the problem that machine abnormal sound rarely occurs and has highly variability and nonstationary, an unsupervised detection method of anomalous sounds for machine based on auto-encoder is proposed.First, the frequency spectrum features of normal sound are used to train the auto-encoder and reconstructed.Then, the error between the features of the sound and the reconstructed features is used to perform anomaly detection. The experimental results on DCASE2020 Challenge Task2 dataset show that compared with the Task2 baseline system, the proposed method not only guarantees a good classification accuracy, but also significantly improves the AUC value, and has improvement in the performance of machine abnormal sound detection.