An End-to-End Unsupervised Approach to Acoustic Anomaly Detection with Deep Support Networks

Abstract

The invention discloses an end-to-end unsupervised deep support network acoustic anomaly detection method, the steps of which are as follows: converting an acoustic signal into an acoustic spectrogram signal of a Mel spectrum; dividing the collected sound signal into training, verification, The test set, in which the verification set is responsible for determining the abnormal threshold; constructing an unsupervised deep support network, including feature learning network is responsible for extracting acoustic features, and deep detection network is responsible for judging whether the sound signal is abnormal; constructing the loss function of deep support network, including feature The learned least squares loss function and the soft interval hinge loss function of the deep detection network; train the verification set to calculate the optimal detection threshold; use the trained deep support network to quantitatively calculate the acoustic outliers. The invention reduces the calculation cost, uses precision rate, recall rate and F1 value as judgment scales, reduces the human intervention of the algorithm, and improves the precision of hyperspectral quantitative analysis.

Description

Technical field: [0001] The invention belongs to the technical field of non-destructive detection of acoustic frequency spectrum, and in particular relates to an end-to-end unsupervised deep support network acoustic anomaly detection method. Background technique: [0002] Acoustic anomaly detection and analysis technology has a wide range of application scenarios, including food packaging detection, pronunciation rehabilitation treatment, laryngoscopy detection, industrial production detection, etc. Commonly used acoustic anomaly detection algorithms are regarded as unsupervised learning problems, that is, abnormal acoustic samples are treated as unknown samples a priori, and most of the training data are normal data; in the test phase, the acoustic data that is different from the distribution of training data Identified as abnormal data; acoustic anomaly detection analysis techniques include convolutional autoencoder network combined with single-class support vector machine...

AI Technical Summary

Problems solved by technology

In order to transfer the advantages of deep learning to the field of acoustic anomaly detection, many deep generative models have shown good results in the field of anomaly detection, but these models mainly rely on the recovery error between the heuristic generated signal and the original signal to judge whether the sound data is Anomalies, and there are relatively few deep models that directly take anomaly detection as the target equation

Images