Industrial scene abnormal sound detection and identification method based on auto-encoder

Abstract

The invention provides an industrial scene abnormal sound detection and recognition method based on an auto-encoder. The method comprises three processes of sound feature extraction, industrial scene sound modeling and abnormal sound detection and recognition. The method comprises the following steps of: extracting sound features of an industrial scene, performing complementary set empirical mode decomposition on each frame of preprocessed signal to obtain a plurality of layers of Intrinsic Mode Function (IMF) components, and combining short-time energy of each layer of IMF component, an energy ratio of IMF energy to an original signal of the frame and a Mel-frequency cepstral coefficient of the IMF to form a feature vector of a target sound signal; training an auto-encoder by using the extracted feature vectors, and adjusting parameters of the encoder and a decoder to obtain a normal scene sound model and a specific abnormal scene sound model; and abnormal sound detection and identification: enabling the features of the sound to be detected to pass through a trained auto-encoder, and judging whether the sound to be detected is abnormal sound or certain known abnormal sound through a threshold condition of a reconstruction error.

Description

technical field [0001] The invention relates to sound signal processing technology, in particular to a method for detecting and identifying abnormal sounds in industrial scenes based on an autoencoder. Background technique [0002] Sound is an important information carrier, which contains very important information. At the same time, the occurrence of abnormal events is often accompanied by the generation of abnormal sounds. For example, abnormal events in public scenes are often accompanied by explosions, screams, and glass explosions. Also in the industrial scene, various equipment under normal operation will generate regular vibrations, and its sound characteristics have certain rules. Once the machine is damaged or other abnormal events occur, the sound characteristics of the industrial scene will change. Therefore, detection and Identifying abnormal sounds in audio signals plays an important role in the security monitoring of industrial scenes. [0003] At present, th...

AI Technical Summary

Problems solved by technology

However, abnormal sounds are highly non-stationary and nonlinear, resulting in the extraction of a single feature will reduce the final recognition efficiency

Images