CN-122024766-A - Method, device, equipment, medium and product for detecting equipment abnormality based on audio data

CN122024766ACN 122024766 ACN122024766 ACN 122024766ACN-122024766-A

Abstract

The invention discloses a method, a device, equipment, a medium and a product for detecting equipment abnormality based on audio data. The method comprises the steps of extracting time-frequency characteristics of target audio data according to logarithmic Mel frequency spectrums of the target audio data, extracting channel characteristic vectors matched with time-frequency positions from the time-frequency characteristics of the target audio data, determining distribution parameters of the time-frequency positions according to the channel characteristic vectors matched with the time-frequency positions, calculating distribution offset of the time-frequency positions according to the distribution parameters of the time-frequency positions and reference distribution parameters, aggregating the distribution offset of the time-frequency positions according to frequency dimensions to obtain the distribution offset of the time dimensions, and determining equipment anomaly detection results. The method solves the problem that the existing equipment abnormality detection mode based on the audio data is poor in accuracy, the audio signals are subjected to distribution offset measurement according to the reference distribution parameters of the positive samples, and the equipment abnormality can be accurately detected on the basis that the negative sample characteristics are not required to be referenced.

Inventors

ZHENG HAITING
JIA QILONG
XU WEI
WANG JIE
Liu Caixi
WANG XIAOYANG
YUAN LEI
WANG WEIWEI
WANG ZHENJIANG
ZHU XUANMING

Assignees

中国电气装备集团科学技术研究院有限公司

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (10)

1. A method for detecting device anomalies based on audio data, the method comprising: Acquiring target audio data, wherein the target audio data comprises at least one audio fragment to be detected, and the audio fragment is a time domain discrete audio signal; Extracting time-frequency characteristics of the target audio data according to the logarithmic Mel frequency spectrum of the target audio data, wherein the time-frequency characteristics comprise time, frequency and 3 characteristic dimensions of a channel; Extracting channel feature vectors matched with each time-frequency position from time-frequency features of target audio data, and determining distribution parameters of each time-frequency position according to the channel feature vectors matched with each time-frequency position, wherein the distribution parameters comprise a mean value vector and a covariance matrix; Calculating the distribution offset of each time-frequency position according to the distribution parameters of each time-frequency position and the reference distribution parameters of each time-frequency position acquired in advance, wherein the reference distribution parameters comprise a reference mean value vector and a reference covariance matrix of each time-frequency position determined based on the audio data of the normal operation period of the equipment; And aggregating the distribution offset of each time-frequency position according to the frequency dimension to obtain the distribution offset of the time dimension, and determining the equipment abnormality detection result according to the distribution offset of the time dimension.
2. The method of claim 1, wherein the number of audio clips in the target audio data is one; Extracting channel feature vectors matched with each time-frequency position from time-frequency features of the target audio data, and determining distribution parameters of each time-frequency position according to the channel feature vectors matched with each time-frequency position, wherein the method comprises the following steps: Extracting channel feature vectors matched with each time-frequency position from the time-frequency features of the audio fragment; And aiming at each time-frequency position, forming a statistical sample set by each time-frequency position in a preset time neighborhood of the time-frequency position, and determining the distribution parameters of the time-frequency position according to the channel feature vectors matched with each time-frequency position in the statistical sample set.
3. The method of claim 1, wherein the target audio data has at least two audio segments; Extracting channel feature vectors matched with each time-frequency position from time-frequency features of the target audio data, and determining distribution parameters of each time-frequency position according to the channel feature vectors matched with each time-frequency position, wherein the method comprises the following steps: Extracting channel feature vectors matched with each time-frequency position from the time-frequency features of each audio fragment; And aiming at each time-frequency position, forming a statistical sample set by each audio fragment, and determining the distribution parameters of the time-frequency positions according to the channel characteristic vectors matched with the time-frequency positions in each audio fragment.
4. The method of claim 1, wherein after extracting the time-frequency characteristics of the target audio data, the method further comprises: If the channel dimension in the time-frequency characteristic of the target audio data is larger than a preset dimension threshold, performing dimension reduction processing on the channel dimension of the time-frequency characteristic of the target audio data based on a preset dimension reduction mode, wherein the preset dimension reduction mode is one of random channel sampling, linear projection or principal component analysis.
5. The method of claim 1, wherein after determining the distribution parameters for each time-frequency location, the method further comprises: And for each time-frequency position, performing contraction processing on the covariance matrix of the time-frequency position based on a preset contraction coefficient and an identity matrix of a target dimension, wherein the dimension of the covariance matrix of the time-frequency position is the same as that of the target dimension.
6. The method according to claim 1, wherein calculating the distribution offset of each time-frequency location according to the distribution parameter of each time-frequency location and the reference distribution parameter of each time-frequency location acquired in advance includes: for each time-frequency position, calculating the Wasserstein distance of the time-frequency position according to the distribution parameter of the time-frequency position and the reference distribution parameter of the time-frequency position, and taking the Wasserstein distance of the time-frequency position as the distribution offset of the time-frequency position.
7. An apparatus abnormality detection device based on audio data, the device comprising: the audio data acquisition module is used for acquiring target audio data, wherein the target audio data comprises at least one audio fragment to be detected, and the audio fragment is a time domain discrete audio signal; The time-frequency characteristic extraction module is used for extracting the time-frequency characteristic of the target audio data according to the logarithmic Mel frequency spectrum of the target audio data, wherein the time-frequency characteristic comprises time, frequency and 3 characteristic dimensions of a channel; The system comprises a distribution parameter determining module, a distribution parameter determining module and a processing module, wherein the distribution parameter determining module is used for extracting channel characteristic vectors matched with each time-frequency position from time-frequency characteristics of target audio data and determining distribution parameters of each time-frequency position according to the channel characteristic vectors matched with each time-frequency position; The offset determining module is used for calculating the distribution offset of each time-frequency position according to the distribution parameter of each time-frequency position and the reference distribution parameter of each time-frequency position acquired in advance, wherein the reference distribution parameter comprises a reference mean value vector and a reference covariance matrix of each time-frequency position which are determined based on the audio data of the normal operation period of the equipment; The detection result determining module is used for aggregating the distribution offset of each time-frequency position according to the frequency dimension to obtain the distribution offset of the time dimension, and determining the equipment abnormality detection result according to the distribution offset of the time dimension.
8. An electronic device, the electronic device comprising: and a memory communicatively coupled to the at least one processor, wherein the memory stores a computer program executable by the at least one processor to enable the at least one processor to perform the audio data based device anomaly detection method of any one of claims 1-6.
9. A computer readable storage medium storing computer instructions for causing a processor to implement the audio data based device anomaly detection method of any one of claims 1-6 when executed.
10. A computer program product comprising a computer program which, when executed by a processor, implements the audio data based device anomaly detection method of any one of claims 1-6.

Description

Method, device, equipment, medium and product for detecting equipment abnormality based on audio data Technical Field The present invention relates to the field of data processing technologies, and in particular, to a method, an apparatus, a device, a medium, and a product for detecting an abnormality of a device based on audio data. Background The equipment anomalies may be reflected by audio data during operation of the equipment, such as bearing wear, structural looseness, and electrical faults. The acoustic sensor is arranged near the equipment to collect the audio data in the running process of the equipment, and the audio data is analyzed, so that the monitoring and fault early warning of the running state of the equipment can be realized under the condition of not contacting the equipment. At present, the existing equipment abnormality detection mode based on audio data mainly comprises the steps of (1) judging equipment abnormality according to the comparison result of the short-time energy, zero crossing rate and other characteristics of an audio signal and a corresponding threshold value, (2) training a deep learning model to perform characteristic learning on a large number of audio samples, and performing equipment abnormality detection by using the trained deep learning model. For the mode (1), the audio data in the running process of the equipment has obvious non-stationarity and time-frequency coupling characteristics, the frequency spectrum structure and energy distribution of sound under different working conditions can be changed, the expression capability of artificial design characteristics such as short-time energy, zero crossing rate and the like on complex time-frequency structure changes in audio signals is insufficient, and the difference between the normal state and the abnormal state of the equipment is difficult to stably describe, so that the abnormal detection precision of the equipment is influenced. For the mode (2), the occurrence probability of the abnormality of the whole operation life of the equipment is low, the type is complex, and the audio data of the equipment abnormality period is difficult to collect in a targeted manner, so that the audio sample data of the equipment abnormality is deficient, and the application of the supervised learning model in equipment abnormality detection is limited. Disclosure of Invention The invention provides a method, a device, equipment, a medium and a product for detecting equipment abnormality based on audio data, which are used for solving the problem of poor accuracy of the existing equipment abnormality detection mode based on the audio data. According to an aspect of the present invention, there is provided an apparatus abnormality detection method based on audio data, the method including: Acquiring target audio data, wherein the target audio data comprises at least one audio fragment to be detected, and the audio fragment is a time domain discrete audio signal; Extracting time-frequency characteristics of the target audio data according to the logarithmic Mel frequency spectrum of the target audio data, wherein the time-frequency characteristics comprise time, frequency and 3 characteristic dimensions of a channel; Extracting channel feature vectors matched with each time-frequency position from time-frequency features of target audio data, and determining distribution parameters of each time-frequency position according to the channel feature vectors matched with each time-frequency position, wherein the distribution parameters comprise a mean value vector and a covariance matrix; Calculating the distribution offset of each time-frequency position according to the distribution parameters of each time-frequency position and the reference distribution parameters of each time-frequency position acquired in advance, wherein the reference distribution parameters comprise a reference mean value vector and a reference covariance matrix of each time-frequency position determined based on the audio data of the normal operation period of the equipment; And aggregating the distribution offset of each time-frequency position according to the frequency dimension to obtain the distribution offset of the time dimension, and determining the equipment abnormality detection result according to the distribution offset of the time dimension. According to another aspect of the present invention, there is provided an apparatus abnormality detection apparatus based on audio data, the apparatus comprising: the audio data acquisition module is used for acquiring target audio data, wherein the target audio data comprises at least one audio fragment to be detected, and the audio fragment is a time domain discrete audio signal; The time-frequency characteristic extraction module is used for extracting the time-frequency characteristic of the target audio data according to the logarithmic Mel frequency spectrum of the target audio data, wherein the time