CN-122024765-A - Electrical equipment abnormality detection method and device, electrical equipment, medium and product

CN122024765ACN 122024765 ACN122024765 ACN 122024765ACN-122024765-A

Abstract

The invention relates to the technical field of intelligent home, and discloses an electrical equipment anomaly detection method, an electrical equipment, a medium and a product, wherein the invention collects the original audio data of the electrical equipment, and the multi-scale time domain features, the logarithmic mel spectrogram features and the amplitude spectrogram features are extracted to carry out cascade fusion of channel dimensions, so that the comprehensive characterization of the running state of the equipment from the time domain, the frequency domain and the time-frequency domain in multi-dimensions is realized, and the richness and the discrimination of the features are obviously improved. The channel and spatial position self-adaptive weighting is carried out on the composite feature map by utilizing the pre-trained feature enhancement network, so that key abnormal features can be effectively focused, background noise interference can be restrained, and the robustness of the model under a complex environment and the sensitivity to weak abnormal signals can be greatly enhanced. Finally, the abnormal probability prediction is realized through the classifier, and normal working sounds and various potential fault sounds can be effectively distinguished while high detection precision is ensured, so that the accuracy of abnormal recognition is improved.

Inventors

HE ZIJIAN
YAN WANG
ZHANG XIAOCI
LIAO YONGPING
Zheng Linfu

Assignees

珠海格力电器股份有限公司

Dates

Publication Date: 20260512
Application Date: 20260304

Claims (12)

1. An electrical equipment anomaly detection method, characterized in that the method comprises: collecting original audio data of the running electrical equipment; Extracting multi-scale time domain features, logarithmic mel spectrogram features and amplitude spectrogram features of the original audio data, and carrying out cascade fusion on the multi-scale time domain features, the logarithmic mel spectrogram features and the amplitude spectrogram features in channel dimension to obtain a composite feature map; Weighting each channel and each spatial position of the composite feature map by utilizing a pre-trained feature enhancement network to obtain a deep feature vector; and inputting the deep feature vector into a pre-trained classifier, and predicting the probability that the original audio data belong to various sound categories by using the classifier to obtain an abnormality detection result.
2. The method for detecting anomaly of electrical equipment according to claim 1, wherein the feature enhancement network comprises a convolution block attention module and a parameter-free attention module, wherein the weighting each channel and each spatial position of the composite feature map by the pre-trained feature enhancement network to obtain a deep feature vector comprises: Inputting the composite feature map to the convolution block attention module to obtain a first channel weight of the composite feature map in each channel and a spatial weight of each spatial position, and obtaining an intermediate feature vector according to the first channel weight, the spatial weight and the composite feature map; and determining a second channel weight of the intermediate feature vector in each channel by using the parameter-free attention module, and carrying out channel weighting on the intermediate feature vector by using the second channel weight to obtain a deep feature vector.
3. The method for detecting abnormality of electrical equipment according to claim 2, wherein the convolution block attention module includes a channel attention branch and a spatial attention branch, the inputting the composite feature map to the convolution block attention module obtains a first channel weight of the composite feature map in each channel and a spatial weight of each spatial position, and obtains an intermediate feature vector according to the first channel weight, the spatial weight and the composite feature map, the method comprises: carrying out global average pooling and maximum pooling on the composite feature map in a space dimension by utilizing the channel attention branches to generate a first channel weight of each channel; Weighting each channel of the composite feature map based on the first channel weight to obtain a channel weighted feature map, and carrying out global average pooling and convolution on the channel weighted feature map in a channel dimension by utilizing the spatial attention branches to generate the spatial weight of each spatial position; and weighting each spatial position of the channel weighted feature map based on the spatial weight to obtain an intermediate feature vector.
4. The electrical device anomaly detection method of claim 2, wherein determining the second channel weight of the intermediate feature vector at each channel using the parameter-less attention module comprises: Inputting the intermediate feature vector to the parameter-free attention module, and calculating the energy value of each neuron based on the feature mean value and the feature standard deviation of the channel where each neuron is located in the intermediate feature vector; For each channel of the intermediate feature vector, a second channel weight for the channel is determined based on energy values of all neurons in the channel.
5. The electrical device anomaly detection method of any one of claims 1-4, wherein the extracting the multi-scale time-domain features of the raw audio data comprises: Inputting the original audio data into a pre-trained multi-scale time domain feature extraction network to obtain multi-scale time domain features, wherein the multi-scale time domain feature extraction network comprises a plurality of expansion convolution layers with gradually increased expansion rate, and each expansion convolution layer is sequentially connected with a batch normalization layer and an activation function layer.
6. The electrical device anomaly detection method of any one of claims 1-4, wherein the training process of the classifier comprises: Acquiring an original training sample and an original classifier model, wherein the original training sample comprises sample audio data of different sound categories; Performing linear interpolation on sample audio data of any two sound categories in the original training sample to obtain new sample audio data, and performing data enhancement on the original training sample by using the new sample audio data to obtain an enhanced training sample; And training the original classifier model by using the enhanced training sample to obtain a pre-trained classifier.
7. The electrical device anomaly detection method of any one of claims 1-4, wherein prior to weighting each channel and each spatial location of the composite feature map with a pre-trained feature enhancement network to obtain deep feature vectors, the method further comprises: performing dimension reduction treatment on the composite feature map; Before inputting the deep feature vector into the pre-trained classifier, the method further comprises: and carrying out feature compression on the deep feature vector.
8. An electrical equipment anomaly detection device, characterized in that the device comprises: the first processing module is used for collecting original audio data when the equipment runs; The second processing module is used for extracting multi-scale time domain features, logarithmic mel spectrogram features and amplitude spectrogram features of the original audio data, and carrying out cascade fusion on the multi-scale time domain features, the logarithmic mel spectrogram features and the amplitude spectrogram features in a channel dimension to obtain a composite feature map; The third processing module is used for weighting each channel and each space position of the composite feature map by utilizing a pre-trained feature enhancement network to obtain a deep feature vector; and the fourth processing module is used for inputting the deep feature vector into a pre-trained classifier, and predicting the probability that the original audio data belong to various sound categories by using the classifier to obtain an abnormality detection result.
9. An electrical device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory having stored therein computer instructions, the processor executing the computer instructions to perform the electrical device anomaly detection method of any one of claims 1 to 7.
10. The electrical appliance according to claim 9, wherein the electrical appliance is a range hood.
11. A computer-readable storage medium, characterized in that the computer-readable storage medium has stored thereon computer instructions for causing a computer to execute the electrical device abnormality detection method according to any one of claims 1 to 7.
12. A computer program product comprising computer instructions for causing a computer to perform the electrical device anomaly detection method of any one of claims 1 to 7.

Description

Electrical equipment abnormality detection method and device, electrical equipment, medium and product Technical Field The invention relates to the technical field of intelligent home, in particular to an electrical equipment abnormality detection method, an electrical equipment abnormality detection device, an electrical equipment, a medium and a product. Background With the rapid development of intelligent household technology, the range hood is used as one of kitchen core appliances, and the stability and safety of the running state of the range hood are increasingly focused by users. However, in the actual use process, the range hood often generates abnormal sounds due to motor wear, bearing aging, blockage of foreign matters of a fan, blockage of an air duct, circuit faults and the like. If the anomalies are not found and processed in time, the cooking experience can be affected, equipment damage can be caused, and even potential safety hazards such as fire disaster and the like can be caused. The traditional detection method of abnormal sounds of the range hood mainly depends on manual experience or rule judgment based on simple acoustic characteristics, and is difficult to comprehensively capture complex and changeable sound signals generated by the range hood under different working conditions, so that the abnormal modes of the range hood are incompletely and inaccurately identified. In addition, the traditional deep learning model lacks an effective characteristic selection and enhancement mechanism, and can not automatically distinguish key abnormal information and redundant noise in signals, and the detection false alarm rate and the omission rate of abnormal sounds of the range hood are high. Disclosure of Invention The invention provides an electrical equipment abnormality detection method, an electrical equipment abnormality detection device, an electrical equipment, a medium and a product, and aims to solve the problems that in the prior art, abnormal recognition of a range hood is incomplete and the accuracy is low. In a first aspect, the present invention provides a method for detecting abnormality of electrical equipment, the method comprising: collecting original audio data of the running electrical equipment; Extracting multi-scale time domain features, logarithmic mel spectrogram features and amplitude spectrogram features of original audio data, and carrying out cascade fusion on the multi-scale time domain features, the logarithmic mel spectrogram features and the amplitude spectrogram features in channel dimension to obtain a composite feature map; Weighting each channel and each space position of the composite feature map by utilizing a pre-trained feature enhancement network to obtain a deep feature vector; and inputting the deep feature vector into a pre-trained classifier, and predicting the probability that the original audio data belong to various sound categories by using the classifier to obtain an abnormality detection result. According to the invention, the original audio data of the electrical equipment are collected, the multi-scale time domain features, the logarithmic mel spectrogram features and the amplitude spectrogram features of the electrical equipment are extracted to carry out cascade fusion of channel dimensions, so that the comprehensive characterization of the running state of the equipment from the time domain, the frequency domain and the time-frequency domain is realized, and the richness and the discrimination of the features are obviously improved. The channel and spatial position self-adaptive weighting is carried out on the composite feature map by utilizing the pre-trained feature enhancement network, so that key abnormal features can be effectively focused, background noise interference can be restrained, and the robustness of the model under a complex environment and the sensitivity to weak abnormal signals can be greatly enhanced. Finally, the abnormal probability prediction is realized through the classifier, and normal working sounds and various potential fault sounds can be effectively distinguished while high detection precision is ensured, so that the accuracy of abnormal recognition is improved. In some alternative embodiments, the feature enhancement network includes a convolution block attention module and a parameter-free attention module, and weighting each channel and each spatial location of the composite feature map with the pre-trained feature enhancement network to obtain deep feature vectors, comprising: inputting the composite feature map to a convolution block attention module to obtain a first channel weight of the composite feature map in each channel and a space weight of each space position, and obtaining an intermediate feature vector according to the first channel weight, the space weight and the composite feature map; And determining a second channel weight of the intermediate feature vector in each channel by using the par