A partial discharge classification method for power equipment based on multi-channel acoustic signal space-time correlation analysis

Abstract

The application discloses a power equipment partial discharge classification method based on multi-channel sound signal space-time correlation analysis, which classifies the power equipment partial discharge type by respectively mining the space correlation and time correlation of the multi-channel sound signal. The application proposes a one-dimensional convolutional neural network for space-time correlation mining, which firstly mines the space correlation of the multi-channel sound signal, obtains space weight information and performs space correlation weighting on the multi-channel sound signal features; then mines the time correlation of each channel signal based on the space correlation mining, thereby obtaining time weight information and performing time correlation weighting on each channel sound signal; finally, the multi-channel sound signal subjected to the space and time correlation mining is further subjected to feature extraction and partial discharge type classification. The application can effectively classify the power equipment partial discharge based on the sound signal, thereby providing support for the fault diagnosis and prevention of the power equipment.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of electric power operation and safety, and particularly relates to a partial discharge classification method for power equipment based on multi-channel acoustic signal space-time correlation analysis. BACKGROUND

[0002] When partial discharge occurs in power equipment, it may cause insulation breakdown, aging, and even personal safety accidents. Therefore, timely detection of partial discharge in power equipment is of great significance to the safe operation of power equipment.

[0003] The generation of partial discharge is caused by the rapid heating and cooling of gas cavities or voids in the insulation layer. The acoustic signals generated by this phenomenon can be detected by acoustic sensors, and the location and type can be determined by analysis. Using acoustic signals for partial discharge detection has many advantages, such as non-invasiveness, high sensitivity, high accuracy, good repeatability, and low cost.

[0004] In practice, an acoustic sensor array is usually used to collect partial discharge signals. Specifically, multiple acoustic sensors are arranged at different positions according to certain standards and sampled continuously for a certain period of time. When analyzing these signals, two kinds of correlation should be considered: time correlation and space correlation. The first is time correlation, because acoustic signals are essentially time series, and there should be differences between the measured signals obtained during partial discharge and during no discharge. At the same time, different types of partial discharge also have differences in the process of changing over time. Therefore, it is necessary to analyze the correlation of acoustic signals at different sampling times in the time dimension, extract the correlation relationship, and thus detect and analyze the type of partial discharge. In addition, signals from different channels or sensors may also have certain spatial correlation. Acoustic sensors at different positions collect information about the same partial discharge signal from different angles, and their signals can complement and enhance each other.

[0005] In summary, when using an acoustic array to detect partial discharge in power equipment, the time correlation and spatial correlation between multiple acoustic sensors should be fully considered in order to more accurately identify the partial discharge phenomenon and the type of partial discharge in power equipment. SUMMARY

[0006] The present application provides a partial discharge classification method for power equipment based on multi-channel acoustic signal space-time correlation analysis, which solves the problem of insufficient consideration of multi-channel acoustic signal data correlation in existing partial discharge detection methods for power equipment based on acoustic signals, low detection accuracy, and difficulty in meeting the requirements of safe and stable operation of power equipment.

[0007] The technical solutions adopted by the present application to solve its technical problems are as follows:

[0008] A power equipment partial discharge classification method based on multi-channel sound signal space-time correlation analysis, comprising the following steps:

[0009] S1, spatial correlation mining of multi-channel sound signals is performed to obtain spatial correlation weights;

[0010] S2, spatial correlation weighting of multi-channel sound signals is performed;

[0011] S3, time correlation mining of multi-channel sound signals weighted by spatial correlation is performed to obtain time correlation weights:

[0012] S4, time correlation weighting of each channel sound signal is performed;

[0013] S5, further feature extraction is performed on the multi-channel sound signals subjected to spatial and time correlation mining, and the power equipment partial discharge is classified.

[0014] In the step S1, the spatial correlation mining of multi-channel sound signals mainly includes the following key steps:

[0015] S11, the n-channel p-length multi-channel sound signal is arranged into an X n×p multi-time sequence matrix;

[0016] S12, X n×p is subjected to 32 one-dimensional convolutional neural networks with a size of 1 and a step of 1 to perform information feature extraction in the spatial dimension, to obtain a feature X n×32 with a size of n*32;

[0017] S13, a global average pooling operation is performed on X n×32 in the channel direction, each sound sensor being a channel, to obtain a feature X n with a size of n;

[0018] S14, the correlation between channels in X n is extracted by two full connection layers with a size of n respectively, to obtain spatial correlation weights S w .

[0019] In the step S2, the spatial correlation weighting of multi-channel sound signals is as follows, using the spatial correlation weights S w to weight X n×p to obtain The specific calculation method is

[0020] In the step S3, when the time correlation mining of multi-channel sound signals is performed, the following key steps are mainly included:

[0021] S31, After 32 one-dimensional convolutional neural networks with a size of 5 and a step of 5, information feature extraction is carried out in the time dimension, and a feature X with a size of 32*(p / 5) is obtained 32×(p / 5) ;

[0022] S32, X 32×(p / 5) Global average pooling operation is carried out in the filter direction, and as known from step S31, 32 filters with a size of 5*5 are used, and a feature X with a size of p / 5 is obtained p / 5 ;

[0023] S33, X n×(p / 5) is subjected to two fully connected layers with a size of p / 10 and p / 5 respectively, and the correlation between each time period is extracted to obtain a time correlation weight T w .

[0024] In the step S4, the time correlation of each channel sound signal is weighted as follows:

[0025] The time correlation weight T w is used to weight X n×(p / 5) to obtain The specific calculation method is

[0026]

[0027] In the step S5, two fully connected layers with a size of 1024 are used to extract features from X respectively, and finally, a fully connected layer with a size of c and a softmax are used to classify the partial discharge type of the power equipment.

[0028] The beneficial effects that can be achieved by the technical solutions of the present application are: the present application fully considers the spatial and temporal correlation of multi-channel sound signals, realizes the detection and classification of power equipment partial discharge based on multi-channel sound signals, and can provide effective decision support for the operation and maintenance of power equipment. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The model schematic diagram of the present application. DETAILED DESCRIPTION

[0030] The present application will be further described below by combining the accompanying drawings Figure 1 The present application will be further described below by combining the accompanying drawings

[0031] The application provides a power equipment partial discharge classification method based on multi-channel sound signal space-time correlation analysis, realizes detection and classification of power equipment partial discharge by mining the space-time correlation of multi-channel sound signals.

[0032] In the embodiment of the application, gap discharge, corona discharge and spherical discharge of power equipment are simulated in a laboratory respectively, and 16-channel sound arrays are used for sampling at a frequency of 44.1 kHz. The collected samples are divided into samples with a size of 16*4000, and the samples are divided into four categories of no partial discharge phenomenon, gap discharge, corona discharge and spherical discharge by considering that no partial discharge phenomenon occurs in the early stage of each simulation test. The data distribution of each category is shown in Table 1.

[0033] Table 1 Data distribution

[0034]

[0035] The specific steps are as follows:

[0036] The model training parameters are set, wherein the learning rate is 0.0001, the RMSprop optimization algorithm is used, and rho is 0.9.

[0037] Step 1: Mining the space correlation of multi-channel sound signals and obtaining the space correlation weight. In this embodiment, a 16-channel sound array is used, and the constructed sample length is 4000, so the input model data is X 16×4000 After 32 one-dimensional convolutional neural networks with a size of 1 and a step of 1, information features are extracted in the spatial dimension to obtain a feature X 16×32 with a size of 16*32; a global average pooling operation is performed in the channel direction (each sound sensor is a channel) to obtain a feature X 16 with a size of 16; and the correlation between channels in X 16 is extracted by two fully connected layers with a size of 16 to obtain the space correlation weight S w .

[0038] Step 2, spatial correlation weighting is performed on the multi-channel sound signal, and a spatial correlation weight S is used w X is weighted to obtain 16×4000

[0039] Step 3, time correlation mining is performed on the multi-channel sound signal subjected to spatial correlation weighting, and a time correlation weight T is obtained After 32 one-dimensional convolutional neural networks with a size of 5 and a step of 5, information feature extraction is performed in the time dimension, and a feature X with a size of 32x800 is obtained 32×800 Global average pooling operation is performed in the filter direction (as can be known from step S31, a total of 32 filters with a size of 5x5 are used), and a feature X with a size of 800 is obtained 800 Two fully connected layers with sizes of 400 and 800 are used to extract the correlation between time periods in X 800 , and a time correlation weight T is obtained w .

[0040] Step 4, time correlation weighting is performed on each channel sound signal, and a time correlation weight T is used w X is weighted to obtain 32×800

[0041] Step 5, further feature extraction is performed on the multi-channel sound signal subjected to spatial and time correlation mining, and the feature is characterized in that two layers of fully connected layers with a size of 1024 are used to extract features from X , and finally, a fully connected layer with a size of 4 and a softmax are used to classify the type of partial discharge of the power equipment.

[0042] In this embodiment, the evaluation indexes for model testing are precision (P), recall (R), and F1 score. In order to verify the necessity of spatial and time correlation extraction of the present application, three types of models considering spatial and time correlation, only time correlation, and only spatial correlation are tested respectively. The comparison results are shown in Tables 2 to 5.

[0043] Table 2: Considering spatial and time correlation

[0044]

[0045] Table 3: Considering time correlation

[0046]

[0047] Table 4: Considering spatial correlation

[0048] ​​

[0049] Table 5 Comparison of three methods

[0050]

[0051] As can be seen from Tables 2 to 5, the best results are obtained when both time and spatial correlation are considered. When spatial correlation is not considered, the precision, recall and F1 score all decrease by 2%. When time correlation is not considered, the precision, recall and F1 decrease by 52%, 27% and 73% respectively. This shows that both time and spatial correlation are important in the classification of partial discharge in power equipment based on multi-channel acoustic signals. However, time correlation is more important because the network performance deteriorates significantly if this is not considered.

[0052] The specific embodiments described herein are merely illustrative of the spirit of the application. Various modifications or changes in addition or substitution to the described specific embodiments can be made by those skilled in the art without departing from the spirit of the application or exceeding the scope of the appended claims.

Claims

1. A method for classifying partial discharge of power equipment based on spatiotemporal correlation analysis of multi-channel acoustic signals, characterized in that, Includes the following steps: S1. Perform spatial correlation mining on multi-channel acoustic signals and obtain spatial correlation weights; In step S1, the spatial correlation mining of multi-channel acoustic signals mainly includes the following key steps: S11. Arrange the n-channel, p-length multi-channel acoustic signals into X. n×p Multi-time series matrix; S12, X n×p After passing through 32 one-dimensional convolutional neural networks with a size of 1 and a stride of 1, information feature extraction is performed in the spatial dimension to obtain a feature X of size n×32. n×32 ; S13, Regarding X n×32 Global average pooling is performed along the channel direction, with each acoustic sensor representing one channel, to obtain a feature X of size n. n ; S14, X passes through two fully connected layers of size n respectively. n The correlation between each channel is extracted to obtain the spatial correlation weight S. w ; S2. Spatial correlation weighting of multi-channel acoustic signals; In step S2, the multi-channel acoustic signals are spatially correlated and weighted as follows, using spatial correlation weight S. w For X n×p Weighted average The specific calculation method is as follows: S3. Perform time correlation mining on the spatially correlated multi-channel acoustic signals and obtain the time correlation weights: In step S3, the time correlation mining method for multi-channel acoustic signals mainly includes the following key steps: S31, will After passing through 32 one-dimensional convolutional neural networks with a size of 5 and a stride of 5, information feature extraction is performed in the time dimension to obtain a feature X of size 32×(p / 5). 32×(p / 5) ; S32, Regarding X 32×(p / 5) A global average pooling operation is performed along the filter direction. As shown in step S31, a total of 32 filters of size 5×5 are used to obtain a feature X of size p / 5. p / 5 ; S33, X passes through two fully connected layers of sizes p / 10 and p / 5 respectively. n×(p / 5) The correlation between different time periods is extracted to obtain the time correlation weight T. w ; S4. Perform time-correlation weighting on the acoustic signals of each channel; In step S4, the time-correlation weighting of the acoustic signals in each channel is performed as follows: using the time-correlation weight T w For X n×(p / 5) Weighted average The specific calculation method is as follows: S5. Further feature extraction is performed on the multi-channel acoustic signals after spatial and temporal correlation mining, and partial discharge of power equipment is classified. In step S5, the multi-channel acoustic signal, after spatial and temporal correlation mining, undergoes further feature extraction, passing through two fully connected layers of size 1024. Feature extraction is performed, and finally, the partial discharge types of power equipment are classified through a fully connected layer of size c and a softmax function.

Images