Electric energy quality disturbance signal identification method and device based on double-path feature fusion network

Abstract

The invention relates to an electric energy quality disturbance signal identification method and device based on a double-path feature fusion network. Belongs to the technical field of electric energy quality analysis and monitoring. The method comprises the steps of collecting power quality disturbance signals in a power system; constructing a double-path feature image; one path is used for performing S transformation on the power quality disturbance signal to obtain an S transformation matrix graph; the other path is used for performing periodic reconstruction on the power quality disturbance signal to obtain a periodic reconstruction matrix graph; inputting the two-way feature image into the trained two-way feature fusion network, and performing power quality disturbance signal identification; and the two-way feature fusion network performs feature extraction, feature fusion and classification on the input S transformation matrix graph and the periodic reconstruction matrix graph, and outputs a classification identification result of the disturbance event in the power system. The method can effectively overcome the defects that in the prior art, the accuracy of electric energy quality disturbance signal recognition is low, and the model structure is complex.

Description

Technical Field

[0001] The present invention relates to the technical field of power quality analysis and monitoring, and in particular to a method and device for identifying power quality disturbance signals based on a dual-path feature fusion network. Background Art

[0002] In modern society, the power system is an important pillar of industry and residents' life, and power quality is the key to ensure the safe and reliable operation of the power system. The power quality disturbance (PQD) caused by nonlinear loads and large-scale distributed grid connection will seriously affect the safe and stable operation of precision computers and microprocessors in the power grid, and in some cases may cause serious safety accidents and economic losses. Therefore, the accurate identification and classification of power quality disturbance signals are of great significance for improving power supply quality, monitoring the status of power equipment, and predicting and eliminating faults.

[0003] As the complexity of PQD increases, traditional methods can no longer meet the classification needs of PQD in various scenarios. Deep learning methods mainly use convolutional neural networks (CNN) and recurrent neural networks (RNN) to simulate the structure and self-training ability of the human brain to complete autonomous learning of a large amount of PQD sample data, which can be applied to the identification of complex PQDs. It has become the current mainstream method to convert one-dimensional signals into two-dimensional images through signal processing methods, and then combine deep learning image processing models to extract features from two-dimensional images to complete PQD classification.

[0004] The time-frequency graph obtained by time-frequency analysis is identified by using a neural network. Although this method has achieved certain results, the single time-frequency feature has low discrimination. Qian Beiqi et al. first used the Markov Transition Field (MTF) to convert the time series data of the PQD signal into image data, and then input the image data into the CNN convolutional neural network for feature extraction. Finally, the key features were extracted and PQD classification was performed. Experiments show that this method has better classification effect and anti-noise ability. In addition, some people have proposed a method for composite disturbance identification using Gram Angular Fields (GAF) and deep residual network (Residual Network, ResNet). Experiments show that this method has a high accuracy rate for composite disturbance identification. However, the network parameters of the above two methods are very large, which increases the difficulty of model training and requires high computing resources. Summary of the invention

[0005] In view of the above analysis, the present invention aims to provide a method and device for identifying power quality disturbance signals based on a dual-path feature fusion network, which can effectively overcome the defects of the prior art in terms of low accuracy in identifying power quality disturbance signals and a relatively complex model structure.

[0006] The purpose of the present invention is mainly achieved through the following technical solutions:

[0007] On the one hand, the present invention discloses a method for identifying power quality disturbance signals based on a dual-path feature fusion network, comprising the following steps:

[0008] Step S1, collecting power quality disturbance signals in the power system;

[0009] Step S2, constructing a dual-path feature image; one path is to perform S transformation on the power quality disturbance signal to obtain an S transformation matrix diagram; the other path is to perform periodic reconstruction on the power quality disturbance signal to obtain a periodic reconstruction matrix diagram;

[0010] Step S3, inputting the dual-path feature image into the trained dual-path feature fusion network to perform power quality disturbance signal recognition;

[0011] The dual-path feature fusion network performs feature extraction, feature fusion and classification on the input S-transformation matrix diagram and the periodic reconstruction matrix diagram respectively, and outputs the classification and recognition results of the disturbance events in the power system.

[0012] On the other hand, the present invention also discloses a power quality disturbance signal recognition device based on a dual-path feature fusion network, comprising: a dual-path feature image generation module and a dual-path feature fusion module;

[0013] The dual-path feature image generation module performs S transformation and period reconstruction on the collected power quality disturbance signal in the power system to obtain an S transformation matrix diagram and a period reconstruction matrix diagram;

[0014] The dual-path feature fusion module has a built-in trained dual-path feature fusion network, which extracts features from the input S-transformation matrix diagram and the periodic reconstruction matrix diagram respectively, performs feature fusion and classification, and outputs classification and recognition results of disturbance events in the power system.

[0015] The beneficial effects of the present invention are as follows:

[0016] Compared with the prior art, the power quality disturbance signal recognition method and device based on the dual-path feature fusion network provided by the present invention convert the one-dimensional power quality disturbance signal into an S-transformation matrix diagram and a periodic reconstruction matrix diagram respectively, use the ShuffterNet_V2 lightweight convolutional neural network to extract features of the two images respectively, then use the multi-head attention mechanism module to perform feature fusion, and finally the classification layer performs feature recognition to obtain the power quality disturbance signal recognition result. At the same time, the present invention can accurately identify single and compound power quality disturbances in a high noise environment through dual-path input features and dual-path feature fusion of the ShuffterNet_V2 lightweight convolutional neural network, and the dual-path feature fusion network used has a simple structure, is convenient for model training, and does not require high computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0018] Figure 1 It is a flow chart of a method for identifying power quality disturbance signals based on a dual-path feature fusion network in an embodiment of the present invention;

[0019] Figure 2 This is a pseudo-color image of an S transformation matrix obtained by performing S transformation on 17 types of power quality disturbance signals in an embodiment of the present invention;

[0020] Figure 3 It is a pseudo-color image of an S-periodic reconstruction matrix obtained by periodically reconstructing 17 types of power quality disturbance signals in an embodiment of the present invention;

[0021] Figure 4 A schematic diagram of model training of a dual-path feature fusion network in an embodiment of the present invention;

[0022] Figure 5 It is a schematic diagram of a dual-path feature fusion network structure and a data processing flow in an embodiment of the present invention;

[0023] Figure 6 The experimental confusion matrix of the two-way feature fusion network based on the ShuffterNet_V2 lightweight convolutional neural network feature extractor;

[0024] Figure 7 The experimental confusion matrix of the model based on the feature extractor of ResNet18 residual network;

[0025] Figure 8 The experimental confusion matrix of the model based on the feature extractor of the EfficientNet_B0 convolutional neural network;

[0026] Fig. 9 Experimental confusion matrix for the model based on the MobileNet_V3 lightweight convolutional neural network feature extractor.

[0027] Fig.10 The present invention is a schematic diagram showing the connection components of a power quality disturbance signal identification device based on a dual-path feature fusion network in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used to illustrate the principles of the present invention together with the embodiments of the present invention.

[0029] Embodiment 1

[0030] This embodiment discloses a method for identifying power quality disturbance signals based on a dual-path feature fusion network. Figure 1 As shown, including:

[0031] Step S1, collecting power quality disturbance signals in the power system;

[0032] Step S2, constructing a dual-path feature image; one path is to perform S transformation on the power quality disturbance signal to obtain an S transformation matrix diagram; the other path is to perform periodic reconstruction on the power quality disturbance signal to obtain a periodic reconstruction matrix diagram;

[0033] Step S3, inputting the dual-path feature image into the trained dual-path feature fusion network to perform power quality disturbance signal recognition;

[0034] The dual-path feature fusion network performs feature extraction, feature fusion and classification on the input S-transformation matrix diagram and the periodic reconstruction matrix diagram respectively, and outputs the classification and recognition results of the disturbance events in the power system.

[0035] Specifically, the power quality disturbance signal collected in step S1 is time series data, which records the operation status of the power grid and reflects the instantaneous changes of key physical quantities such as voltage and current on the time axis. These changes may be caused by various factors in the power system, such as rapid changes in load, faults within the power grid, harmonic interference caused by nonlinear loads, etc.

[0036] Specifically, in step S2, it includes:

[0037] 1) The power quality disturbance signal is processed locally in the time and frequency domain by S transform, and the signal is analyzed at different times and frequencies by changing the window function and modulation function to obtain the S transform matrix diagram;

[0038] 2) According to the maximum frequency component in the power spectrum density of the power quality disturbance signal, the original signal is periodically reconstructed, and the original signal is transformed into a two-dimensional matrix to obtain a periodic reconstruction matrix diagram.

[0039] S transform is a time-frequency analysis method that combines the advantages of short-time Fourier transform and wavelet transform, and can provide a time-frequency representation of a signal.

[0040] The power quality disturbance signal is processed locally in the time-frequency domain by S transform, and the signal is analyzed at different times and frequencies by changing the window function and modulation function to obtain the S transform matrix diagram; the window function adopts the Gaussian window function, and the modulation function adopts the linear modulation function. The S transform can better capture the nonlinear and non-steady-state characteristics of the signal, so it has great advantages in processing nonlinear signals and transient signals.

[0041] In step S2, a more specific periodic reconstruction matrix diagram generation process includes:

[0042] 1) Segment the input power quality disturbance signal time series data, perform window function processing, Fourier transform and average processing on each signal segment to obtain the power spectrum density of the power quality disturbance signal;

[0043] The time series data of the input power quality disturbance signal is divided into multiple segments, and the length of each segment should be sufficient to contain one or more cycles; a window function, such as a Hanning window or a Hamming window, is applied to each segment of the signal to reduce the spectrum leakage effect. The signal processed by the window function is subjected to a fast Fourier transform (FFT) to obtain the spectrum of the signal. The spectrum of all segments is averaged to obtain the power spectral density (PSD) of the power quality disturbance signal.

[0044] 2) determining the maximum frequency component in the power spectrum density, and calculating the reconstructed period length of the power quality disturbance signal according to the period of the maximum frequency component;

[0045] The power spectrum density is analyzed to find the maximum frequency component therein; the maximum frequency component corresponds to the main periodic component of the power quality disturbance signal. According to the maximum frequency component, its corresponding period is calculated, and this period is used as the reconstruction period length.

[0046] 3) Calibrate and fill the power quality disturbance signal according to the reconstruction period length, so that the length of the reconstructed signal is an integer multiple of the reconstruction period length, transform the original signal into a two-dimensional matrix, and obtain a periodic reconstruction matrix diagram.

[0047] During signal calibration and padding, the original signal is calibrated according to the calculated reconstruction period length, that is, the signal length is adjusted to be an integer multiple of the reconstruction period length. If the original signal length is insufficient, it can be supplemented by copying, interpolating or padding with zeros.

[0048] The calibrated and filled signal is transformed into a two-dimensional matrix, where each row represents the signal value of an integer period; using the two-dimensional matrix data, a period reconstruction matrix diagram is drawn. In this diagram, each row corresponds to a signal of an integer period, which can intuitively represent the periodic characteristics of the signal and possible disturbances.

[0049] The complexity of power quality disturbance signals in power systems lies in their inherent multi-periodic characteristics, that is, multiple periodic patterns of different frequencies are intertwined and influence each other at the same time, resulting in extremely high complexity in the timing changes of the signal.

[0050] In the traditional one-dimensional time series analysis method, due to its single-dimensional characteristics, it is easy to ignore or confuse the timing changes of these different periods, which may seriously affect the effectiveness and accuracy of the analysis results. Therefore, according to the maximum frequency component in the power spectral density (PSD) of the power quality disturbance signal, the original signal is reconstructed periodically and transformed into a two-dimensional matrix, which can deeply explore the multi-periodic characteristics of the power quality disturbance signal.

[0051] In power quality analysis, the significant frequency components in the power spectrum density are crucial for understanding and processing power quality disturbance signals. The information in the power spectrum density reveals the frequency part with the most concentrated power in the power quality disturbance signal, that is, the frequency component with the largest power. This frequency component implies the inherent periodic nature of the power quality disturbance signal. Through this frequency component, the reconstructed period length of the power quality disturbance signal can be calculated. After calculating the reconstructed period length, the reconstructed period length of the power quality disturbance signal is calibrated and filled so that the length of the reconstructed signal is an integer multiple of the reconstructed period length.

[0052] The above process aims to preserve the periodic length characteristics of signal reconstruction while avoiding the introduction of non-original frequency components in the signal reconstruction process, thus ensuring the original periodic integrity of the power quality disturbance signal. This method ensures that in the signal processing process, the reconstructed signal not only retains the time characteristics of the original signal, but also introduces the periodic dimension, so that the signal changes can be observed in different periods, which is of great significance for capturing periodic disturbances (such as harmonics, flicker, etc.), and provides richer and more discriminating features for the identification of power quality disturbances.

[0053] Specifically, in step S3, the dual-path feature fusion network includes a feature extraction module, a feature fusion module and a classification module; wherein,

[0054] The feature extraction module is used to extract features from the input S-transformation matrix diagram and the periodic reconstruction matrix diagram through two lightweight convolutional neural networks, and splice and output the extracted two-way features to obtain the extracted power quality disturbance signal features;

[0055] More specifically, the feature extraction module includes two branches of ffterNet_V2 lightweight convolutional neural network for feature extraction; a single branch corresponds to the input S-transformation matrix diagram and the periodic reconstruction matrix diagram for feature extraction; and the extracted features are flattened and spliced ​​to form a set of feature sequences;

[0056] The feature fusion module adopts a multi-head attention mechanism to enhance the ability to capture and understand the complex interactions between features in the extracted feature sequence, and achieves in-depth fusion of features by learning and fusing information in parallel in multiple subspaces, and outputs the fused features;

[0057] The feature fusion module adopts a multi-head attention mechanism to perform feature fusion, maps the input feature sequence through three sets of weight matrices, and generates three sets of vectors of query Q, key K and value V respectively; divides the three sets of vectors of query Q, key K and value V into multiple heads, each head processes a subset of the vectors, each head obtains the attention weight by calculating the dot product of query Q and key K, and uses these attention weights to weight the value V to obtain the output result of each head; merges the output results of each head, and integrates them using a linear transformation layer to obtain a fused feature.

[0058] The classification module is used to perform separate recognition tasks on the fused features and output classification and recognition results.

[0059] ShuffterNet_V2 is an advanced lightweight neural network architecture used in the feature extraction module, which improves computational efficiency through innovative channel segmentation technology. This technology divides the feature channel from the input into two parts, one of which is directly connected to the output, while the other is transformed through a process of deep separable convolution to reduce the number of network parameters and computational complexity.

[0060] The depthwise separable convolution operation disassembles the traditional convolution process. First, each channel is spatially convolved independently, and then the output is fused through 1*1 convolution, which not only reduces the size of the model but also maintains the efficiency of the operation. Then, the segmented channels are reorganized by channel shuffling, ensuring that the information of different channels in the network can interact with each other, enhancing the model's ability to capture features. At the same time, ShuffleNet_V2 also incorporates a cleverly designed downsampling module, which aims to reduce the size of the feature map while appropriately increasing the number of channels, thereby ensuring smooth information transmission and minimizing performance loss. By integrating point-by-point convolution technology into the network design, ShuffleNet_V2 effectively fuses features between branches and balances the relationship between model size and information retention. Therefore, ShuffleNet_V2 achieves a good balance between speed, accuracy, and efficiency.

[0061] The multi-head attention mechanism used in the feature fusion module is an important mechanism in the field of natural language processing. It is used to capture the relationship between elements in sequence data. Its core idea is to learn information in different subspaces at the same time and combine this information to obtain a richer and more comprehensive representation. Such a mechanism enables the model to effectively capture the multivariate attributes of data in different feature subspaces and fuse these features in multiple dimensions to finally obtain a comprehensive output representation.

[0062] Specifically, the training process of the dual-path feature fusion network includes:

[0063] 1) Randomly generate time series data of multiple power quality disturbance signals, perform S transformation and period reconstruction on each power quality disturbance signal, obtain the corresponding S transformation matrix diagram and period reconstruction matrix diagram, and construct a data set;

[0064] 2) Divide the data set into training set, validation set and test set according to the preset ratio;

[0065] 3) Use the training set to train the dual-path feature fusion network model, and use the validation set and test set to verify and test the model respectively.

[0066] In the technical solution of this embodiment, the Matlab mathematical simulation model can be used to generate 17 kinds of time series data of power quality disturbance signals using random parameters, and each power quality disturbance signal is subjected to S transformation and period reconstruction to obtain the corresponding S transformation matrix diagram and period reconstruction matrix diagram, such as Figure 2 , Figure 3 As shown, a dataset is constructed and divided into a training set, a validation set, and a test set in a ratio of 6:2:2.

[0067] Specifically, the training set is used to train the dual-path feature fusion network model to obtain a trained dual-path feature fusion network. The process includes:

[0068] 1) Setting the loss function and optimizer of the dual-path feature fusion network, and inputting the training set into the dual-path feature fusion network for model training;

[0069] 2) The loss value is calculated based on the loss function, and the optimizer updates the network parameters according to the loss value and network gradient information;

[0070] 3) If the loss value is less than the preset threshold, the model training ends, and the current two-way feature fusion network is the trained two-way feature fusion network. Otherwise, return to step 1) and continue to use the training set for model training.

[0071] Specifically, Figure 4 As shown in the figure, when the dual-path feature fusion network is trained on the training set, the generalization ability of the model is evaluated by observing the performance on the validation set, and the network parameters and network structure are tuned;

[0072] After obtaining the trained two-way feature fusion network, the performance of the model is evaluated using the test set.

[0073] like Figure 5 As shown in FIG. , a more specific schematic diagram of a dual-path feature fusion network structure and a data processing flow is shown.

[0074] The technical solution of this application is verified by comparative experiments as follows:

[0075] In the MATLAB platform, four groups of power quality disturbance signals with different signal-to-noise ratios (SNR) were generated by random variables, simulating the noise-free state and the signal disturbance conditions affected by noise of SNR=20dB, SNR=40dB and SNR=60dB respectively. Each group of data was further subdivided into 17 power quality disturbance types, covering 9 single disturbance types and 7 compound disturbance types except the standard sine wave. Each disturbance type contained 500 groups of data, totaling 8500 groups of data.

[0076] Each power quality disturbance signal consists of 8 periodic signals, with a total of 1024 data points. Among them, the corresponding relationship between the disturbance type numbers C1~C17 in the confusion matrix and the disturbance type of the power quality disturbance signal is: C1: sine; C2: sag; C3: swell; C4: interruption; C5: flicker; C6: oscillation; C7: pulse; C8: harmonic; C9: notch; C10: peak; C11: sag + harmonic + oscillation; C12: swell + harmonic + oscillation; C13: sag + harmonic; C14: swell + harmonic; C15: sag + oscillation; C16: swell + oscillation; C17: harmonic + oscillation.

[0077] The model parameters are set as follows: initial learning rate is 0.001, batch size is 32, training cycle is 50 epochs, optimizer is SGD, to avoid model overfitting, the penalty coefficient value is set to 0.0001, the loss function uses the cross entropy loss function, and the learning rate scheduler is CosineAnnealing R.

[0078] In order to shorten the training time and improve the model training effect, the transfer learning strategy is used to train the dual-path feature fusion network. Taking the model weights pre-trained on the large-scale dataset ImageNet as the starting point, the dual-path feature fusion network is further trained for power quality disturbance classification and recognition.

[0079] In order to verify the recognition performance of the dual-path feature fusion network in a noisy environment, experiments were conducted on power quality disturbance signals with four levels of signal-to-noise ratio and the data were analyzed. The specific experimental results are shown in Tables 1 and 2 below.

[0080] Table 1 Results of classification of power quality disturbances by different methods in noise-free and SNR=20dB noise environments

[0081]

[0082]

[0083] As can be seen from the above table, the classification accuracy of the present invention in a noise-free environment and a noise environment with SNR=20dB respectively reaches 100% and 99.65%, which is significantly higher than the S transform method and the periodic reconstruction method, indicating that the present invention has extremely high accuracy in identifying power quality disturbance signals. The method proposed in the present invention is more robust and can effectively enhance the expression of power quality disturbance characteristics.

[0084] Although the S-transform and cycle reconstruction methods also show high accuracy, under noisy conditions, the accuracy of the S-transform drops significantly for the two disturbance types of notch and spike; for the two disturbance types of flicker and temporary rise + oscillation, the accuracy of the cycle reconstruction drops slightly. The present invention performs better in various noise environments, which further proves its superiority in processing complex disturbance signals and resisting noise interference. By integrating information in the time domain and frequency domain, a more comprehensive disturbance signal representation can be provided, thereby enhancing the model's ability to judge and identify power quality events.

[0085] Table 2 Results of classification of power quality disturbances by the present invention under different noise environments

[0086]

[0087]

[0088] As can be seen from the above table, in most cases, with the decrease of SNR, that is, the increase of noise level, the accuracy of power quality disturbance classification and identification mostly remains at 100%. However, in the environment of SNR = 20dB, the accuracy of interruption identification drops to 96.55%, and at the noise level of SNR = 40dB, the accuracy of signal classification and identification of certain composite disturbance types (such as swell + harmonic + oscillation, swell + oscillation) drops to 98%, indicating that the present invention has good robustness and accuracy in a higher noise environment, and can effectively identify and classify power quality disturbance signals.

[0089] In the case of low signal-to-noise ratio, the signal used for feature extraction is partially obscured by noise. In the case of interrupt disturbances, because the noise may have higher energy than the signal in a specific frequency band, it will be difficult to accurately distinguish the signal features from the noise. In addition, the identifiability of certain key features in the signal (such as peaks or periodic changes in transient disturbances) may be impaired when the noise level rises, especially those features that change slightly in the time or frequency domain. This may be caused by the instantaneous frequency characteristics or amplitude changes of the signal being masked by noise, resulting in a slight decrease in recognition accuracy.

[0090] Figure 6 to Figure 9The experimental confusion matrix of the model with different feature extractors is shown, indicating that the recognition accuracy of different models for power quality disturbances is relatively high in single and compound disturbances, especially in disturbance types related to sags and interruptions. Specifically, when the voltage amplitude drops to a critical value of about 0.1pu, the features of the sag image and the interruption image are difficult to distinguish. At this time, the subtle features of the feature image are the key to recognition. Due to the difference in processing power of different feature extractors, the model recognition effect is different, indicating that the feature extractor based on the ShuffterNet_V2 lightweight convolutional neural network used in the present invention has a more powerful feature extraction capability in power quality disturbance recognition.

[0091] In order to evaluate the performance of different models in the task of power quality disturbance identification and classification in a noise-free environment, multiple models were experimented in a unified experimental environment. The experimental results are shown in the following table (model performance evaluation is mainly based on four key indicators: accuracy, precision, recall and F1 score):

[0092] Table 3 Performance evaluation of different models for classifying power quality disturbances

[0093] Network Model Accuracy Precision Recall F1-score Parameter quantity / M Training time / s ResNet18 99.47 99.51 99.47 99.47 12.3 82.3 EfficientNet_B0 98.88 98.92 98.88 98.88 10.6 75.6 MobileNet_V3 98.52 98.62 98.52 98.52 6.7 71.5 ShuffleNet_V2 100 100 100 100 5.4 51.6

[0094] As can be seen from the table above, the dual-path feature fusion network based on the single-branch feature extractor of the ShuffterNet_V2 lightweight convolutional neural network has achieved 100% in all evaluation indicators, showing its high accuracy in identifying power quality disturbances in a noise-free environment. At the same time, the number of parameters of ShuffleNet_V2 is only 5.4M, which is significantly lower than 12.3M of ResNet18 and 10.6M of EfficientNet_B0, making ShuffleNet_V2 outstanding in terms of model complexity and computational efficiency.

[0095] In addition, the training time of ShuffleNet_V2 is only 51.6s, which is about 30-40% shorter than ResNet18's 82.3s and EfficientNet_B0's 75.6s, further proving its advantage in training efficiency. Compared with the MobileNet_V3 lightweight convolutional neural network, although MobileNet_V3 also shows competitiveness in terms of parameter volume and training time, ShuffleNet_V2 reduces the parameter volume by about 25% while still maintaining high classification performance, showing its superiority in lightweight models.

[0096] Comprehensive analysis shows that the dual-path feature fusion network proposed in the present invention can more accurately capture and express the characteristics of power quality disturbances in the power quality disturbance classification and identification tasks in practical application scenarios that require fast and efficient processing of large amounts of data, showing excellent performance.

[0097] Embodiment 2

[0098] Another embodiment of the present invention discloses a power quality disturbance signal identification device of a dual-path feature fusion network, such as Fig.10 As shown, a dual-path feature image generation module and a dual-path feature fusion module;

[0099] The dual-path feature image generation module performs S transformation and period reconstruction on the collected power quality disturbance signal in the power system to obtain an S transformation matrix diagram and a period reconstruction matrix diagram;

[0100] The dual-path feature fusion module has a built-in trained dual-path feature fusion network, which extracts features from the input S-transformation matrix diagram and the periodic reconstruction matrix diagram respectively, performs feature fusion and classification, and outputs classification and recognition results of disturbance events in the power system.

[0101] Specifically, the dual-path feature image generation module includes a first-path feature image generation module and a second-path feature image generation module;

[0102] The first characteristic image generation module is used to perform local processing on the power quality disturbance signal in the time-frequency domain, perform signal analysis in time and frequency, and obtain an S-transformation matrix diagram;

[0103] The second characteristic image generation module is used to periodically reconstruct the original signal according to the maximum frequency component in the power spectrum density of the power quality disturbance signal, transform the original signal into a two-dimensional matrix, and obtain a periodic reconstruction matrix diagram.

[0104] The second characteristic image generation module includes a power spectrum density module, a reconstruction period calculation module and an image generation module;

[0105] The power spectrum density module is used to segment the input power quality disturbance signal time series data, perform window function processing, Fourier transform and average processing on each signal segment, and obtain the power spectrum density of the power quality disturbance signal;

[0106] The reconstruction period calculation module is used to determine the maximum frequency component in the power spectrum density, and calculate the reconstruction period length of the power quality disturbance signal according to the period of the maximum frequency component;

[0107] The image generation module is used to calibrate and fill the power quality disturbance signal according to the reconstruction period length, so that the length of the reconstructed signal is an integer multiple of the reconstruction period length, and transform the original signal into a two-dimensional matrix to obtain a periodic reconstruction matrix diagram.

[0108] Specifically, the dual-path feature fusion network includes a feature extraction module, a feature fusion module and a classification module; wherein,

[0109] The feature extraction module is used to extract features from the input S-transformation matrix diagram and the periodic reconstruction matrix diagram through two lightweight convolutional neural networks, and splice and output the extracted two-way features to obtain the extracted power quality disturbance signal features;

[0110] More specifically, the feature extraction module includes two branches of ffterNet_V2 lightweight convolutional neural network for feature extraction; a single branch corresponds to the input S-transformation matrix diagram and the periodic reconstruction matrix diagram for feature extraction; and the extracted features are flattened and spliced ​​to form a set of feature sequences;

[0111] The feature fusion module adopts a multi-head attention mechanism to enhance the ability to capture and understand the complex interactions between features in the extracted feature sequence, and achieves in-depth fusion of features by learning and fusing information in parallel in multiple subspaces, and outputs the fused features;

[0112] The feature fusion module adopts a multi-head attention mechanism to perform feature fusion, maps the input feature sequence through three sets of weight matrices, and generates three sets of vectors of query Q, key K and value V respectively; divides the three sets of vectors of query Q, key K and value V into multiple heads, each head processes a subset of the vectors, each head obtains the attention weight by calculating the dot product of query Q and key K, and uses these attention weights to weight the value V to obtain the output result of each head; merges the output results of each head, and integrates them using a linear transformation layer to obtain a fused feature.

[0113] The classification module is used to perform separate recognition tasks on the fused features and output classification and recognition results.

[0114] The more specific technical details, technical effects and comparative experiments of this embodiment are the same as those in Embodiment 1. Please refer to them for details and will not be described in detail here.

[0115] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for identifying power quality disturbance signals based on a dual-path feature fusion network, characterized in that: The following steps are involved: Step S1, collecting power quality disturbance signals in the power system; Step S2, constructing a dual-path feature image; one path is to perform S transformation on the power quality disturbance signal to obtain an S transformation matrix diagram; the other path is to perform periodic reconstruction on the power quality disturbance signal to obtain a periodic reconstruction matrix diagram; Step S3, inputting the dual-path feature image into the trained dual-path feature fusion network to perform power quality disturbance signal recognition; The dual-path feature fusion network performs feature extraction, feature fusion and classification on the input S-transformation matrix diagram and the periodic reconstruction matrix diagram respectively, and outputs the classification and recognition results of the disturbance events in the power system.

2. The method for identifying power quality disturbance signals using a dual-path feature fusion network according to claim 1, characterized in that: Step S2 includes: The power quality disturbance signal is processed locally in the time and frequency domain by S transform, and the signal is analyzed at different times and frequencies by changing the window function and modulation function to obtain the S transform matrix diagram; According to the maximum frequency component in the power spectrum density of the power quality disturbance signal, the original signal is periodically reconstructed, and the original signal is transformed into a two-dimensional matrix to obtain a periodic reconstruction matrix diagram.

3. The method for identifying power quality disturbance signals using a dual-path feature fusion network according to claim 2 is characterized in that: The process of generating the periodic reconstruction matrix diagram includes: 1) Segment the input power quality disturbance signal time series data, perform window function processing, Fourier transform and average processing on each signal segment to obtain the power spectrum density of the power quality disturbance signal; 2) determining the maximum frequency component in the power spectrum density, and calculating the reconstructed period length of the power quality disturbance signal according to the period of the maximum frequency component; 3) Calibrate and fill the power quality disturbance signal according to the reconstruction period length, so that the length of the reconstructed signal is an integer multiple of the reconstruction period length, transform the original signal into a two-dimensional matrix, and obtain a periodic reconstruction matrix diagram.

4. The method for identifying power quality disturbance signals using a dual-path feature fusion network according to claim 1, characterized in that: The dual-path feature fusion network includes a feature extraction module, a feature fusion module and a classification module; wherein, The feature extraction module is used to extract features from the input S-transformation matrix diagram and the periodic reconstruction matrix diagram through two lightweight convolutional neural networks, and splice and output the extracted two-way features to obtain the extracted power quality disturbance signal features; The feature fusion module adopts a multi-head attention mechanism to enhance the ability to capture and understand the complex interactions between features in the extracted feature sequence, and achieves in-depth fusion of features by learning and fusing information in parallel in multiple subspaces, and outputs the fused features; The classification module is used to perform separate recognition tasks on the fused features and output classification and recognition results.

5. The method for identifying power quality disturbance signals using a dual-path feature fusion network according to any one of claims 1 to 4, characterized in that: The training process of the dual-path feature fusion network includes: 1) Randomly generate time series data of multiple power quality disturbance signals, perform S transformation and period reconstruction on each power quality disturbance signal, obtain the corresponding S transformation matrix diagram and period reconstruction matrix diagram, and construct a data set; 2) Divide the data set into training set, validation set and test set according to the preset ratio; 3) Use the training set to train the dual-path feature fusion network model, and use the validation set and test set to verify and test the model respectively.

6. The method for identifying power quality disturbance signals using a dual-path feature fusion network according to claim 5, characterized in that: The training set is used to train the two-way feature fusion network model to obtain a trained two-way feature fusion network. The process includes: 1) Setting the loss function and optimizer of the dual-path feature fusion network, and inputting the training set into the dual-path feature fusion network for model training; 2) The loss value is calculated based on the loss function, and the optimizer updates the network parameters according to the loss value and network gradient information; 3) If the loss value is less than the preset threshold, the model training ends, and the current two-way feature fusion network is the trained two-way feature fusion network. Otherwise, return to step 1) and continue to use the training set for model training.

7. A power quality disturbance signal identification device based on a dual-path feature fusion network, characterized in that: include: Two-way feature image generation module and two-way feature fusion module; The dual-path feature image generation module performs S transformation and period reconstruction on the collected power quality disturbance signal in the power system to obtain an S transformation matrix diagram and a period reconstruction matrix diagram; The dual-path feature fusion module has a built-in trained dual-path feature fusion network, which extracts features from the input S-transformation matrix diagram and the periodic reconstruction matrix diagram respectively, performs feature fusion and classification, and outputs classification and recognition results of disturbance events in the power system.

8. The power quality disturbance signal identification device of the dual-path feature fusion network according to claim 7 is characterized in that: The dual-path feature image generation module includes a first-path feature image generation module and a second-path feature image generation module; The first characteristic image generation module is used to perform local processing on the power quality disturbance signal in the time-frequency domain, perform signal analysis in time and frequency, and obtain an S-transformation matrix diagram; The second characteristic image generation module is used to periodically reconstruct the original signal according to the maximum frequency component in the power spectrum density of the power quality disturbance signal, transform the original signal into a two-dimensional matrix, and obtain a periodic reconstruction matrix diagram.

9. The power quality disturbance signal identification device of the dual-path feature fusion network according to claim 8 is characterized in that: The dual-path feature fusion network includes a feature extraction module, a feature fusion module and a classification module; wherein, The feature extraction module is used to extract features from the input S-transformation matrix diagram and the periodic reconstruction matrix diagram through two lightweight convolutional neural networks, and splice and output the extracted two-way features to obtain the extracted power quality disturbance signal features; The feature fusion module adopts a multi-head attention mechanism to enhance the ability to capture and understand the complex interactions between features in the extracted feature sequence, and achieves in-depth fusion of features by learning and fusing information in parallel in multiple subspaces, and outputs the fused features; The classification module is used to perform separate recognition tasks on the fused features and output classification and recognition results.

10. The power quality disturbance signal identification device of the dual-path feature fusion network according to claim 9, characterized in that: The feature extraction module includes two branches of ffterNet_V2 lightweight convolutional neural network for feature extraction; the single branch corresponds to the input S-transformation matrix diagram and the periodic reconstruction matrix diagram for feature extraction; and the extracted features are flattened and spliced ​​to form a set of feature sequences; The feature fusion module adopts a multi-head attention mechanism to perform feature fusion; the input feature sequence is mapped through three sets of weight matrices to generate three sets of vectors of query Q, key K and value V respectively; the three sets of vectors of query Q, key K and value V are divided into multiple heads, each head processes a subset of the vectors, each head obtains the attention weight by calculating the dot product of the query Q and the key K, and uses these attention weights to weight the value V to obtain the output result of each head; the output results of each head are merged and integrated using a linear transformation layer to obtain a fused feature.

Images