Power Quality Disturbance Recognition Method Based on Feature Image Combination and Improved ResNet-18

Abstract

The present invention proposes a power quality disturbance recognition method based on feature image combination and improved ResNet-18. First, variational mode decomposition is performed on each power quality disturbance signal to obtain a series of intrinsic mode functions and residual components; secondly, the intrinsic mode functions, residual components, original disturbance signals and Subtract components are vertically spliced into a component matrix, and a feature component color map is generated by using a signal-image conversion method; thirdly, continuous wavelet transform is performed on the original disturbance signal to generate a wavelet time-frequency map; finally, the feature component color map and the wavelet time-frequency map are input into the improved six-channel ResNet-18 in parallel to complete disturbance recognition. The power quality disturbance recognition method is analyzed through simulation signals, and the results show that the proposed method has good anti-noise performance and can better extract the feature information of power quality disturbances, achieving a higher recognition accuracy rate.

Description

Technical Field

[0001] The present invention belongs to the technical field of power quality disturbance identification, and particularly relates to a power quality disturbance identification method based on feature image combination and improved ResNet-18. Background Art

[0002] Driven by the dual impetus of the energy revolution and the digital revolution, the electrification of new power systems and the high proportion of new energy infiltrating into the power grid are the main trends in the development of current power systems. However, with the continuous development of new power systems, power quality disturbance problems have gradually emerged: 1) New power quality disturbances, such as voltage spikes and voltage notches, continuously appear in new energy systems; 2) Various single disturbances are compounded and multiplexed, such as the compound of voltage sags and harmonics. The power quality deterioration problems caused by these power quality disturbances will not only affect the power consumption side, such as problems like substandard equipment quality and malfunction of precision instruments, but may even endanger the stable operation of the power system, such as the disconnection of distributed micro-sources from the grid. The power system's requirements for power quality are gradually increasing, and the prerequisite for improving power quality problems is to quickly and accurately classify and identify disturbances. Therefore, it is of great significance to identify the types of power quality disturbances that occur in new power systems with new energy as the main body.

[0003] In recent years, many studies have proposed new methods for identifying power quality disturbances based on deep learning. These methods avoid the subjectivity of traditional manual feature selection and instead utilize the powerful automatic learning and feature mining capabilities of neural networks for feature extraction, enabling them to better adapt to the problem of power quality disturbance identification in new power systems and achieve a relatively high recognition accuracy. Such methods mainly analyze from two aspects: signal preprocessing and power quality disturbance identification. However, the feature discrimination of the signal after preprocessing and the feature extraction ability of the neural network are crucial. Existing studies have continuously deepened and improved in the above two aspects, and the recognition effect has been improved to some extent. However, there are still the following problems: 1) The complementarity of features between signal preprocessing methods has not been concerned, and complementary features have not been organically combined. Therefore, it is impossible to construct an image with relatively complete feature information. For example, in reference [1], continuous wavelet transform is used to generate a wavelet-time frequency diagram, and the generated image and the original disturbance signal are respectively input into a convolutional neural network and a long short-term memory network for parallel recognition, initially using the combination of time-frequency analysis method and the original disturbance signal. However, this method only introduces the original disturbance signal with insufficient features based on the time-frequency image, and the recognition effect for composite power quality disturbances is poor, still having a large room for improvement. 2) Many transfer learning methods simply add fully connected layers after the network without making adaptive modifications to the network structure. Although transfer learning can effectively improve the training speed of the network, it inevitably limits the composition of the input image and the optimization of the network structure, making it difficult to improve the accuracy. For example, in reference [2], a method for identifying power quality disturbances based on dual-channel Gram angle field and ResNet is proposed. Due to the limitation of the three channels of ResNet in transfer learning, in order to combine two pictures, only single-channel image combination can be used, without making adaptive modifications to the ResNet structure, reducing the implicit information volume of the pictures and making it difficult to further improve the recognition accuracy.

[0004] [1] Zhang Zhentao. Research on Denoising and Identification of Power Quality Disturbance Signals Based on Wavelet Transform and Deep Learning [D]. Nanchang University, 2021. DOI: 10.27232 / d.cnki.gnchu.2021.002327.

[0005] [2] He Caijun, Li Kaicheng, Yang Wangwang, Dong Yufei, Song Zhaoxia, Fan Weixin, Wang Wei. Identification of Composite Power Quality Disturbances Based on Dual-Channel GAF and Deep Residual Network [J / OL]. Power System Technology: 1-10 [2022-08-12]. DOI: 10.13335 / j.1000-3673.pst.2022.0644. Summary of the Invention

[0006] Aiming at the problems of limited single-image feature information and insufficient algorithm recognition ability in the traditional power quality disturbance recognition system, based on the idea of feature fusion, the present invention aims to provide a power quality disturbance recognition method based on feature image combination and improved ResNet-18. First, variational mode decomposition is performed on each power quality disturbance signal to obtain a series of intrinsic mode functions and residual components. Second, the intrinsic mode functions, residual components, original disturbance signals, and Subtract components are vertically concatenated into a component matrix, and a signal-image conversion method is used to generate a feature component color map. Third, continuous wavelet transform is performed on the original disturbance signal to generate a wavelet time-frequency map. Finally, the feature component color map and the wavelet time-frequency map are input in parallel to the improved six-channel ResNet-18 to complete disturbance recognition. The power quality disturbance recognition method is analyzed through simulation signals, and the results show that the proposed method has good anti-noise performance and can better extract power quality disturbance feature information, achieving a higher recognition accuracy.

[0007] Aiming at the two major problems of the power quality recognition method, the present invention proposes a power quality disturbance recognition method based on feature image combination and improved ResNet-18. The key steps are as follows:

[0008] First, generate each PQD sampling signal and perform VMD decomposition on it to obtain IMFs and residual components.

[0009] The Subtract component is proposed and the original disturbance signal component is introduced to form a component matrix to suppress the endpoint effect of the signal decomposition method and amplify the implicit features of the PQD signal. Then, the component matrix normalization method and the pseudo-color coding technology are used to convert the component matrix into a three-channel n×n×3 feature component color map, obtaining the signal fluctuation features and deep features of each component;

[0010] Second, perform CWT on each PQD sampling signal using the 'db4' wavelet to generate a three-channel n×n×3 wavelet time-frequency map to obtain the time-frequency energy features of the original disturbance signal;

[0011] Third, according to the feature fusion idea, the feature component color map and the wavelet time-frequency map are combined to form a six-channel n×n×6 image with both time-frequency energy features and deep features of the disturbance signal;

[0012] Finally, considering various noise and disturbance situations, generate an image set with sufficient samples, and input the combined image into the improved six-channel ResNet-18 for training and complete recognition.

[0013] While learning the time-frequency energy features of the wavelet, the network learns the deep features of the disturbance signal in the feature component color map, thereby constructing a high-accuracy recognition model.

[0014] The technical solution adopted by the present invention to solve its technical problems is as follows:

[0015] A power quality disturbance recognition method based on feature image combination and improved ResNet-18:

[0016] First, perform variational mode decomposition (VMD) on various types of power quality disturbance signals to obtain intrinsic mode functions (IMFs) and residual components;

[0017] Vertically splice the intrinsic mode functions, residual components, original disturbance signals, and Subtract components into a component matrix, and use a signal-image conversion method to generate a three-channel n×n×3 feature component color map; the Subtract component refers to the absolute value difference between the original disturbance signal and the normal signal;

[0018] Secondly, perform continuous wavelet transform on the original disturbance signal to generate a wavelet time-frequency map; finally, parallelly input the feature component color map and the wavelet time-frequency map into the improved six-channel ResNet-18 to complete disturbance recognition.

[0019] Furthermore, the various types of power quality disturbance signals at least include the normal operating state and the power quality disturbance signal model constructed according to IEEE Std. 1159-2019. This is used as the recognition object.

[0020] Furthermore, the process of vertically splicing the intrinsic mode functions, residual components, original disturbance signals, and Subtract components into a component matrix specifically includes:

[0021] Vertically splice the decomposed components, Subtract components, and original disturbance signals to form a component matrix to suppress the end effect of the signal decomposition method and amplify the implicit features of the power quality disturbance signal; the composition method of the Subtract component is shown in Equation (1) and is composed of the absolute value subtraction of the disturbance signal (y(t)) and the normal operating signal (y Normal (t)):

[0022] Subtract(t) = |y(t)| - |y Normal (t)| (1)

[0023] Each component is vertically spliced to form an augmented matrix X, as shown in Equation (2):

[0024]

[0025] Furthermore, the process of using a signal-image conversion method to generate a three-channel n×n×3 feature component color map specifically includes:

[0026] The augmented matrix X is transformed into a three-channel n×n×3 feature component color map by using the component matrix normalization method and the pseudo-color coding technique, as shown in Equation (3):

[0027]

[0028] In Equation (3), P(m,n) represents the pixel value of the pixel matrix in the m-th row and the n-th column; x i represents the value of the i-th data point; x min and x max represent the minimum value and the maximum value of each component respectively; the values in the component matrix are normalized to the interval (0, 255) through Equation (3), so that each value in the component matrix has a certain pixel intensity, and a grayscale image is drawn according to the size of each pixel point value, and then the grayscale image is transformed into a feature component color map by using pseudo-color coding.

[0029] Furthermore, the continuous wavelet transform is performed on the original disturbance signal to generate a wavelet time-frequency map, specifically: the continuous wavelet transform is performed on each power quality disturbance signal by using the 'db4' wavelet, and the frequency display range is 0 - 800 Hz, and a three-channel n×n×3 wavelet time-frequency map is generated to obtain the time-frequency energy characteristics of the original disturbance signal.

[0030] Furthermore, the parallel input of the feature component color map and the wavelet time-frequency map specifically means: the feature component color map and the wavelet time-frequency map are combined to form a six-channel n×n×6 image with both time-frequency energy characteristics and deep characteristics of the disturbance signal.

[0031] Furthermore, the improved six-channel ResNet-18 is based on the three-channel ResNet-18 structure:

[0032] The number of channels of the input two-dimensional image is increased to twice, and the number of convolution kernels in the convolutional layer is also increased to twice to ensure that the number of channels is adapted to the image size;

[0033] Two fully connected layers are added to the network, namely a fully connected layer with 2000 full connections and a fully connected layer with 1000 full connections to ensure that the feature quantity is adapted to the number of channels; finally, a fully connected layer with the number of disturbances as the full connection number is added for classification and recognition.

[0034] Compared with the prior art, the present invention and its preferred embodiments provide an image composition method with sufficient spatio-temporal information and complete feature information for the input image: 1) Based on the variational mode decomposition method, the original perturbation signal and the Subtract component (the absolute value difference between the original perturbation signal and the normal signal) are introduced, and the matrix splicing and pseudo-color coding technology are used to generate the color map of the feature component; 2) Since the wavelet time-frequency map fully retains the time-frequency energy characteristics of the perturbation signal, and the color map of the feature component further explores the deep features of the perturbation signal by using the signal decomposition method. Based on the feature fusion idea, the present invention utilizes the feature complementarity between the wavelet time-frequency map and the color map of the feature component, and combines the two images to form a six-channel image with sufficient feature information, realizing the completion of features;

[0035] In terms of the network recognition structure, the present invention proposes an improved multi-channel ResNet-18 network structure based on the combined image to classify and identify power quality disturbances. The improved ResNet-18 fully retains the structural advantages of ResNet, and at the same time can adapt to multi-channel image input and multiple feature extraction, can achieve better recognition effects in various noise environments and has a certain universality. Brief Description of the Drawings

[0036] The following further details the present invention in conjunction with the drawings and specific embodiments:

[0037] Figure 1 It is the basic flowchart of the PQD recognition based on the feature image combination of the color map of the feature component and the wavelet time-frequency map and the improved ResNet-18 proposed by the present invention;

[0038] Figure 2 Schematic diagram of the typical power quality disturbance waveform in the embodiment of the present invention;

[0039] Figure 3 Schematic diagram of the generation process of the color map of the feature component in the embodiment of the present invention;

[0040] Figure 4 It is the typical color map of the feature component of various power quality disturbance signals in the embodiment of the present invention (after grayscale processing);

[0041] Figure 5 It is the typical wavelet time-frequency map of various power quality disturbance signals in the embodiment of the present invention;

[0042] Figure 6 Schematic diagram of the combined image of the typical power quality disturbance in the embodiment of the present invention;

[0043] Figure 7 It is the schematic diagram of the accuracy change curve during the training process of ResNet-18 in the embodiment of the present invention. Detailed Description of the Invention

[0044] To make the features and advantages of this patent more obvious and understandable, specific embodiments are given below for detailed description as follows:

[0045] It should be noted that the following detailed description is illustrative and aims to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application belongs.

[0046] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to this application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0047] The basic process of PQD recognition for the feature image combination based on the feature component color map and wavelet time-frequency map and the improved ResNet-18 proposed in the embodiments of the present invention is as Figure 1 shown, and the main content is as follows:

[0048] 1 Power quality disturbance signal generation

[0049] According to IEEE Std.1159-2019, a power quality disturbance signal model including normal operating conditions is constructed, including voltage sag, voltage swell, voltage interruption, harmonics, flicker, transient oscillation, voltage notch and voltage spike, voltage sag + harmonics, voltage swell + harmonics, voltage interruption + harmonics, flicker + harmonics, flicker + voltage sag and flicker + voltage swell, as shown in Table 1. According to Table 1, use Matlab to generate power quality disturbance signals y(t) with no noise, 40dB, 35dB, 30dB, and 25dB signal-to-noise ratios within 10T with random amplitudes, random occurrence times, and random durations. The fundamental frequency is set to 50Hz, the sampling frequency is set to 3.2kHz, and the number of sampling points is set to 640 points. Typical power quality disturbance waveforms are as Figure 2 shown.

[0050] Table 1 Power quality disturbance mathematical model

[0051]

[0052]

[0053]

[0054] 2 Composition of the feature component color map

[0055] Variational mode decomposition is performed on various power quality disturbance signals to obtain intrinsic mode functions and residual components (collectively referred to as decomposition components). The decomposition components are vertically concatenated with the Subtract component and the original disturbance signal to form a component matrix to suppress the end effect of the signal decomposition method and amplify the implicit features of the power quality disturbance signal. The composition method of the Subtract component is shown in Equation (1) and is composed of the absolute value subtraction of the disturbance signal (y(t)) and the normal operation signal (y Normal (t)).

[0056] Subtract(t) = |y(t)| - |y Normal (t)| (1)

[0057] Each component is vertically concatenated to form an augmented matrix X, as shown in Equation (2):

[0058]

[0059] The augmented matrix X is transformed into a three-channel n×n×3 feature component color map by using the component matrix normalization method and the pseudo-color coding technique, as shown in Equation (3).

[0060]

[0061] In Equation (3), P(m,n) represents the pixel value of the pixel matrix in the m-th row and the n-th column; x i represents the value of the i-th data point; x min and x max represent the minimum and maximum values of each component, respectively. By normalizing the values in the component matrix to the range of (0, 255) through Equation (3) (since different maximum and minimum values are selected for normalization and the recognition effect is different, Table 2 gives the optimal component normalization method adopted in the present invention), each value in the component matrix has a certain pixel intensity. According to the magnitude of each pixel point value, it can be drawn as a grayscale image, and then the grayscale image can be transformed into a feature component color map by using pseudo-color coding. The overall generation process is as Figure 3 shown.

[0062] Table 2 Normalization method of components in the augmented matrix X

[0063]

[0064] The feature component color map constructs a feature image rich in spatio-temporal information by using the combination method of multiple feature components: The variational mode decomposition method uses the decomposition components to deeply analyze the fluctuation characteristics and deep features of each disturbance signal. At the same time, with the assistance of the Subtract component and the original disturbance signal, it not only overcomes the problems brought by the end effect of the decomposition method but also enhances the expressiveness of the signal features in the noise environment. Its typical image is asFigure 4 as shown

[0065] 3 Generation of Wavelet Time-Frequency Diagram

[0066] Perform continuous wavelet transform on each power quality disturbance signal using the 'db4' wavelet. The most optimal frequency display range is selected from 0 to 800 Hz, and a three-channel n×n×3 wavelet time-frequency diagram is generated to obtain the time-frequency energy characteristics of the original disturbance signal. The 'db4' wavelet is more similar to the sine wave, so it can show more obvious energy characteristics in the wavelet time-frequency diagram. Its typical image is as Figure 5 as shown

[0067] 4 Image Combination Based on Feature Complementarity

[0068] The color map of feature components can better perceive the mutation characteristics of the signal when analyzing the decomposed components, the original disturbance signal, and the Subtract component, but it is difficult to reflect the small fluctuations of some high-frequency components. The wavelet time-frequency diagram performs time-frequency domain analysis on the original disturbance signal to extract the time-frequency energy characteristics of the disturbance signal, but the wavelet transform is not sensitive to some singular points and mutation points. The two have natural feature complementarity.

[0069] Therefore, aiming at the problem of incomplete spatio-temporal information of images in the existing methods and based on the idea of feature fusion, the color map of feature components and the wavelet time-frequency diagram are combined to form a six-channel n×n×6 image with both time-frequency energy characteristics and deep characteristics of the disturbance signal;

[0070] 5 Improvement of ResNet-18

[0071] Aiming at the problems of traditional transfer learning networks, the present invention proposes an improved ResNet-18 network structure based on the six-channel feature image. The improvement method is as follows:

[0072] Based on the traditional three-channel ResNet-18 structure, the input image size is modified from 224×224×3 to 224×224×6, and the number of channels of the input two-dimensional image becomes twice the original. Since the number of convolutional kernels is closely related to the number of channels and the network satisfies Equation (4) during convolutional operations, the number of convolutional kernels also needs to be modified accordingly. Therefore, the number of convolutional kernels in the convolutional layer is also changed to twice the original to ensure that the number of channels adapts to the image size.

[0073]

[0074] where out (l) is the l-dimensional feature matrix output after the input feature matrix is calculated by l convolutional kernels; x l i,j is the element in the input l-dimensional feature matrix; kl u-i,v-j is the internal element of l convolutional kernels; B l i,j is the bias of the convolutional layer of this layer; l in and l out are the dimensions of the input feature matrix and the output feature matrix.

[0075] At the same time, due to the changes in the number of image channels and convolutional kernels, the feature quantities extracted by the network through operations such as convolution, pooling, and batch normalization are doubled compared to the unimproved ResNet-18. Therefore, two fully connected layers are added to the network, namely a fully connected layer with 2000 neurons and a fully connected layer with 1000 neurons to ensure that the feature quantities are adapted to the number of channels; finally, a fully connected layer with the number of neurons equal to the number of disturbances is added for classification and recognition. Thus, the improved multi-channel ResNet-18 can be obtained. Its overall structure is still similar to that of the traditional ResNet-18, retaining the structural advantages of ResNet-18, and modifying the internal parameters of the network to enable it to adapt to the recognition of multi-channel images, expanding the composition method of the input images. The present invention adopts the improved ResNet-18 structure, and its structure is shown in Tables 3 and 4.

[0076] Table 3 Improved ResNet-18 residual structure

[0077]

[0078]

[0079] Table 4 Improved ResNet-18 bottom structure

[0080] Hierarchy Parameter Merged layer - Relu function - Average pooling layer Size: 3×3; Stride: 2; Padding: 1 Fully connected layer Number of fully connections: 2000 Fully connected layer Number of fully connections: 1000 Fully connected layer Number of fully connections: 15 SoftMax layer - Output layer -

[0081] 6 Power Quality Disturbance Recognition Based on Feature Image Combination and Improved ResNet-18

[0082] For the generated 15 types of PQD signals, the feature component color maps and wavelet time-frequency maps are respectively generated according to the above image generation method, and image combination is performed as training and test samples. Typical sample images are as Figure 6 shown. Each type of disturbance includes 300 samples under the same noise environment, and the training set and test set are divided according to the ratio of 7:3. The combined images are input into the improved six-channel ResNet-18 for training and testing. The number of iterations is 2500 times, and the training parameters adopted by the network are shown in Table 5. While the network learns the wavelet time-frequency energy features, it also learns the deep features of the disturbance signals in the feature component color maps, and continuously performs self-adjustment of the internal weights and parameters to construct a high-accuracy recognition model. The accuracy change curve during the network training process is as Figure 7As shown. The recognition accuracy rates in various noise environments are shown in Table 6.

[0083] Table 5 Network training parameters

[0084] Hierarchy Parameter Batch size 64 Initial learning rate 0.0001 Loss function Cross entropy loss function Optimizer adam

[0085] Table 6 Average recognition accuracy rates of various perturbations of the advanced network

[0086]

[0087] Through Figure 7 It can be seen from Table 6 that the proposed method has good convergence performance: when the network iterates about 700 times, it can reach a recognition accuracy rate of more than 90% and fluctuate slightly at 100%; it can reach a high recognition accuracy rate in various noise environments: the improved ResNet-18 has strong feature extraction ability. With the feature combination of two feature images, whether in a weak noise environment or a strong noise environment, the overall average recognition accuracy rate of its test samples can reach more than 97.78%, and the recognition accuracy rate of C7 can be maintained above 94.44% except in the environment with a signal-to-noise ratio of 25 dB. It can be seen that the proposed method has good anti-noise performance and recognition ability.

[0088] 7 Summary

[0089] In summary, in order to improve the recognition accuracy rate of various PQDs, the embodiments of the present invention overcome the limitations of traditional methods and propose a new PQD recognition method based on the combination of feature images and improved ResNet-18. This recognition system uses the feature image combination of the feature component color map and the wavelet time-frequency map and the improved six-channel ResNet-18 to extract the fusion features of the images, and has a high recognition accuracy rate. The conclusions obtained through experiments are as follows:

[0090] 1) In terms of the input image, the present invention proposes a method of combining two images based on the idea of feature complementarity. The wavelet time-frequency map has the time-frequency energy characteristics of the PQD signal, and the feature component color map has good component fluctuation characteristics. The combination of the two images has more sufficient spatio-temporal feature information and stronger feature expressiveness. Among them, the feature component color map provides a new composition method of component splicing and generating an image, expanding the previous image generation method of PQD signals.

[0091] 2) In terms of the recognition network, an improved multi-channel ResNet method is proposed to classify and recognize PQDs. The improved ResNet-18 can better suppress the gradient problem and the network degradation problem and can automatically extract the deep features of multiple combined images, and can achieve better recognition effects.

[0092] 3) The method proposed by the present invention has strong multi-image analysis capabilities. It is not only applicable to the PQD classification field, but also can be applied to time series prediction problems affected by various factors, such as photovoltaic power prediction, after appropriately modifying the underlying structure of ResNet.

[0093] As described above, it is only a preferred embodiment of the present invention, and it is not a limitation of the present invention in other forms. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes. However, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution content of the present invention still belong to the protection scope of the technical solution of the present invention.

[0094] This patent is not limited to the above best implementation mode. Anyone inspired by this patent can obtain various other forms of power quality disturbance identification methods based on feature image combination and improved ResNet-18. All equal changes and modifications made according to the scope of the patent application of the present invention shall fall within the scope covered by this patent.

Claims

1. A method for identifying power quality disturbances based on feature image combination and improved ResNet-18, characterized in that: First, perform variational mode decomposition (VMD) on various types of power quality disturbance signals to obtain intrinsic mode functions (IMFs) and residual components; Vertically splice the intrinsic mode functions, residual components, original disturbance signals, and Subtract components (where the Subtract component refers to the absolute value difference between the original disturbance signal and the normal signal) to form a component matrix, and use a signal-image conversion method to generate a three-channel n×n×3 feature component color map; Secondly, perform continuous wavelet transform on the original disturbance signal to generate a wavelet time-frequency map; finally, combine the feature component color map and the wavelet time-frequency map and input them into the improved six-channel ResNet-18 to complete disturbance identification; The process of vertically splicing the intrinsic mode functions, residual components, original disturbance signals, and Subtract components to form a component matrix specifically includes: Vertically splice the decomposed components with the Subtract component, the original disturbance signal, the IMFs, and the residual component to form a component matrix, so as to suppress the end effect of the signal decomposition method and amplify the implicit features of the power quality disturbance signal; among them, the composition method of the Subtract component is shown in Equation (1), which is composed of taking the absolute value subtraction of the disturbance signal (y(t)) and the normal operation signal (y Normal (t)): Subtract(t)=|y(t)|-|y Normal (t)| (1) Vertically splice each component to form an augmented matrix X, as shown in Equation (2): The process of using a signal-image conversion method to generate a three-channel n×n×3 feature component color map specifically includes: Use a component matrix normalization method and a pseudo-color coding technique to convert the augmented matrix X into a three-channel n×n×3 feature component color map, as shown in Equation (3): In Equation (3), P(m,n) represents the pixel value of the pixel matrix in the m-th row and the n-th column; x i represents the value of the i-th data point; x min and x max represent the minimum value and the maximum value of each component respectively; the values in the component matrix are normalized to the range of (0, 255) through Equation (3), so that each value in the component matrix has a certain pixel intensity. According to the magnitudes of the values of each pixel point, they are plotted as a grayscale image, and then the grayscale image is converted into a color image of characteristic components by using pseudo-color coding; Feature image combination specifically refers to: combining the feature component color map and the wavelet time-frequency map to form a six-channel n×n×6 image with both time-frequency energy characteristics and deep characteristics of the disturbance signal; The improved six-channel ResNet-18 is based on the three-channel ResNet-18 structure: Increase the number of channels of the input two-dimensional image to twice, and also increase the number of convolution kernels in the convolutional layer to twice to ensure that the number of channels adapts to the image size; add two fully connected layers to the network, namely a fully connected layer with 2000 neurons and a fully connected layer with 1000 neurons to ensure that the feature quantity adapts to the number of channels; finally, add a fully connected layer with the number of neurons equal to the number of disturbances for classification and identification.

2. The power quality disturbance recognition method based on feature image combination and improved ResNet-18 according to claim 1, characterized in that: The various types of power quality disturbance signals at least include the normal operating state and the power quality disturbance signal model constructed according to IEEE Std. 1159-2019.

3. The power quality disturbance recognition method based on feature image combination and improved ResNet-18 according to claim 1, characterized in that: Performing continuous wavelet transform on the original disturbance signal to generate a wavelet time-frequency map specifically means: performing continuous wavelet transform on each power quality disturbance signal using the 'db4' wavelet, with the frequency display range of 0 - 800 Hz, to generate a three-channel n×n×3 wavelet time-frequency map to obtain the time-frequency energy characteristics of the original disturbance signal.

Images