Insulator contamination quantitative detection system and method based on residual structure convolutional neural network

Abstract

The invention proposes a system for quantitatively detecting the pollution degree of an insulator based on a residual structure convolutional neural network, which belongs to the technical field of insulator pollution degree detection. The system extracts high-spectral spectrum line data from a pollution sampling area and divides the data into a training set and a test set according to the ratio of spectrum division; randomly initializes network parameters, inputs full-band spectrum data into a feature extraction network to extract pollution degree features; inputs the extracted feature vector into a regression layer for regression training, uses the mean square error as a loss function, and calculates the error between the predicted value and the true label; uses an optimizer to update the network parameters, and accelerates the training process by adjusting the adaptive learning rate; inputs unknown samples to be tested into a trained detection model to obtain detection results; the invention realizes quantitative detection of the pollution degree of insulators, helps to optimize maintenance plans, improve the operation and maintenance efficiency of power systems, and timely discover and eliminate potential faults, thereby improving the reliability and safety of power systems.

Description

Technical Field

[0001] The present invention belongs to the technical field of insulator contamination degree detection, and in particular, relates to an insulator contamination degree quantitative detection system and method based on a residual structure convolutional neural network. Background Art

[0002] Insulators are important insulating tools in electrical equipment systems. They have the functions of support, isolation and insulation, and can ensure the safe and stable operation of power systems in various places. Since insulators are generally exposed to open air, their surfaces are corroded by various pollutants over a long period of time, which leads to a decrease in the lightning resistance and insulation level of the line, making it easy to cause flashover accidents. Flashover accidents can cause large-scale power outages and bring losses to the society and economy.

[0003] In order to accurately detect and evaluate the degree of contamination on the surface of insulators, domestic and foreign scholars have proposed different detection methods for different characteristic quantities over the years. The International Conference on Large Electric Systems (CIGRE) recommends five contamination detection methods: equivalent salt density method, contamination layer conductivity method, pollution flashover voltage gradient method, pulse counting method and leakage current method. However, traditional detection methods using electrical parameters such as conductivity, voltage gradient, leakage current and other electrical parameters as characteristic quantities generally have problems such as low efficiency, complex processes and high costs, and cannot achieve accurate and efficient live detection of the degree of contamination on large-scale line insulators. In recent years, with the development of various optical imaging technologies, researchers have successively developed methods such as ultrasonic, infrared imaging, and ultraviolet imaging for determining the degree of insulator contamination. The characteristic quantity has gradually changed from electrical quantity detection to non-electrical quantity detection, and the detection technology has also developed from the initial contact detection to non-contact detection. Among the non-contact detection methods, ultraviolet discharge imaging and infrared imaging have proven their effectiveness in detecting abnormal corona and heating problems on insulators. However, since significant leakage current is required to trigger insulator corona and heating, even severe contamination may not meet the conditions for their generation. Therefore, infrared imaging and ultraviolet solar-blind imaging have limited capabilities in detecting the degree of contamination on the insulator surface. Summary of the invention

[0004] To address the above problems, the present invention combines hyperspectral spectrum with deep learning and proposes a quantitative detection system and method for insulator contamination based on residual structure convolutional neural network, which realizes quantitative detection of insulator contamination, helps to optimize maintenance plans, improve the operation and maintenance efficiency of power systems, and promptly discover and eliminate potential faults, thereby improving the reliability and safety of power systems.

[0005] The present invention is achieved through the following technical solutions:

[0006] A quantitative detection system for insulator contamination based on residual structure convolutional neural network:

[0007] The insulator contamination quantitative detection system includes Backbone layer, Neck layer and Head layer:

[0008] The Backbone layer is used to perform preliminary feature extraction on the input full-band spectral data, including four residual convolution modules.

[0009] The Neck layer further extracts, upsamples, and concatenates the features extracted from the fourth, third, and second residual blocks to form three new feature maps.

[0010] The Head layer first flattens the three feature maps transmitted by the Neck layer. The three unfolded feature vectors are concatenated into one feature vector after weighted calculation by two linear layers. The feature vector is then integrated through information of five linear layers to finally realize the prediction from spectral information to salt and grayscale values.

[0011] Furthermore, the residual convolution module is composed of a main path and a jump-layer connection path;

[0012] The main path includes convolution layer 1, activation layer 1, convolution layer 2, average pooling layer, and activation layer 2;

[0013] Define the network input feature as a one-dimensional feature vector x, first pass it to the convolution layer 1 to preliminarily extract the local feature information of the feature vector x, output it to the activation layer 1 to introduce nonlinear factors through the activation function, and the features processed by the activation layer 1 are input to the convolution layer 2 for feature extraction and combination. The features output by the convolution layer 2 enter the average pooling layer to adjust the size, and then enter the activation layer 2. The output feature after the activation layer 2 is F(x);

[0014] The jump-layer connection path includes only one convolution layer. Because the dimension of the main path is inconsistent with the input after two layers of convolution, in order to satisfy the identity mapping, the dimension of the input x is adjusted through the convolution layer of the jump-layer connection path and then added to F(x), and finally the added feature vector is output.

[0015] Furthermore, the input of the residual structure convolutional neural network is the preprocessed full-band hyperspectral spectrum line data, including feature vectors and corresponding true labels; the network output is the predicted salt ash content value; and the mean square error (MSE) is used as the loss function.

[0016] A quantitative detection method for insulator contamination based on residual structure convolutional neural network:

[0017] The method specifically comprises the following steps:

[0018] Step 1: Extract hyperspectral line data for the polluted sampling cloth area, and divide the full-band spectral reflectance data that has been spectrally preprocessed into a training set and a test set according to the ratio of spectral division;

[0019] Step 2: Randomly initialize network parameters, including weights and biases of feature extraction network and regression layer;

[0020] Step 3: Input the full-band spectrum data into the feature extraction network to extract the pollution degree feature and obtain the feature vector;

[0021] Step 4: Input the extracted feature vector into the regression layer for regression training, use the mean square error as the loss function, and calculate the error between the predicted value and the true label;

[0022] Step 5: Back propagation and parameter update: Use the optimizer to update the network parameters and accelerate the training process by adaptively adjusting the learning rate so that the regression model can show better predictive performance on the data.

[0023] Step 6: Model prediction: The unknown sample to be tested is input into the trained detection model to obtain the detection result.

[0024] Furthermore, in step one,

[0025] Place a special contamination sampling cloth in the contamination area of ​​the insulator and collect contamination sampling cloth samples after a period of time;

[0026] Use a hyperspectral imager to collect spectral data of the polluted sampling cloth to obtain full-band spectral reflectance data;

[0027] The collected raw spectral data were denoised to eliminate interference, baseline corrected to eliminate baseline offset, and then normalized to enhance data stability.

[0028] Furthermore, in step three,

[0029] The four residual convolution modules of the Backbone layer are used to extract the dirtiness feature and obtain the feature vector.

[0030] The main path of the residual convolution module extracts feature information, and the jump-layer connection path adjusts the dimension to ensure identical mapping.

[0031] Furthermore, in step 4,

[0032] The calculation formula of the MSE is:

[0033] Where n is the number of samples, y i is the true label value of the i-th sample, is the predicted value of the i-th sample.

[0034] Furthermore, in step five,

[0035] Based on the calculated loss function value, the back-propagation algorithm is used to calculate the gradient; starting from the regression layer, the gradient is back-propagated along the network structure, and the gradient values ​​of the weights and biases of each layer are calculated in turn; during the calculation process, according to the chain rule, the partial derivative of the loss function with respect to each neuron output is multiplied by the partial derivative of the neuron with respect to the input, and then gradually passed to the input layer of the network to obtain the gradient information of each parameter.

[0036] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0037] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the steps of the above method are implemented.

[0038] Beneficial effects of the present invention

[0039] The feature extraction capability of the present invention is stronger: the Backbone part adopts the residual convolution module and the residual structure of the jump-layer connection, which avoids the degradation problem of the deep neural network and can better extract the pollution degree features in the full-band spectral data. Compared with the traditional network with simple stacking of convolution layers, its feature extraction capability is better and it can capture more complex and abstract feature information; the Neck part further extracts, upsamples and concatenates the features extracted by different residual blocks to form a new feature map; this multi-dimensional feature fusion strategy helps the model capture features of different levels and dimensions and the contextual relationships between them, makes full use of the correlation between the features, and more accurately understands and characterizes the relevant information of the insulator pollution degree.

[0040] The model training of the present invention is more efficient: the residual structure enables the network to converge faster during the training process, reduces the time and number of iterations required for training, and improves the training efficiency of the model. During the back propagation and parameter update process, the optimizer is used and the adaptive learning rate adjustment is used to accelerate the training, so that the model can find the optimal parameter combination more quickly, further improving the training efficiency and model performance, and enhancing the adaptability of the model to different data distributions and characteristics.

[0041] The prediction accuracy of the present invention is higher: through the coordinated work of Backbone, Neck and Head, accurate prediction from spectral information to salt and grayscale values ​​is achieved; compared with traditional detection methods or shallower network models, this deep neural network can automatically learn a more complex mapping relationship between spectral data and contamination degree, thereby improving the accuracy and reliability of the prediction.

[0042] The present invention uses the mean square error as the loss function, which can effectively measure the difference between the predicted value and the true label. By continuously optimizing the network parameters, the mean square error between the predicted value and the true value is minimized, thereby further improving the prediction accuracy, making the salt and grayscale values ​​output by the model closer to the true value, and more accurately reflecting the degree of contamination of the insulator. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic diagram of the structure of the quantitative monitoring model of insulator contamination degree of the present invention.

[0044] Figure 2 The residual convolution module of the present invention

[0045] Figure 3 The salt and ash amount detection effect of the field pollution sampling cloth of the present invention, including (a) salt amount detection effect, (b) ash amount detection effect. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] The experimental methods used in the following examples are conventional methods unless otherwise specified. The materials, reagents, methods and instruments used are conventional materials, reagents, methods and instruments in the art unless otherwise specified, and can be obtained through commercial channels by those skilled in the art.

[0048] A quantitative detection system for insulator contamination based on residual structure convolutional neural network:

[0049] The insulator contamination quantitative detection system includes Backbone layer, Neck layer and Head layer:

[0050] The Backbone layer is used to perform preliminary feature extraction on the input full-band spectral data, including four Figure 2 The residual convolution module shown in the figure is composed of a main path and a jump-layer connection path;

[0051] The main path is responsible for deep feature transformation and abstraction of the input features, while the jump-layer connection path is used to establish a direct connection between the input and the specific layer of the main path to achieve identity mapping and promote effective propagation of gradients.

[0052] The main path includes convolution layer 1, activation layer 1, convolution layer 2, average pooling layer, and activation layer 2;

[0053] Define the network input feature as a one-dimensional feature vector x, first pass it to the convolution layer 1 to preliminarily extract the local feature information of the feature vector x, output it to the activation layer 1 to introduce nonlinear factors through the activation function, and the features processed by the activation layer 1 are input to the convolution layer 2 for feature extraction and combination. The features output by the convolution layer 2 enter the average pooling layer to adjust the size, and then enter the activation layer 2. The output feature after the activation layer 2 is F(x);

[0054] The jump-layer connection path includes only one convolution layer. Because the dimension of the main path is inconsistent with the input after two layers of convolution, in order to satisfy the identity mapping, the dimension of the input x is adjusted through the convolution layer of the jump-layer connection path and then added to F(x), and finally the added feature vector is output.

[0055] The input feature of the network is defined as a one-dimensional feature vector x, and the model fitting ability is enhanced through two convolutional layers and an activation function. In addition, a residual structure with jump-layer connection is adopted. Since the dimension after two layers of convolution is inconsistent with the input, in order to meet the identity mapping, a convolutional layer is added to adjust the dimension of the input x and then added to F(x), and finally the added feature vector is output. This output feature vector contains both the feature information F(x) after deep extraction and abstraction of the main path, and integrates the key information of the original input feature vector x. Through this residual structure design, the network can effectively avoid the problem of gradient vanishing or gradient explosion when the depth increases, realize the identity mapping, and improve the network's feature extraction ability and training effect.

[0056] The present invention selects a residual structure of jump-layer connection instead of a simple stacked convolutional layer, so that the network has an identity mapping capability, which is well avoided in terms of the degradation problem of deep neural networks and shows a faster convergence speed during training.

[0057] The Neck layer further extracts, upsamples, and concatenates the features extracted from the fourth, third, and second residual blocks to form three new feature maps. Upsampling adjusts the shape of high-dimensional features so that high-level features and low-level features can be concatenated together. On the other hand, it can reconstruct the detailed features of the hyperspectrum, thereby improving the network's feature expression capabilities. The reason why the feature fusion strategy is adopted in this step is that this fusion helps the model capture features of different dimensions and their contextual relationships, thereby improving the model's feature extraction capabilities and ultimately achieving the goal of improving task accuracy.

[0058] Head: This layer first flattens the three feature maps passed by the neck layer. The three feature vectors after flattening are concatenated into one feature vector after weighted calculation by two linear layers. The feature vector is then integrated through information of five linear layers to finally realize the prediction from spectral information to salt and grayscale values.

[0059] The network input is the preprocessed full-band hyperspectral line data, including feature vectors and corresponding true labels; the network output is the predicted salt ash content value. The mean square error (MSE) is used as the loss function.

[0060] In order to evaluate the effect of the model constructed by the present invention, a certain number of insulator sampling cloth hyperspectral data were collected, and after processing and annotation, a quantitative monitoring data set for insulator contamination was formed. The data set was input into XGBoost, convolutional neural network and the residual network constructed by the present invention for comparison to evaluate the model performance, and the root mean square error, mean absolute error and determination coefficient results on the validation set samples were used as evaluation indicators.

[0061] A quantitative detection method for insulator contamination based on residual structure convolutional neural network:

[0062] The method specifically comprises the following steps:

[0063] Step 1: Extract hyperspectral line data for the polluted sampling cloth area, and divide the full-band spectral reflectance data that has been spectrally preprocessed into a training set and a test set according to the ratio of spectral division;

[0064] A special contamination sampling cloth is placed in the insulator contamination area, and samples of the contamination sampling cloth are collected after a period of time; a hyperspectral imager is used to collect spectral data of the contamination sampling cloth to obtain full-band spectral reflectance data; the collected raw spectral data is denoised to eliminate interference, baseline correction is performed to eliminate baseline offset, and then normalization is performed to enhance data stability.

[0065] Step 2: Randomly initialize network parameters, including weights and biases of feature extraction network and regression layer;

[0066] Step 3: Input the full-band spectrum data into the feature extraction network to extract the pollution degree feature and obtain the feature vector;

[0067] The four residual convolution modules of the Backbone layer are used to extract the dirtiness feature and obtain the feature vector.

[0068] The main path of the residual convolution module extracts feature information, and the jump-layer connection path adjusts the dimension to ensure identical mapping.

[0069] Step 4: Input the extracted feature vector into the regression layer for regression training, use the mean square error as the loss function, and calculate the error between the predicted value and the true label;

[0070] Use mean square error (MSE) as the loss function to calculate the error between the predicted value and the true label (known salt and grayscale values);

[0071] The calculation formula of the MSE is:

[0072] Where n is the number of samples, y i is the true label value of the i-th sample, is the predicted value of the i-th sample.

[0073] By calculating the MSE, we can intuitively reflect the degree of deviation between the model prediction value and the true value, and provide a basis for subsequent parameter updates.

[0074] Step 5: Back propagation and parameter update: Use the optimizer to update the network parameters and accelerate the training process by adaptively adjusting the learning rate so that the regression model can show better predictive performance on the data.

[0075] Based on the calculated loss function value, the back propagation algorithm is used to calculate the gradient. Starting from the regression layer, the gradient is back propagated along the network structure, and the gradient values ​​of the weights and biases of each layer are calculated in turn. During the calculation process, according to the chain rule, the partial derivative of the loss function to each neuron output is multiplied by the partial derivative of the neuron to the input, and gradually passed to the input layer of the network to obtain the gradient information of each parameter.

[0076] Choose a suitable optimizer, such as the Adam optimizer, and use different learning rates when updating different parameters to improve the efficiency and stability of parameter updates.

[0077] Set the initial learning rate and adjust it dynamically during training according to the set learning rate adjustment strategy.

[0078] According to the calculated gradient and the optimizer's update rules, the network's weights and biases are updated. By repeating the back-propagation and parameter update process, the network's loss function value gradually decreases, and the model's predictive performance on the training data continues to improve.

[0079] Step 6: Model prediction: The unknown sample to be tested is input into the trained detection model to obtain the detection result.

[0080] The feature vector of the sample to be tested is input into the trained detection model. According to the forward propagation process of the model, after the feature extraction of Backbone, the feature fusion and processing of Neck, and the information integration and prediction of Head, the predicted salt and grayscale values ​​are finally obtained as the detection results.

[0081] During the experiment, the data sets were divided into training set and validation set in a ratio of 7:3. The model training rounds were set to 250, and the Adam optimizer was used. The stepLR scheduler was used, and the initial learning rate was set to 0.001 and decayed to the original 0.8 every fifty rounds of training. The MSE loss function was used as the loss function.

[0082] The final training results are shown in the following table:

[0083] Table 1 Performance parameters of three detection models

[0084]

[0085] According to the table analysis, the residual convolutional neural network shows significant performance advantages in both salt and gray detection. In terms of salt detection, the root mean square error of the test set of the GA-XGBoost model is 6.50mg, the mean absolute error is 5.80mg, and the determination coefficient is 0.84; while the performance indicators of the residual convolutional neural network are 2.71mg, 2.07mg, and 0.95. This shows that the residual convolutional neural network significantly and effectively reduces the error in salt detection and significantly improves the accuracy and reliability of prediction. In terms of gray detection, the advantages of the residual convolutional neural network can also be observed. The root mean square error of the GA-XGBoost model is 15.87mg, the mean absolute error is 12.63mg, and the determination coefficient is 0.84; while the performance indicators of the residual convolutional neural network are 10.27mg, 8.48mg, and 0.93. This shows that the residual convolutional neural network has lower error and higher fitting degree in gray volume detection, and has superior performance compared with other models.

[0086] In general, the residual convolutional neural network showed strong feature extraction ability and data fitting effect on the pollution sampling cloth dataset, and showed lower error and higher overall fitting degree in salt and ash detection. The visualization of its true value and predicted value is shown in the figure below. Figure 3 .

[0087] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0088] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the steps of the above method are implemented.

[0089] The memory in the embodiments of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory, ROM, a programmable read-only memory, PROM, an erasable programmable read-only memory, EPROM, an electrically erasable programmable read-only memory, EEPROM, or a flash memory. The volatile memory may be a random access memory, RAM, which is used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static random access memory static RAM, SRAM, dynamic random access memory dynamic RAM, DRAM, synchronous dynamic random access memory synchronous DRAM, SDRAM, double data rate synchronous dynamic random access memory double data rate SDRAM, DDR SDRAM, enhanced synchronous dynamic random access memory enhanced SDRAM, ESDRAM, synchronous link dynamic random access memory synchlink DRAM, SLDRAM, and direct memory bus random access memory direct rambus RAM, DR RAM. It should be noted that memory of the methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0090] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website site, computer, server or data center through a wired method such as coaxial cable, optical fiber, digital subscriber line digital subscriber line, DSL or wireless such as infrared, wireless, microwave, etc. to another website site, computer, server or data center. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media integrated. The available medium can be a magnetic medium such as a floppy disk, a hard disk, a tape, an optical medium such as a high-density digital video disc digital video disc, DVD, or a semiconductor medium such as a solid state hard disk solid state disc, SSD, etc.

[0091] In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in a processor or an instruction in the form of software. The steps of the method disclosed in conjunction with the embodiment of the present application can be directly embodied as a hardware processor for execution, or a combination of hardware and software modules in a processor for execution. The software module can be located in a storage medium mature in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in a memory, and the processor reads the information in the memory and completes the steps of the above method in conjunction with its hardware. To avoid repetition, it is not described in detail here.

[0092] It should be noted that the processor in the embodiment of the present application can be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method embodiment can be completed by an integrated logic circuit of hardware in the processor or an instruction in the form of software. The above processor can be a general-purpose processor, a digital signal processor DSP, an application-specific integrated circuit ASIC, a field programmable gate array FPGA or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in the embodiment of the present application can be directly embodied as a hardware decoding processor to perform, or the hardware and software modules in the decoding processor can be combined to perform. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in a memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0093] The above is a detailed introduction to the insulator contamination quantitative detection system and method based on the residual structure convolutional neural network proposed in the present invention, and the principles and implementation methods of the present invention are explained. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. The quantitative detection system of insulator contamination based on residual structure convolutional neural network is characterized by: The insulator contamination quantitative detection system includes Backbone layer, Neck layer and Head layer: The Backbone layer is used to perform preliminary feature extraction on the input full-band spectral data, including four residual convolution modules. The Neck layer further extracts, upsamples, and concatenates the features extracted from the fourth, third, and second residual blocks to form three new feature maps. The Head layer first flattens the three feature maps transmitted by the Neck layer. The three unfolded feature vectors are concatenated into one feature vector after weighted calculation by two linear layers. The feature vector is then integrated through information of five linear layers to finally realize the prediction from spectral information to salt and grayscale values.

2. The insulator contamination quantitative detection system according to claim 1 is characterized in that: The residual convolution module consists of a main path and a jump-layer connection path; The main path includes convolution layer 1, activation layer 1, convolution layer 2, average pooling layer, and activation layer 2; Define the network input feature as a one-dimensional feature vector x, first pass it to the convolution layer 1 to preliminarily extract the local feature information of the feature vector x, output it to the activation layer 1 to introduce nonlinear factors through the activation function, and the features processed by the activation layer 1 are input to the convolution layer 2 for feature extraction and combination. The features output by the convolution layer 2 enter the average pooling layer to adjust the size, and then enter the activation layer 2. The output feature after the activation layer 2 is F(x); The jump-layer connection path includes only one convolution layer. Because the dimension of the main path is inconsistent with the input after two layers of convolution, in order to satisfy the identity mapping, the dimension of the input x is adjusted through the convolution layer of the jump-layer connection path and then added to F(x), and finally the added feature vector is output.

3. The insulator contamination quantitative detection system according to claim 2 is characterized in that: The input of the residual structure convolutional neural network is the preprocessed full-band hyperspectral spectrum line data, including feature vectors and corresponding true labels; the network output is the predicted salt ash content value; and the mean square error (MSE) is used as the loss function.

4. A detection method for executing the insulator contamination quantitative detection system based on residual structure convolutional neural network according to any one of claims 1 to 3, characterized in that: The method specifically comprises the following steps: Step 1: Extract hyperspectral line data for the polluted sampling cloth area, and divide the full-band spectral reflectance data that has been spectrally preprocessed into a training set and a test set according to the ratio of spectral division; Step 2: Randomly initialize the network parameters, including the weights and biases of the feature extraction network and regression layer; Step 3: Input the full-band spectrum data into the feature extraction network to extract the pollution degree feature and obtain the feature vector; Step 4: Input the extracted feature vector into the regression layer for regression training, use the mean square error as the loss function, and calculate the error between the predicted value and the true label; Step 5: Back propagation and parameter update: Use the optimizer to update the network parameters and accelerate the training process by adaptively adjusting the learning rate so that the regression model can show better predictive performance on the data. Step 6: Model prediction: The unknown sample to be tested is input into the trained detection model to obtain the detection result.

5. The method according to claim 4, characterized in that: In step one, Place a special contamination sampling cloth in the contamination area of ​​the insulator and collect contamination sampling cloth samples after a period of time; Use a hyperspectral imager to collect spectral data of the polluted sampling cloth to obtain full-band spectral reflectance data; The collected raw spectral data were denoised to eliminate interference, baseline corrected to eliminate baseline offset, and then normalized to enhance data stability.

6. The method according to claim 5, characterized in that: In step three, The four residual convolution modules of the Backbone layer are used to extract the dirtiness feature and obtain the feature vector. The main path of the residual convolution module extracts feature information, and the jump-layer connection path adjusts the dimension to ensure identical mapping.

7. The method according to claim 6, characterized in that: In step four, The calculation formula of the MSE is: Where n is the number of samples, y i is the true label value of the i-th sample, is the predicted value of the i-th sample.

8. The method according to claim 6, characterized in that: In step five, Based on the calculated loss function value, the back-propagation algorithm is used to calculate the gradient; starting from the regression layer, the gradient is back-propagated along the network structure, and the gradient values ​​of the weights and biases of each layer are calculated in turn; during the calculation process, according to the chain rule, the partial derivative of the loss function with respect to each neuron output is multiplied by the partial derivative of the neuron with respect to the input, and then gradually passed to the input layer of the network to obtain the gradient information of each parameter.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method described in claim 8 are implemented.

10. A computer-readable storage medium for storing computer instructions, characterized in that: When the computer instructions are executed by a processor, the steps of the method described in claim 8 are implemented.

Images