A method and device for accurately identifying contamination of porcelain insulators in high-voltage direct current transmission lines

Abstract

The present application provides a method and device for accurately identifying contamination of porcelain insulators of a high-voltage direct current transmission line, and relates to the field of insulator contamination detection. The method comprises: constructing a contamination data set through a contamination reagent; preprocessing the contamination data set; constructing a training set and a validation set through the preprocessed contamination data set and formatting and tensorizing them; constructing a classification network model; training the classification network model through the training set and the validation set; obtaining a hyperspectral contamination image of an insulator to be classified; accurately identifying the contamination components of the hyperspectral contamination image of the insulator through the trained classification network model; and constructing a model based on a multi-scale three-dimensional convolutional block to efficiently extract two-dimensional multi-scale spatial features and one-dimensional spectral features of hyperspectral image data of insulator contamination, thereby realizing accurate identification of contamination components of different types of insulators.

Description

Technical Field

[0001] The present application relates to the field of insulator contamination detection, and in particular to a method and device for accurately identifying contamination of porcelain insulators in high-voltage direct current transmission lines. Background Art

[0002] At present, the accurate identification of the contamination components of porcelain insulators of high-voltage DC transmission lines mainly relies on high-precision instruments, and the operation process is complicated and costly. Although the non-contact identification technology based on hyperspectral images has achieved certain results on simulated samples in laboratory environments, its application on naturally contaminated high-voltage DC insulators is still insufficient. At the same time, the traditional method of detecting the degree of contamination requires large equipment and rich operating experience, and has the disadvantages of large errors and high costs.

[0003] In the technical field of insulator contamination component identification, the emerging non-contact image processing technology has made some progress. The dimensions of insulator contamination image data are diverse, including one-dimensional data (such as time series data), two-dimensional data (such as grayscale images or color images) and three-dimensional data (such as hyperspectral images or stereo images). Each dimension of data contains specific information, which is of great significance for improving the recognition accuracy.

[0004] However, the existing insulator contamination image dataset is obtained by shooting, the dataset quality is low, and the complementarity between data of different dimensions is ignored, resulting in the model's inaccurate recognition of contaminants. Summary of the invention

[0005] The purpose of the present invention is to provide a method and device for accurately identifying contamination on porcelain insulators of high-voltage direct current transmission lines in order to solve the problem that the existing insulator image processing technology is not accurate enough in identifying contaminants.

[0006] The above-mentioned purpose of the present application is achieved through the following technical solutions: S1: Construct a contamination dataset through contamination reagents; S2: Preprocess the dirty data set; S3: Use the preprocessed dirty data set to construct the training set and validation set and perform formatting and tensorization; S4: Build a classification network model; S5: Train the classification network model using the training set and validation set; S5: Obtain the insulator hyperspectral pollution image to be classified; accurately identify the pollution components of the insulator hyperspectral pollution image through the trained classification network model.

[0007] Optionally, step S1 includes: S11: Glass and rubber are selected as the base materials of insulators; S12: selecting a contamination reagent, the contamination reagent comprising: kaolin, iron sulfate, sodium chloride, iron oxide, anhydrite, quartz sand and calcium carbonate; S13: Each contamination reagent is evenly applied on the substrate to prepare an insulator sample; S14: Place the prepared insulator sample on the platform of the data acquisition system to obtain hyperspectral imaging and construct a pollution data set.

[0008] Optionally, step S2 includes: preprocessing including: lens correction, black and white correction, manual extraction and denoising.

[0009] Optionally, step S3 includes: S31: normalize the hyperspectral image data in the pollution dataset; S32: construct a sample data set by combining the normalized hyperspectral image data with the neighborhood set pixel generation method; S33: Enhance the sample data set by rotation, scaling, translation and flipping; S34: Generate a label set corresponding to the sample data set according to the type and concentration of the contamination reagent; S35: Construct training set and validation set through label set and sample data set.

[0010] Optionally, step S4 includes: A classification network model is constructed through the residual connection module mechanism and convolutional neural network.

[0011] Optionally, step S5 includes: S51: Introduce a learning rate scheduling mechanism based on dynamic adjustment of the validation set accuracy and set a learning rate halving strategy; S52: The classification network model is trained by combining the training set and the validation set through L2 regularization technology, random depth technology and learning rate halving strategy.

[0012] An electronic device comprises a processor, a memory, a user interface and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes a method for accurately identifying contamination of porcelain insulators of high-voltage direct current transmission lines.

[0013] A computer-readable storage medium stores instructions. When the instructions are executed, a method for accurately identifying contamination of porcelain insulators of a high-voltage direct current transmission line is executed.

[0014] The beneficial effects of the technical solution provided by this application are: The key point of the present invention is to propose an innovative insulator pollution component recognition model, which cleverly integrates multi-scale three-dimensional convolution blocks and constructs a convolutional neural network model based on the multi-scale three-dimensional convolution block and the residual connection module mechanism. It can jointly extract the two-dimensional multi-scale spatial features and one-dimensional spectral features of the hyperspectral image data of insulator pollution in an end-to-end manner. Through this unique design, the model can fully capture the complex information of insulator pollution and achieve accurate distinction between different types of pollution components. In addition, the present invention also focuses on the optimization of data processing and model training. Through a series of preprocessing, enhancement, formatting and tensorization operations, as well as strategies such as dynamic adjustment of learning rate, L2 regularization, gradient clipping and random depth, the recognition performance and generalization ability of the model are significantly improved.

[0015] Through the design of multi-scale three-dimensional convolution blocks, optimization of data processing flow, adjustment of model training strategy and hyperparameter tuning methods, etc. At the same time, the present invention also protects the technical means and innovations used in key links such as data collection and preparation, data preprocessing, data sample label generation and enhancement, sample formatting and tensorization, network construction and hyperparameter setting. These protection points together constitute the unique technical barriers of the present invention, ensuring its leading position and practical application value in the field of insulator pollution component identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present application will be further described below with reference to the accompanying drawings and embodiments, in which: Figure 1 is a step diagram in an embodiment of the present application; Figure 2 It is a model structure diagram in the embodiment of the present application; Figure 3 It is a schematic diagram of the structure of an electronic device in an embodiment of the present application. DETAILED DESCRIPTION

[0017] In order to have a clearer understanding of the technical features, purposes and effects of the present application, the specific implementation methods of the present application are now described in detail with reference to the accompanying drawings.

[0018] The embodiment of the present application provides a method for accurately identifying contamination of porcelain insulators in a high-voltage direct current transmission line.

[0019] Please refer to Figure 1 , Figure 1 This is a step diagram of a method for accurately identifying contamination of porcelain insulators of a high-voltage direct current transmission line in an embodiment of the present application, comprising: S1: Construct a contamination dataset through contamination reagents; S2: Preprocess the dirty data set; S3: Use the preprocessed dirty data set to construct the training set and validation set and perform formatting and tensorization; The present application provides an embodiment as follows: according to the dimension (one-dimensional or three-dimensional) of the selected model, a data set in a format corresponding to the interface is generated. Tensorization operations are performed on all samples to generate multi-dimensional vectors to improve the training efficiency of the neural network.

[0020] According to the input requirements of the selected model, the format of the data set is unified. For example, if the model requires the input to be a one-dimensional vector, the original data hyperspectral image (three-dimensional, containing two-dimensional spatial information and one-dimensional spectral information) needs to be converted to a one-dimensional format (arrange each pixel of the image in sequence into a long vector).

[0021] For one-dimensional data, if the original data is multidimensional (such as time series data), dimensionality reduction is required to convert it into a one-dimensional vector. PCA (Principal Component Analysis) or other dimensionality reduction techniques can be used to reduce the dimensionality of the data while retaining key information.

[0022] For three-dimensional data (such as image data), if the original data is two-dimensional (such as grayscale images), it needs to be converted to a three-dimensional format, usually by adding a channel dimension to the two-dimensional data. If the original data is already three-dimensional (such as color images), its three-dimensional structure is directly retained.

[0023] According to the input requirements of the model, the cleaned and converted data is constructed into a multidimensional tensor (a three-dimensional tensor with three dimensions: height, width, and color channels). The dimensions of the tensor should match the dimensions of the model's input layer to ensure that the data can be correctly passed to the model. The tensor data is standardized or normalized to improve the training efficiency and performance of the neural network. Standardization usually scales the data to a range of mean 0 and variance 1; while normalization scales the data to between 0 and 1.

[0024] Perform format verification on the processed data set to ensure that the dimension, type, and range of the data are consistent with the input requirements of the model. If format mismatch is found, data cleaning and conversion need to be performed again.

[0025] Dataset division: The processed dataset is divided into training set, validation set and test set. The training set is used for model training; the validation set is used for model selection and parameter adjustment during training; the test set is used to evaluate the final performance of the model. The division ratio should be determined according to the specific task and the size of the dataset. 2D multi-scale spatial feature extraction: The model first uses a 2D convolutional layer to extract spatial features for each spectral channel of the hyperspectral image. By stacking multiple 2D convolutional layers and introducing a pooling layer to reduce the resolution of the feature map, the model can gradually abstract the spatial distribution characteristics of the pollution components.

[0026] One-dimensional spectral feature extraction: While extracting spatial features, the model also uses a one-dimensional convolutional layer to extract features from the spectral dimension of the hyperspectral image. This helps capture the unique responses of pollutant components in different spectral bands and provides key information for subsequent identification.

[0027] Feature fusion: The extracted two-dimensional spatial features and one-dimensional spectral features are fused through a weighted sum fusion strategy to form a joint feature representation. This step ensures that the model can fully utilize all the information of the hyperspectral image data and improve the comprehensiveness and accuracy of recognition.

[0028] The specific formula of the weighted sum fusion strategy is as follows: For any sample point, the corresponding feature vectors in the one-dimensional data and two-dimensional data are α and β respectively, and the fused feature vector is γ. The weighted sum fusion formula can be expressed as:

[0029] Among them, w1 and w2 are the weights of α and β, respectively, indicating the importance of their respective feature vectors in the fusion process. The value range of the weight is usually between 0 and 1, and satisfies w1 + w2 = 1.

[0030] The fused feature vector γ is the weighted sum of the original feature vectors α and β, and the weights w1 and w2 determine their contribution to the fused feature.

[0031] S4: Build a classification network model; S5: Train the classification network model using the training set and validation set; S5: Obtain the insulator hyperspectral pollution image to be classified; accurately identify the pollution components of the insulator hyperspectral pollution image through the trained classification network model.

[0032] Step S1 includes: S11: Glass and rubber are selected as the base materials of insulators; S12: selecting a contamination reagent, the contamination reagent comprising: kaolin, iron sulfate, sodium chloride, iron oxide, anhydrite, quartz sand and calcium carbonate; As an example, in order to comprehensively evaluate the pollution components of insulators, seven typical pollution reagents are selected, including kaolin, iron sulfate, sodium chloride, iron oxide, anhydrite, quartz sand and calcium carbonate. These reagents cover various types such as inorganic substances, organic substances and salts, and can reflect the pollution conditions that insulators may encounter in actual environments.

[0033] S13: Each contamination reagent is evenly applied on the substrate to prepare an insulator sample; As an example, the contamination method is to evenly apply the contamination reagent on the insulator substrate with a brush or a scraper. The single and mixed contamination samples with a size of 5cm×5cm are prepared by the contamination method. For single contamination samples, each contamination reagent is evenly applied on the insulator substrate to ensure a consistent coating thickness. For mixed contamination samples, different proportions of contamination reagents are mixed and applied on the insulator substrate to simulate the complex contamination conditions that may exist in the actual environment.

[0034] S14: Place the prepared insulator sample on the platform of the data acquisition system to obtain hyperspectral imaging and construct a pollution data set.

[0035] As an embodiment, an advanced hyperspectral data acquisition system is used, which has high-resolution spectral imaging capabilities and can capture the spectral information of samples in different bands. The system usually includes components such as a spectrometer, a light source, a lens, and an image acquisition card, which can achieve accurate spectral imaging of the sample. Place the prepared insulator sample on the platform of the data acquisition system to ensure that the sample surface is flat and the light is uniform. Then start the data acquisition system to perform hyperspectral imaging on the sample. During the imaging process, the system automatically adjusts the working status of the light source, lens, and spectrometer according to the set parameters to obtain high-quality hyperspectral image data. The collected hyperspectral image data needs to be saved in a computer for subsequent data processing and analysis. This patent will save the data in a specific file format (such as ENVI, HDF5, etc.), which can support the storage and reading of hyperspectral data.

[0036] Step S2 includes: preprocessing including: lens correction, black and white correction, manual extraction and denoising.

[0037] As an embodiment, first, preliminary data analysis is performed by MATLAB software or Python software, including lens correction and black and white correction, to ensure the accuracy of subsequent processing. Subsequently, manual extraction (such as using the ROI extraction function in ENVI software, circled one by one on the template by a circle, rectangle, custom curve, etc.) or automatic extraction (such as a method based on threshold segmentation) is used to extract the region of interest. Smoothing method (eliminating the error caused by random and mechanical noise, making the wavelength information obtained smoother), derivative method (by doing derivative processing on spectral information, the interference of background information can be effectively reduced, the image information is clearer, and the resolution is improved) and other filtering methods are used to remove the noise in the image, and smooth the spectral curve to improve the accuracy and reliability of the data.

[0038] Step S3 includes: S31: normalize the hyperspectral image data in the pollution dataset; As an example, the hyperspectral image data in the pollution data set is quality checked to ensure that the image is clear and free of noise interference. The background information in the image is removed, and only the insulator and its pollution part are retained to reduce the interference of irrelevant information. The image data is normalized and the pixel values ​​are scaled to between 0 and 1 to eliminate the brightness differences between different images.

[0039] S32: construct a sample data set by combining the normalized hyperspectral image data with the neighborhood set pixel generation method; As an embodiment, in order to make full use of the spatial information of the hyperspectral image, a neighborhood set pixel generation method is adopted. For each pixel point, the pixel values ​​within a certain range around it are extracted as neighborhood features to form a new feature vector. The size of the neighborhood range can be adjusted according to actual needs to balance the utilization of spatial information and computational complexity. The neighborhood feature vector of each pixel point is used as part of the spectral line set of the pixel point. For the entire image, the spectral line sets of all pixels are merged to form a complete sample data set.

[0040] S33: Enhance the sample data set by rotation, scaling, translation and flipping; S34: Generate a label set corresponding to the sample data set according to the type and concentration of the contamination reagent; S35: Construct training set and validation set through label set and sample data set.

[0041] As an embodiment, in order to increase the diversity and robustness of samples, this patent adopts a variety of data enhancement methods, including rotation, scaling, translation, flipping, etc., to simulate various changes that may occur in the insulator pollution image in the actual environment, thereby improving the generalization ability of the model. Different data enhancement methods are applied to each original sample to generate multiple enhanced samples. All enhanced samples are merged into the original pollution data set to form a final set of spectral lines of different pollution components with a total number of samples of 5000 to 20000.

[0042] As an example, a corresponding label set is generated for each set of sample data according to the actual components of the insulator contamination. The label set includes key information such as the type and concentration of the contamination components, which is used for subsequent classification and regression tasks.

[0043] Step S4 includes: A classification network model is constructed through the residual connection module mechanism and convolutional neural network.

[0044] The present application provides an embodiment as follows: a classification network model is defined, a training set and a validation set are used as input, and network hyperparameters are set, including the selection of pixel set size (patch size), learning rate, optimizer, and loss function. In particular, the Kaiming initialization method is used to initialize the network for the ReLU activation function to avoid gradient explosion or gradient disappearance problems. The structural diagram of the classification network model is as follows: Figure 2 shown.

[0045] The present application provides an embodiment as follows: according to the specific requirements of the classification task, a suitable network architecture is selected. For example, for image classification tasks, a convolutional neural network (CNN) with a residual connection module mechanism is selected. Ensure that the network architecture can handle the dimensions and formats of the input data, such as processing two-dimensional images or three-dimensional data. Determine the number of layers of the network and the type of each layer, such as convolutional layers, pooling layers, fully connected layers, etc. Adjust parameters such as the number of neurons in each layer and the size of the convolution kernel according to the complexity of the task. For networks that process image data, set a suitable pixel set size to capture local features in the image. Adjust the size of the pixel set size according to the image resolution and feature scale.

[0046] The present application provides an embodiment as follows: according to the task requirements and model complexity, select a suitable optimizer, such as SGD, Adam or RMSprop, etc. The parameters of the optimizer (such as momentum, weight decay, etc.) also need to be adjusted according to the specific situation. Select a loss function that matches the classification task, such as cross entropy loss, focal loss, etc. Adjust the parameters of the loss function according to the task requirements and data distribution.

[0047] The present application provides an embodiment as follows, using the Kaiming initialization method (also known as He initialization) to initialize the network weights for the ReLU activation function.

[0048] Step S5 includes: S51: Introduce a learning rate scheduling mechanism based on dynamic adjustment of the validation set accuracy and set a learning rate halving strategy; S52: The classification network model is trained by combining the training set and the validation set through L2 regularization technology, random depth technology and learning rate halving strategy.

[0049] The present application provides an embodiment as follows, introducing a learning rate scheduling mechanism based on dynamic adjustment of the validation set accuracy, setting a learning rate halving strategy, that is, when the validation set accuracy fails to improve in multiple consecutive cycles and reaches the set stagnation threshold, the current learning rate is halved. In addition, L2 regularization technology is used to limit model complexity and reduce the risk of overfitting. At the same time, a gradient clipping strategy is implemented to limit the range of gradient values ​​to avoid the gradient explosion problem, and combined with the use of random depth technology, that is, randomly selecting a different number of network layers for forward and backward propagation in each training iteration, further enhancing the generalization ability of the model.

[0050] The present application provides an embodiment as follows. To ensure a comprehensive evaluation of the model performance, the current state of the model is saved every 5 training cycles (epochs), and the performance of the model is verified using an independent validation set. In addition to conventional prediction accuracy calculations, multiple classification performance indicators such as F1 scores and AUC-ROC curves are introduced to more comprehensively measure the classification effect and generalization ability of the model.

[0051] This application provides an embodiment as follows: in terms of hyperparameter tuning, a strategy combining grid search and random search is used to systematically explore the impact of different hyperparameter combinations on model performance within a preset hyperparameter space. At the same time, this patent introduces the Bayesian Optimization method, using previous experimental results to guide subsequent search directions, so as to more efficiently find the optimal or approximately optimal hyperparameter configuration.

[0052] This application provides an embodiment as follows, and also uses cross-validation technology to evaluate the stability of the model under different hyperparameter combinations to reduce the performance evaluation deviation caused by improper data division. Through these comprehensive measures, this patent aims to continuously improve the performance of the model and provide a solid guarantee for its reliability in practical applications.

[0053] The present application provides an embodiment as follows: the model parameters are saved every 10 epochs, the test set sample spectrum is input into the model interface to test the network performance, and the prediction accuracy is calculated.

[0054] Table 1 Prediction accuracy of different combinations of contamination reagents

[0055] Table 1 shows the prediction accuracy of insulating sheet samples prepared with different contamination compounds by the smearing method. It can be seen that as the number of contamination reagent combinations increases, the contamination identification prediction accuracy increases. This is mainly because the combination of multiple contamination reagents increases the diversity of data samples. This diversity allows the model to be exposed to more different contamination situations during training, thereby improving its generalization ability and prediction accuracy.

[0056] Table 2

[0057] Table 2 shows the prediction accuracy comparison of the convolutional neural network (CNN) with the addition of the residual connection module mechanism. The residual connection module implements residual learning by adding the input data directly to the output of certain network layers through skip connections. This mechanism makes the output of the network layer become the changing part of the input data, while the shortcut connection transmits the unchanged part. In this way, even in the case of a deep network, the gradient can be effectively propagated to the shallower layers, thus solving the gradient vanishing problem and improving the prediction accuracy of the model.

[0058] The present application also discloses an electronic device. Figure 3 , Figure 3 The electronic device 500 may include: at least one processor 501 , at least one network interface 504 , a user interface 503 , a memory 505 , and at least one communication bus 502 .

[0059] The communication bus 502 is used to realize the connection and communication between these components.

[0060] The user interface 503 may include a display screen, and the optional user interface 503 may also include a standard wired interface or a wireless interface.

[0061] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0062] The present application also discloses a computer-readable storage medium storing a plurality of instructions suitable for loading by a processor to execute the above-mentioned method for accurately identifying contamination of porcelain insulators of a high-voltage direct current transmission line.

[0063] The above are only exemplary embodiments of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure.

[0064] This application is intended to cover any variation, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art not described in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A method for accurately identifying contamination of porcelain insulators in high-voltage direct current transmission lines, characterized in that: The method comprises the following steps: S1: Construct a contamination dataset through contamination reagents; S2: Preprocess the dirty data set; S3: Use the preprocessed dirty data set to construct the training set and validation set and perform formatting and tensorization; S4: Build a classification network model; S5: Train the classification network model using the training set and validation set; S6: Obtain the insulator hyperspectral pollution image to be classified; accurately identify the pollution components of the insulator hyperspectral pollution image through the trained classification network model.

2. A method for accurately identifying contamination of porcelain insulators in a high-voltage direct current transmission line according to claim 1, characterized in that: Step S1 includes: S11: Glass and rubber are selected as the base materials of insulators; S12: selecting a contamination reagent, the contamination reagent comprising: kaolin, iron sulfate, sodium chloride, iron oxide, anhydrite, quartz sand and calcium carbonate; S13: Each contamination reagent is evenly applied on the substrate to prepare an insulator sample; S14: Place the prepared insulator sample on the platform of the data acquisition system to obtain hyperspectral imaging and construct a pollution data set.

3. The method for accurately identifying contamination of porcelain insulators in a high-voltage direct current transmission line according to claim 1, characterized in that: Step S2 includes: preprocessing including: lens correction, black and white correction, manual extraction and denoising.

4. A method for accurately identifying contamination of porcelain insulators in a high-voltage direct current transmission line according to claim 1, characterized in that: Step S3 includes: S31: normalize the hyperspectral image data in the pollution dataset; S32: construct a sample data set by combining the normalized hyperspectral image data with the neighborhood set pixel generation method; S33: Enhance the sample data set by rotation, scaling, translation and flipping; S34: Generate a label set corresponding to the sample data set according to the type and concentration of the contamination reagent; S35: Construct training set and validation set through label set and sample data set.

5. The method for accurately identifying contamination of porcelain insulators in a high-voltage direct current transmission line according to claim 1, characterized in that: Step S4 includes: A classification network model is constructed through the residual connection module mechanism and convolutional neural network.

6. A method for accurately identifying contamination of porcelain insulators in a high-voltage direct current transmission line according to claim 1, characterized in that: Step S5 includes: S51: Introduce a learning rate scheduling mechanism based on dynamic adjustment of the validation set accuracy and set a learning rate halving strategy; S52: The classification network model is trained by combining the training set and the validation set through L2 regularization technology, random depth technology and learning rate halving strategy.

7. An electronic device, characterized in that: The electronic device comprises a processor (501), a memory (505), a user interface (503) and a network interface (504), wherein the memory (505) is used to store instructions, the user interface (503) and the network interface (504) are used to communicate with other devices, and the processor (501) is used to execute the instructions stored in the memory (505) so that the electronic device executes the method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed by a computer, the method according to any one of claims 1 to 6 is executed.

Images