Substation meter identification method and system for sandstorm environment engineering

Abstract

The invention discloses a substation meter identification method and system for wind and sand environment engineering. The method comprises the following steps: acquiring a substation meter image; the collected images are preprocessed; feature extraction is carried out on the preprocessed image by using a deep learning model, and key features in the meter image are automatically learned and extracted; detecting and positioning the transformer substation meter in the image according to the extracted key features; performing pattern recognition processing on the positioned meter image by adopting a pattern recognition algorithm to recognize the meter reading, and integrating an automatic correction algorithm to improve the reading accuracy; a meter image is analyzed through a designed damage detection algorithm, and an early warning signal is sent out once a damage condition is detected. Through the combination of the deep learning algorithm and the mode recognition technology, the accuracy and stability of recognition are improved, the influence of environmental factors on the recognition result is reduced, and the transformer substation meter can be efficiently and accurately recognized in the wind and sand environment.

Description

Technical Field

[0001] The invention belongs to the field of image processing, relates to substation meter recognition, and specifically relates to a substation meter recognition method and system for a sandstorm environment project. Background Art

[0002] With the rapid development of smart grid technology, substations are key nodes in the power system, and their intelligent inspection level is directly related to the safe and stable operation of the power grid. However, in areas with frequent sandstorms, traditional meter recognition methods are often severely interfered with, resulting in a significant decrease in recognition accuracy. This not only increases the burden of manual inspections, but may also cause safety accidents due to misreading or missing meter data. Therefore, it is particularly important to develop an algorithm that can efficiently and accurately identify substation meters in harsh environments.

[0003] At present, traditional pattern recognition algorithms are used to model each point, mark the size, range, pointer, appearance characteristics, etc. of the meter, generate a recognition model, and use the model to identify meters at fixed points and fixed angles. However, there are still shortcomings. Meters at different points need to be modeled separately, and the model is sensitive to distance, angle, lighting, etc., and the actual application cost is relatively high for scenarios with more points. Summary of the invention

[0004] Purpose of the invention: In order to overcome the deficiencies in the prior art, a method and system for identifying substation meters in a sandstorm environment project is provided. By combining deep learning algorithms and pattern recognition technology, the accuracy and stability of identification are improved, the impact of environmental factors on the identification results is reduced, and substation meters can be efficiently and accurately identified in a sandstorm environment.

[0005] Technical solution: To achieve the above purpose, the present invention provides a method for identifying substation meters in a sandstorm environment project, comprising the following steps:

[0006] S1: Collect substation meter images;

[0007] S2: preprocessing the collected images;

[0008] S3: Use the deep learning model to extract features from the preprocessed images, automatically learn and extract key features from the meter images;

[0009] S4: Detect and locate the substation meters in the image based on the extracted key features;

[0010] S5: Using a pattern recognition algorithm to perform pattern recognition processing on the located meter image, identify the meter reading, and integrate an automatic correction algorithm to improve the reading accuracy;

[0011] S6: The meter image is analyzed by a designed damage detection algorithm, and a warning signal is immediately issued once damage is detected.

[0012] Furthermore, in step S1, the substation meter images are collected by an on-site inspection robot or a fixed camera. The inspection robot can carry a high-definition camera to automatically inspect and capture meter images in the substation; the fixed camera is installed at a key position of the substation to monitor the meter in real time.

[0013] Furthermore, the preprocessing in step S2 includes removing wind and sand, reducing noise, and enhancing contrast.

[0014] Furthermore, in step S2, in order to solve the image blur problem that may be caused by the sandstorm environment, a sandstorm removal algorithm based on dark channel prior is used to remove the sandstorm from the collected image, specifically:

[0015] The algorithm removes fog from the image by calculating the dark channel value of the image, statistical histogram, ambient light value and transmittance, etc. The image quality after removing the wind and sand is significantly improved, providing a better recognition basis for subsequent steps.

[0016] The minimum value of all pixels in each block in the R, G, and B components is taken as the dark channel value of the R, G, and B components of the block;

[0017] The method for calculating the dark channel I_dark is to first divide the input original image into multiple blocks, each block size is n×m, and the image height is h and the width is w, then the number of blocks should be (h×w) / (n×m); based on the above blocks, the dark channel I_dark is calculated as follows:

[0018]

[0019] Among them, min means taking the minimum value, Ω means the set of n×m; the dark channel I_dark is actually to find the Raw c Minimum value of Raw c Refers to the raw image data output by the image sensor containing the three components of R, G, and B.

[0020] Furthermore, in step S3, the deep learning model is used to extract features from the preprocessed image, and key features in the meter image are automatically learned and extracted, specifically:

[0021] Feature extraction includes node feature extraction, edge feature extraction, and image feature extraction;

[0022] The extraction of key features specifically includes:

[0023] The key features of edge feature extraction are extracted through global domain overlap features. The global domain overlap feature refers to the number of paths with different path lengths between two nodes, and the calculation formula of the global domain overlap feature is:

[0024] S=(I-β*A)-1-I

[0025] Where A is the adjacency matrix, and the result S[i,j] represents the number of paths with different path lengths between nodes i and j;

[0026] The key features of node feature extraction include degree and clustering coefficient. Degree reflects the number of connections of a node. The formula for extracting it by counting the number of edges in the adjacency matrix is ​​Degree(v)=∑Avu, where Avu indicates whether there is an edge between nodes and. The clustering coefficient measures the connectivity of node neighbors and is obtained by calculating the number of edges between a node and its neighbors. The calculation formula is:

[0027]

[0028] Where E is the actual number of edges and k is the node degree;

[0029] The key features of image feature extraction include color, texture, shape, local and deep learning features; color features represent pixel distribution through color histogram and calculate the frequency of each color; texture features are extracted through local binary pattern LBP or gray level co-occurrence matrix GLCM. The LBP formula is:

[0030]

[0031] Where s is the threshold function; shape features use edge detection algorithms to extract object contours; local features such as SIFT detect key points and their feature vectors in the image; deep learning features automatically extract high-level features through convolutional neural networks, and forward propagation generates feature maps.

[0032] Furthermore, the detection and positioning of the substation meter in the image in step S4 is specifically as follows:

[0033] Deep learning model training: Use YOLOv5s or similar object detection networks to train images after sand removal. Label the meter locations and types in a large number of sample images as training data input for model training. During the training process, continuously optimize the model parameters to improve the accuracy of positioning and classification.

[0034] Meter location and classification: The trained deep learning model can automatically locate and classify substation meters in the input image. The model extracts image features and matches them with pre-trained weights to achieve efficient and accurate location and classification output.

[0035] Furthermore, the pattern recognition processing in step S5 is as follows: the meter image is separated from the background by image segmentation technology; then the contour features of the meter pointer or numbers are identified by contour extraction technology; and finally the meter reading is calculated in combination with prior knowledge (such as meter range, pointer starting position, etc.). Specifically: first convert the color image into a grayscale image, then apply Gaussian blur to reduce noise, smooth the image, and enhance the contrast between the meter and the background; and binarize the image through threshold processing; next, extract the contour features of the meter pointer or numbers, and implement it through Canny edge detection and findContours function. The Canny edge detection algorithm is used to find the edges in the image, and the findContours function of OpenCV is used to detect the contours in the image; screen out the contours that match the meter pointer or digital features; use the prior knowledge of the meter to calculate the final reading, calculate the angle of the meter pointer through geometric methods, and calculate the reading in combination with the meter range. The angle of the pointer is calculated using the slope formula: angle = arctan (y2-y1 / x2-x1), and then mapped to the meter range according to the angle, and finally display the detected pointer and calculated reading on the image.

[0036] Furthermore, the correction process of the automatic correction algorithm in step S5 is: adding the automatic correction algorithm to adjust the distance, angle and lighting parameters of the image during the recognition process to reduce the impact of environmental factors on the recognition results; at the same time, adding the feature extraction algorithm to further extract node features, edge features, image features, etc. to enhance the robustness of recognition.

[0037] The correction process includes image preprocessing, distance correction, angle correction, illumination correction, feature extraction and enhancement, recognition and post-processing, specifically:

[0038] Image preprocessing: Use Gaussian blur or median filtering to remove noise; resize the image to unify the input size;

[0039] Distance correction: estimate distance through deep learning or traditional computer vision techniques; apply perspective transformation and adjust the image to eliminate the effect of distance;

[0040] Angle correction: Detect straight lines through Hough transform and estimate the image rotation angle; use the rotation matrix to perform image rotation correction.

[0041] Lighting correction: Use histogram equalization to enhance contrast; apply gamma correction to adjust brightness;

[0042] Feature extraction and enhancement: Use edge detection to extract edge point features; use SIFT or ORB to extract key point features; use machine learning algorithms to train the model to further improve the robustness of the features.

[0043] Recognition and post-processing: The extracted features are input into the classifier for target recognition. Based on the recognition results, a threshold or voting mechanism is applied to improve recognition accuracy.

[0044] Furthermore, the damage detection algorithm in step S6 is specifically:

[0045] A1: Damage sample collection and labeling: During the inspection process, images of damaged meters detected are collected as damaged samples and labeled;

[0046] A2: SVM classification model training: The collected damaged samples and normal samples are input into the support vector machine (SVM) classification model for training. The classification accuracy is improved by optimizing the model parameters to achieve automatic identification and alarm of damaged meters.

[0047] SVM classification model training includes:

[0048] Objective function: SVM aims to find an optimal hyperplane so that data points of different categories are separated by the maximum margin; the optimization goal is to minimize the following objective function:

[0049]

[0050] Where w is the normal vector of the hyperplane, C is the penalty parameter, ξ i is a slack variable used to handle the inseparable case.

[0051] Constraints:

[0052]

[0053] Among them, y i is the label of the sample (+1 or -1), and b is the bias term;

[0054] Kernel function: For nonlinearly separable cases, kernel functions can be used to map data to higher dimensions:

[0055] K(x i ,x j )=φ(x i )·φ(x j )

[0056] Final model: After training, the SVM model can be expressed as:

[0057]

[0058] Among them, α i is the Lagrange multiplier.

[0059] The present invention also provides a substation meter identification system, comprising:

[0060] An image acquisition module, used to collect image data of meters in the substation;

[0061] An image preprocessing module is used to perform preprocessing operations on the collected images;

[0062] Feature extraction and recognition module, including a deep learning model, for feature extraction and recognition of preprocessed images;

[0063] Reading recognition and correction module, used to calculate the reading of pointer meter and perform automatic correction;

[0064] Damage detection and early warning module, used to detect whether the meter is damaged and send out early warning signals when abnormalities are found;

[0065] Control system, used to coordinate the work of each module and process the recognition results;

[0066] A display device, used to display information such as recognition results, warning information and meter images;

[0067] The wireless communication module is used to transmit the recognition results, warning information, etc. to the remote monitoring center or inspection personnel in real time.

[0068] System integration and application:

[0069] System integration: Integrate the above algorithm modules into the intelligent inspection robot or monitoring system to form a complete substation meter automatic identification system. During the system integration process, it is necessary to ensure smooth data interaction between modules and stable operation of the system.

[0070] Application scenarios: The system can be widely used in various substation scenarios to realize the automatic identification and inspection functions of meters. Especially in harsh environments such as wind and sand, the system can significantly improve the inspection efficiency and identification accuracy, reduce labor costs and the probability of safety hazards.

[0071] Beneficial effects: Compared with the prior art, the present invention integrates multiple advanced technologies such as image sand removal, deep learning, image processing and pattern recognition to form a new meter recognition algorithm for meter recognition in a sandstorm environment. A large number of samples are collected through on-site collection or other channels, and trained through a deep learning algorithm. The position and type of the meter can be automatically located in the scene. Subsequently, the meter reading is automatically identified by adding automatic correction, feature extraction and other algorithms through the pattern recognition algorithm. The method of the present invention has strong adaptability to conditions such as distance, angle, and lighting, and the same algorithm is applicable to meters of the same type. There is no need to model each point separately, and it is suitable for multi-point scenarios. At the same time, it has higher accuracy and stronger generalization function, and is suitable for automatic recognition of multiple meters in substations, especially for meter recognition in sandstorm environments. The robustness of the algorithm is improved, and the problem of poor recognition effect of existing recognition methods in sandstorm environments is solved. It has broad application prospects and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0073] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0074] like Figure 1 As shown, the present invention provides a method for identifying substation meters in a sandstorm environment project, comprising the following steps:

[0075] S1: Collect substation meter images;

[0076] Substation meter images are collected through on-site inspection robots or fixed cameras. Inspection robots can carry high-definition cameras to automatically inspect and take meter images in substations; fixed cameras are installed at key locations in substations to monitor meters in real time.

[0077] S2: preprocessing the collected images;

[0078] Preprocessing includes removing wind and sand, reducing noise, and enhancing contrast;

[0079] In order to solve the image blur problem that may be caused by the sandstorm environment, a sandstorm removal algorithm based on dark channel prior is used to remove the sandstorm from the collected images. Specifically:

[0080] The algorithm removes fog from the image by calculating the dark channel value of the image, statistical histogram, ambient light value and transmittance, etc. The quality of the image after removing the wind and sand is significantly improved, providing a better recognition basis for subsequent steps.

[0081] The dark channel calculation module is used to calculate the dark channel value of each block of the current image, and the minimum value of the R, G, and B components of all pixels in each block is used as the dark channel value of the R, G, and B components of the block.

[0082] Specifically, the method for calculating the dark channel I_dark is to first divide an input original image into multiple blocks, each block size is n×m. In this embodiment, preferably n is 18 and m is 24. Assuming the height of the image is h and the width is w, the number of blocks should be (h×w) / (18×24). Based on the above blocks, the dark channel I_dark is calculated as follows:

[0083]

[0084] Among them, min means taking the minimum value, and Ω means the set of 18×24. The dark channel I_dark is actually to find the Raw value in the Ω domain. c The minimum value of Raw c Refers to the raw image data output by the image sensor containing the three components of R, G, and B.

[0085] Noise reduction and enhancement: Use median filtering, Gaussian filtering and other technologies to remove image noise, and use histogram equalization and other methods to enhance image contrast and make meter features more prominent.

[0086] S3: Use the deep learning model to extract features from the preprocessed images, automatically learn and extract key features from the meter images;

[0087] Feature extraction includes node feature extraction, edge feature extraction, and image feature extraction;

[0088] The extraction of key features specifically includes:

[0089] The key features of edge feature extraction are extracted through global domain overlap features. The global domain overlap feature refers to the number of paths with different path lengths between two nodes, and the calculation formula of the global domain overlap feature is:

[0090] S=(I-β*A)-1-I

[0091] Among them, A is the adjacency matrix, and the result S[i,j] represents the number of paths with different path lengths between the i node and the j node.

[0092] The key features of node feature extraction include degree, clustering coefficient and other features. Degree reflects the number of connections of a node. The formula for extracting it by counting the number of edges in the adjacency matrix is ​​Degree(v) = ∑Avu, where Avu indicates whether there is an edge between nodes and. The clustering coefficient measures the connectivity of node neighbors and is obtained by calculating the number of edges between a node and its neighbors. The calculation formula is:

[0093]

[0094] Where E is the actual number of edges and k is the node degree.

[0095] The key features of image feature extraction include color, texture, shape, local and deep learning features. Color features represent pixel distribution through color histogram and calculate the frequency of each color. Texture features are extracted through local binary pattern (LBP) or gray level co-occurrence matrix (GLCM). The LBP formula is:

[0096]

[0097] Where s is the threshold function. Shape features use edge detection algorithms to extract object contours. Local features such as SIFT detect key points and their feature vectors in the image. Deep learning features automatically extract high-level features through convolutional neural networks and forward propagate to generate feature maps.

[0098] S4: Detect and locate the substation meters in the image based on the extracted key features;

[0099] The whole process includes multiple steps from image preprocessing, feature extraction, to classifier detection and post-processing. Image preprocessing involves convolution of the image through a Gaussian blur kernel to smooth the image. The specific formula is:

[0100]

[0101] Where G(x,y) represents the Gaussian value at (x,y)(x,y)(x,y), σ is the standard deviation of the Gaussian kernel, which controls the degree of blur. x,y are the coordinates of the relative pixel in the image. Feature extraction calculates the gradient of each pixel in the image and divides the gradient direction into multiple intervals to construct a histogram.

[0102] Deep learning model training: Use YOLOv5s or similar object detection networks to train images after sand removal. Label the meter locations and types in a large number of sample images as training data input for model training. During the training process, continuously optimize the model parameters to improve the accuracy of positioning and classification.

[0103] Meter location and classification: The trained deep learning model can automatically locate and classify substation meters in the input image. The model extracts image features and matches them with pre-trained weights to achieve efficient and accurate location and classification output.

[0104] S5: Using a pattern recognition algorithm to perform pattern recognition processing on the located meter image, identify the meter reading, and integrate an automatic correction algorithm to improve the reading accuracy;

[0105] The pattern recognition processing method is as follows: the meter image is separated from the background through image segmentation technology; then the contour features of the meter pointer or numbers are identified using contour extraction technology; finally, the meter reading is calculated based on prior knowledge (such as meter range, pointer starting position, etc.).

[0106] The purpose of image segmentation is to separate the meter from the background. Common image segmentation techniques include threshold processing, edge detection, and region growing. First, convert the color image into a grayscale image to reduce the complexity of the data. Then apply Gaussian blur to reduce noise, smooth the image, and enhance the contrast between the meter and the background. And binarize the image through threshold processing. You can use the Otsu threshold algorithm or a custom threshold. Next, you need to extract the contour features of the meter pointer or number, which can be achieved through Canny edge detection and findContours function. Use the Canny edge detection algorithm to find the edges in the image. Use OpenCV's findContours function to detect contours in the image. Here we only focus on contours with larger areas (contours of pointers and numbers) and ignore small contours such as noise. Filter out contours that match the meter pointer or number features, usually based on features such as the shape, area, or aspect ratio of the contour. Use prior knowledge of the meter (such as range, pointer starting position) to calculate the final reading. The pointer is usually long and straight, and the Hough Transform can be used to detect straight lines. The angle of the meter pointer is calculated by geometric methods, and the reading is calculated in combination with the meter range. The angle of the pointer can be calculated by the slope formula: angle = arctan (y2-y1 / x2-x1), and then mapped to the meter range according to the angle. Finally, the detected pointer and the calculated reading are displayed on the image.

[0107] The correction process of the automatic correction algorithm is as follows: during the recognition process, the automatic correction algorithm is added to adjust the distance, angle and lighting parameters of the image to reduce the impact of environmental factors on the recognition results; at the same time, the feature extraction algorithm is added to further extract node features, edge features, image features, etc. to enhance the robustness of recognition.

[0108] To implement the automatic correction algorithm, the following steps can be followed:

[0109] Image preprocessing: Use Gaussian blur or median filtering to remove noise. Resize the image to a uniform input size.

[0110] Distance Correction: Estimate distances through deep learning or traditional computer vision techniques such as stereo vision. Apply perspective transformations and adjust images to remove distance effects.

[0111] Angle correction: Detect straight lines through Hough transform and estimate the image rotation angle. Use the rotation matrix to perform image rotation correction.

[0112] Lighting Correction: Use histogram equalization to enhance contrast. Apply gamma correction to adjust brightness.

[0113] Feature extraction and enhancement: Use edge detection (such as the Canny algorithm) to extract edge point features. Use SIFT or ORB to extract key point features. Use machine learning algorithms (such as random forests or support vector machines) to train the model to further improve the robustness of the features.

[0114] Recognition and post-processing: The extracted features are input into the classifier for target recognition. Based on the recognition results, a threshold or voting mechanism is applied to improve recognition accuracy.

[0115] S6: Analyze the meter image through the designed damage detection algorithm, and issue a warning signal immediately once damage is detected; this step, as a pre-processing link of the reading recognition algorithm, effectively reduces the occurrence of false readings and improves the reliability of the overall system;

[0116] The damage detection algorithm is as follows:

[0117] A1: Damage sample collection and labeling: During the inspection process, images of damaged meters detected are collected as damaged samples and labeled;

[0118] A2: SVM classification model training: The collected damaged samples and normal samples are input into the support vector machine (SVM) classification model for training. The classification accuracy is improved by optimizing the model parameters to achieve automatic identification and alarm of damaged meters.

[0119] The basic algorithm expression of Support Vector Machine (SVM) can be summarized by the following steps and formulas:

[0120] Objective function: SVM aims to find an optimal hyperplane so that data points of different categories are separated by the maximum margin; the optimization goal is to minimize the following objective function:

[0121]

[0122] Where w is the normal vector of the hyperplane, C is the penalty parameter, ξ iis a slack variable used to handle the inseparable case.

[0123] Constraints:

[0124]

[0125] Among them, y i is the label of the sample (+1 or -1), and b is the bias term;

[0126] Kernel function: For nonlinearly separable cases, kernel functions can be used to map data to higher dimensions:

[0127] K(x i ,x j )=φ(x i )·φ(x j )

[0128] Final model: After training, the SVM model can be expressed as:

[0129]

[0130] Among them, α i is the Lagrange multiplier.

[0131] By optimizing these parameters, you can improve classification accuracy and achieve automatic identification and alarm of broken meters.

[0132] In order to implement the above method, the present invention also provides a substation meter identification system, comprising:

[0133] An image acquisition module, used to collect image data of meters in the substation;

[0134] An image preprocessing module is used to perform preprocessing operations on the collected images;

[0135] Feature extraction and recognition module, including a deep learning model, for feature extraction and recognition of preprocessed images;

[0136] Reading recognition and correction module, used to calculate the reading of pointer meter and perform automatic correction;

[0137] Damage detection and early warning module, used to detect whether the meter is damaged and send out early warning signals when abnormalities are found;

[0138] Control system, used to coordinate the work of each module and process the recognition results;

[0139] A display device, used to display information such as recognition results, warning information and meter images;

[0140] The wireless communication module is used to transmit the recognition results, warning information, etc. to the remote monitoring center or inspection personnel in real time.

[0141] Integrate and apply the above substation meter identification system:

[0142] System integration: Integrate the above algorithm modules into the intelligent inspection robot or monitoring system to form a complete substation meter automatic identification system; during the system integration process, it is necessary to ensure smooth data interaction between modules and stable operation of the system.

[0143] Application scenarios: The system can be widely used in various substation scenarios to realize the automatic identification and inspection functions of meters. Especially in harsh environments such as wind and sand, the system can significantly improve the inspection efficiency and identification accuracy, reduce labor costs and the probability of safety hazards.

[0144] In order to verify the effectiveness and effect of the present invention, in this embodiment, the above-mentioned method of the present invention is compared with the traditional identification method in a sandstorm environment, as follows:

[0145] Experimental environment: In a moderate sandstorm environment with a wind speed of 10m / s, a set of substation meter image data in a simulated sandstorm environment was collected. The data set includes images that are partially blurred and obscured by sandstorms. This environment will interfere with the text and numbers on the meter, and traditional recognition methods have obvious challenges in these environments.

[0146] Comparison algorithms:

[0147] The method of the present invention is based on an improved image enhancement and noise filtering technology, firstly preprocessing the image to reduce the interference caused by wind and sand, and then performing multiple iterative recognition through a deep learning network.

[0148] Traditional methods: Use traditional edge detection-based algorithms to try to extract meter information directly from images.

[0149] Recognition accuracy:

[0150] The method of the present invention has an accuracy rate of successfully identifying meter information of 92%, that is, among the 100 images processed, 92 images can correctly identify meter data.

[0151] Traditional method: The recognition accuracy rate is 78%. In the case of severe wind and sand obstruction, some images cannot be recognized at all.

[0152] Identify error conditions:

[0153] The method of the present invention: In rare cases, the identified meter information will be slightly affected by the residual wind and sand, but in most cases, the meter numbers can be clearly identified.

[0154] Traditional method: Many images cannot be recognized due to the interference of wind and sand on edge features. Especially when the wind and sand are concentrated in the meter area, the recognition results are obviously biased.

[0155] Processing speed:

[0156] The method of the present invention takes an average of 120 milliseconds to process an image. Due to the use of a parallel processing architecture, the overall processing speed is relatively fast.

[0157] Traditional method: The average processing time is 180 milliseconds, which is relatively slow due to the high algorithm complexity and lack of preprocessing steps.

[0158] This case shows through comparison that in a moderate wind and sand environment, the method of the present invention is significantly superior to the traditional method in terms of recognition accuracy and processing speed. Especially in images with greater wind and sand interference, the recognition effect of the traditional method is poor, while the method of the present invention can adapt well to complex environments.

Claims

1. A method for identifying substation meters in a sandstorm environment project, characterized in that: The steps include: S1: Collect substation meter images; S2: preprocessing the collected images; S3: Use the deep learning model to extract features from the preprocessed images, automatically learn and extract key features from the meter images; S4: Detect and locate the substation meters in the image based on the extracted key features; S5: Using a pattern recognition algorithm to perform pattern recognition processing on the located meter image, identify the meter reading, and integrate an automatic correction algorithm to improve the reading accuracy; S6: The meter image is analyzed by a designed damage detection algorithm, and a warning signal is immediately issued once damage is detected.

2. The method for identifying substation meters in a sandstorm environment project according to claim 1, characterized in that: In step S1, the substation meter images are collected by an on-site inspection robot or a fixed camera.

3. The method for identifying substation meters in a sandstorm environment project according to claim 1, characterized in that: The pre-processing in step S2 includes removing wind and sand, reducing noise, and enhancing contrast.

4. The method for identifying substation meters in a sandstorm environment project according to claim 3 is characterized in that: In step S2, a sand removal algorithm based on dark channel prior is used to remove sand from the collected image, specifically: The minimum value of all pixels in each block in the R, G, and B components is taken as the dark channel value of the R, G, and B components of the block; The method for calculating the dark channel I_dark is to first divide the input original image into multiple blocks, each block size is n×m, and the image height is h and the width is w, then the number of blocks should be (h×w) / (n×m); based on the above blocks, the dark channel I_dark is calculated as follows: Among them, min means taking the minimum value, Ω means the set of n×m; the dark channel I_dark is actually to find the Raw c Minimum value of Raw c Refers to the raw image data output by the image sensor containing the three components of R, G, and B.

5. The method for identifying substation meters in a sandstorm environment project according to claim 1, characterized in that: The feature extraction in step S3 includes node feature extraction, edge feature extraction, and image feature extraction.

6. The method for identifying substation meters in a sandstorm environment project according to claim 5, characterized in that: The extraction of key features in step S3 specifically includes: The key features of edge feature extraction are extracted through global domain overlap features. The global domain overlap feature refers to the number of paths with different path lengths between two nodes, and the calculation formula of the global domain overlap feature is: S=(I-β*A)-1-I Where A is the adjacency matrix, and the result S[i,j] represents the number of paths with different path lengths between nodes i and j; The key features of node feature extraction include degree and clustering coefficient. Degree reflects the number of connections of a node. The formula for extracting it by counting the number of edges in the adjacency matrix is ​​Degree(v)=∑Avu, where Avu indicates whether there is an edge between nodes and. The clustering coefficient measures the connectivity of node neighbors and is obtained by calculating the number of edges between a node and its neighbors. The calculation formula is: Where E is the actual number of edges and k is the node degree; The key features of image feature extraction include color, texture, shape, local and deep learning features; color features represent pixel distribution through color histogram and calculate the frequency of each color; texture features are extracted through local binary pattern LBP or gray level co-occurrence matrix GLCM. The LBP formula is: Where s is the threshold function; the shape feature uses the edge detection algorithm to extract the object contour; the deep learning feature automatically extracts high-level features through the convolutional neural network, and forward propagates to generate feature maps.

7. The method for identifying substation meters in a sandstorm environment project according to claim 1, characterized in that: The pattern recognition processing in step S5 is as follows: the meter image is separated from the background by image segmentation technology; then the contour features of the meter pointer or number are identified by contour extraction technology; and finally the meter indication is calculated in combination with prior knowledge; Specifically: first convert the color image into a grayscale image, then apply Gaussian blur to reduce noise, smooth the image, and enhance the contrast between the meter and the background; and binarize the image through threshold processing; next, extract the contour features of the meter pointer or numbers, and implement it through Canny edge detection and findContours function. The Canny edge detection algorithm is used to find the edges in the image, and the findContours function of OpenCV is used to detect the contours in the image; screen out the contours that match the meter pointer or digital features; use the prior knowledge of the meter to calculate the final reading, calculate the angle of the meter pointer through geometric methods, and calculate the reading in combination with the meter range. The angle of the pointer is calculated using the slope formula: angle = arctan (y2-y1 / x2-x1), and then mapped to the meter range according to the angle, and finally display the detected pointer and calculated reading on the image.

8. The method for identifying substation meters in a sandstorm environment project according to claim 1, characterized in that: The correction process of the automatic correction algorithm in step S5 is: adding the automatic correction algorithm to adjust the distance, angle and illumination parameters of the image during the recognition process to reduce the influence of environmental factors on the recognition result; and adding the feature extraction algorithm to further extract node features, edge features, image features, etc. to enhance the robustness of recognition; The correction process includes image preprocessing, distance correction, angle correction, illumination correction, feature extraction and enhancement, recognition and post-processing, specifically: Image preprocessing: Use Gaussian blur or median filtering to remove noise; resize the image to unify the input size; Distance correction: estimate distance through deep learning or traditional computer vision techniques; apply perspective transformation and adjust the image to eliminate the effect of distance; Angle correction: Detect straight lines through Hough transform and estimate the image rotation angle; use the rotation matrix to perform image rotation correction. Lighting correction: Use histogram equalization to enhance contrast; apply gamma correction to adjust brightness; Feature extraction and enhancement: Use edge detection to extract edge point features; use SIFT or ORB to extract key point features; Use machine learning algorithms to train the model to further improve the robustness of features. Recognition and post-processing: The extracted features are input into the classifier for target recognition. Based on the recognition results, a threshold or voting mechanism is applied to improve the recognition accuracy.

9. The method for identifying substation meters in a sandstorm environment project according to claim 1, characterized in that: The damage detection algorithm in step S6 is specifically: A1: Damage sample collection and labeling: During the inspection process, images of damaged meters detected are collected as damaged samples and labeled; A2: SVM classification model training: The collected damaged samples and normal samples are input into the support vector machine (SVM) classification model for training. The classification accuracy is improved by optimizing the model parameters to achieve automatic identification and alarm of damaged meters. SVM classification model training includes: Objective function: SVM aims to find an optimal hyperplane so that data points of different categories are separated by the maximum margin; the optimization goal is to minimize the following objective function: Where w is the normal vector of the hyperplane, C is the penalty parameter, ξ i is a slack variable used to handle the inseparable case. Constraints: Among them, y i is the label of the sample (+1 or -1), and b is the bias term; Kernel function: For nonlinearly separable cases, kernel functions can be used to map data to higher dimensions: K(x i ,x j )=φ(x i )·φ(x j ) Final model: After training, the SVM model can be expressed as: Among them, α i is the Lagrange multiplier.

10. The substation meter identification system according to the method of claim 1, characterized in that: include: An image acquisition module, used to collect image data of meters in the substation; An image preprocessing module is used to perform preprocessing operations on the collected images; Feature extraction and recognition module, including a deep learning model, for feature extraction and recognition of preprocessed images; Reading recognition and correction module, used to calculate the reading of pointer meter and perform automatic correction; Damage detection and early warning module, used to detect whether the meter is damaged and send out early warning signals when abnormalities are found; Control system, used to coordinate the work of each module and process the recognition results; A display device, used to display recognition results, warning information and meter images; The wireless communication module is used to transmit the recognition results and warning information to the remote monitoring center or inspection personnel in real time.

Images