Intelligent identification system for assembling and connecting specifications of electric energy meter

Abstract

The invention discloses an electric energy meter assembling standard intelligent identification system, which comprises an image acquisition module, an image preprocessing module, a wiring identification module and a standard comparison module, and is characterized in that the wiring identification module carries out NLM denoising and Gamma correction enhancement processing on a to-be-identified image, dynamically calculates a threshold value by using an Otsus method, accurately detects the position of a sleeve in an electric meter image, and carries out standard comparison on the to-be-identified image; and acquiring the coordinate of each casing, and identifying the casing fields by using PaddleOCR (Passive Optical Character Recognition). According to the method, through specific tests, wiring diagram shooting, threshold value dynamic calculation through an Otsus method, contrast enhancement through Gamma correction, NLM denoising, text detection, character recognition, geometric correction and other modes, OCR recognition is carried out on the image, algorithm and formula optimization of different degrees is carried out in the recognition process, the detection precision and efficiency are remarkably improved, and the method is suitable for popularization and application. And an effective solution is provided for automatic detection of the electric meter.

Description

Technical Field

[0001] The present invention relates to the technical field of an electric energy meter installation and identification system, and in particular to an electric energy meter installation specification intelligent identification system. Background Art

[0002] In recent years, the demand for electricity has continued to increase. As electricity consumption continues to expand, the requirements placed on power companies have also increased, especially during meter installation and connection, which impacts the power supply environment and quality. Furthermore, the development of intelligent and digital power systems has placed higher demands on the wiring process and standard inspection of electricity meters. Traditional manual inspection methods are no longer able to meet the efficiency, accuracy, and reliability requirements of modern power systems.

[0003] Current electricity meter wiring process inspections on the market suffer from several drawbacks: 1) Limited scope: Slow inspection speeds make it difficult to meet the needs of large-scale electricity meter inspections, especially during centralized installation or periodic maintenance of power equipment, where efficiency bottlenecks are particularly prominent. Furthermore, this limited scope of inspection fails to simultaneously verify wiring compliance, correctness, and workmanship. 2) Inaccuracy: Traditional image recognition is significantly affected by factors such as lighting and space, resulting in errors in recognition results. 3) High cost and cumbersome procedures: Due to the current limited scope of inspection, additional inspection modules are required to meet additional inspection requirements, resulting in high costs and a cumbersome operational process. Summary of the Invention

[0004] In response to the above problems, the present invention proposes an intelligent recognition system for electric energy meter installation specifications, which solves the defect that the existing traditional manual detection method can no longer meet the requirements of modern power systems for efficiency, accuracy and reliability.

[0005] The technical solution adopted by the present invention is as follows: An intelligent recognition system for electric energy meter installation specifications includes: an image acquisition module for acquiring images of electric energy meter connection terminals; The image preprocessing module is used to preprocess the collected images, and the feature extraction module uses the image recognition algorithm to extract the feature information of the electricity meter terminals and wires; The wiring identification module is used to identify the type and number of the electricity meter terminal blocks and the connection relationship of the wires based on the extracted feature information. The wiring identification module includes a color recognition unit and a wiring accuracy recognition unit. The color recognition unit is used to identify the wire color through color space conversion and threshold segmentation algorithm and compare it with the preset color. The wiring accuracy recognition unit is used to determine whether the wiring of each color is correct and standardized based on the identified connection relationship between the terminal blocks and the wires. The wiring recognition module performs NLM denoising and gamma correction enhancement on the image to be identified, uses Otsu's method to dynamically calculate the threshold, detects the location of the casing in the meter image, obtains the coordinates of each casing, and uses PaddleOCR to identify the casing field; and The specification comparison module is used to compare the identified wiring information with the preset wiring process and specifications. The wiring information includes horizontality, verticality, inclination, and curvature, and determine whether the wiring containing the above wiring information is correct and standardized.

[0006] Through specific experiments, this invention uses Otsu's method to dynamically calculate thresholds using a wiring diagram, gamma correction for contrast enhancement, NLM denoising, text detection, character recognition, and geometric correction to perform optical character recognition (OCR) on the image. The algorithm and formula optimization are performed to varying degrees during the recognition process. Experimental results demonstrate that the invention significantly improves detection accuracy and efficiency, achieving higher reliability, when applied to meter tilt detection, meter wiring detection, and casing field recognition. This provides an effective solution for automated meter detection.

[0007] Optionally, the preprocessing operations include image denoising, grayscale conversion, contrast enhancement, and binarization.

[0008] Optionally, the image denoising operation includes searching in the image block using an NLM algorithm, calculating the Euclidean distance between the sliding window and the specified window to determine the degree of similarity between them, thereby determining a weighted average value and performing a filtering operation; the specific NLM algorithm process adopts the following formula: ;

[0009] Wherein, NL[v](i): the neighborhood aggregation result of node v in the i-th layer (summary of neighbor information), w(i,j) represents the weighted average kernel value, i: the number of layers of the neural network (current layer), v(j): node v is on the target node; j∈i: the position index of neighbor pixel j belongs to the search window i, and the kernel value is determined by the similarity between two blocks, and the similarity between two blocks is determined by calculating the Euclidean distance between them.

[0010] Optionally, the binarization operation is to find an optimal threshold by calculating the inter-class variance, so that the pixels in the image can be divided into two classes, and the variance between the two classes is maximized.

[0011] Optionally, the contrast enhancement operation is specifically to enhance the contrast using Gamma correction, using the following formula: Ienhanced(x,y)=I(x,y)γ; Where ‌x‌ represents the horizontal position of the pixel in the image, ‌y‌ represents the vertical position of the pixel in the image, γ<1 enhances dark details, and γ>1 enhances bright details‌‌.

[0012] Optionally, the extracted feature information includes color, shape, and position.

[0013] Optionally, the electric energy meter installation specification intelligent identification system further includes a result output module, which is used to output detection results, including whether the wiring is correct, error type, location information, and generate a detection report.

[0014] The wiring recognition module also includes a text detection unit, which adopts the EAST model. The EAST model directly predicts the text area in the image through a fully convolutional network and merges the final detection results through non-maximum suppression.

[0015] Optionally, the wiring recognition module further includes a character recognition unit, wherein the character recognition unit adopts a CRNN model, wherein the CRNN model extracts local features of image data by using a convolutional layer, then reduces the spatial dimension of the features by using a pooling layer, and finally performs classification or regression tasks by using a fully connected layer; The CRNN model extracts features from the image through CNN and processes these feature sequences through RNN. It performs time series modeling based on these feature sequences and introduces an attention mechanism into the sequence modeling. The specific attention mechanism adopts the following formula:

[0016] Where Q is the query matrix, K is the key matrix, V is the value matrix, and d k It is a dimension; Alternatively, the wiring recognition module also includes using an N-gram language model or BERT to correct the recognition results and using regular expressions to match fixed formats.

[0017] Optionally, a loss function is used during the binarization process to evaluate the difference between the binarized output of the model and the true label. The loss function can be used to measure the degree of model fit. A loss value can be obtained by calculating the difference between the output predicted by the model and the actual label, and then minimizing this loss value. The formula used is:

[0018] is the function obtained after training, w and b are the two independent variables of the loss function, : is the true label (target value) of the i-th sample, and the two parameters obtained by function training.

[0019] In summary, the present invention has the following beneficial effects: 1. By combining Otsu's and PaddleOCR, this solution successfully addresses the field mismatch problem caused by inaccurate casing positioning in traditional OCR recognition, the problem of individual character recognition failures due to insufficient lighting and image distortion, and the problem of recognition failures caused by jagged text edges. In tests, this solution accurately identified the casing position and characters in over 96% of electricity meter images.

[0020] 2. The present invention adopts the EAST model, and its end-to-end detection method greatly improves the efficiency and accuracy of detection.

[0021] 3. The present invention adds an attention mechanism to CRNN, which greatly improves the efficiency and accuracy of its deep learning, so that it can recognize images through self-learning even when the image feature points are not obvious or unbalanced.

[0022] 4. The present invention uses a combination of multiple detection methods to judge the wiring standardization from various angles, improve the accuracy of standard judgment, and enhance the wiring level of employees. Through a large amount of data experimental analysis, the system's recognition accuracy reaches about 95-98%.

[0023] 5. The present invention adopts a combination of multiple detection methods to reduce the complexity of identification operations and reduce usage costs.

[0024] 6. After the detection starts, the present invention automatically detects with one click and automatically generates a report of the detection results, which is convenient for data storage, query and analysis.

[0025] 7. In the present invention, a wiring diagram to be identified, including the identification of wiring standardization, wiring accuracy, position standardization, etc., can give a conclusion in just 6-8 seconds, and the recognition success rate basically remains above 98%, which greatly reduces the pressure and workload of teachers' subjective judgment, reduces time and improves efficiency. DETAILED DESCRIPTION

[0026] The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0027] Example 1 The technical solution adopted by the present invention is as follows:

[0028] The present invention discloses an intelligent recognition system for electric energy meter installation specifications, comprising: an image acquisition module for acquiring images of electric energy meter terminals; an image preprocessing module for preprocessing the acquired images; a feature extraction module for extracting feature information of electric energy meter terminals and wires using an image recognition algorithm; a wiring recognition module for identifying the type and number of electric energy meter terminals, as well as the connection relationship of the wires based on the extracted feature information; the wiring recognition module comprises a color recognition unit and a wiring accuracy recognition unit, the color recognition unit being used to identify the wire color through color space conversion and threshold segmentation algorithm, and comparing it with a preset color, the wiring accuracy recognition unit being used to judge whether the wiring is correct and standardized based on the identified connection relationship between the terminal and the wire; and a specification comparison module being used to compare the identified wiring information with preset wiring processes and specifications, the wiring information including horizontality, verticality, inclination, and curvature, to judge whether the wiring containing the above wiring information is correct and standardized.

[0029] The preprocessing operations include image denoising, grayscale conversion, contrast enhancement, and binarization.

[0030] The image denoising operation involves searching within an image block using an NLM algorithm. The algorithm calculates the Euclidean distance between a sliding window and a specified window. The Euclidean distance is calculated by taking the square root of the sum of the squared coordinate differences between two points in each dimension. This determines the degree of similarity between the two points and then determines a weighted average value for filtering. The NLM algorithm utilizes all the information in the image to perform a weighted average of the similarities between the pixels.

[0031] The specific NLM algorithm process uses the following formula:

[0032] Wherein, NL[v](i): the neighborhood aggregation result of node v in the i-th layer (summary of neighbor information), w(i,j) represents the weighted average kernel value, i: the number of layers of the neural network (current layer), v(j): node v is on the target node; j∈i: the position index of neighbor pixel j belongs to the search window i, and the kernel value is determined by the similarity between two blocks, and the similarity between two blocks is determined by calculating the Euclidean distance between them.

[0033] NLM searches in the image block and calculates the Euclidean distance between the sliding window and the specified window to determine the similarity between them, thereby determining the value of the weighted average and performing the filtering operation.

[0034] This value is determined by the similarity between two blocks, and the similarity between two blocks is determined by calculating the Euclidean distance between them. However, in an image, for each pixel, the closer the pixel is to it, the greater the similarity is to itself, and the farther the distance is, the smaller the similarity is.

[0035] The binarization operation involves finding the optimal threshold by calculating the inter-class variance, so that the pixels in the image can be divided into two classes with the maximum variance between the two classes.

[0036] The wiring recognition module performs NLM denoising and gamma correction enhancement on the image to be identified, uses Otsu's method to dynamically calculate the threshold, accurately detects the location of the casing in the meter image, obtains the coordinates of each casing, uses PaddleOCR to identify the casing field, and optimizes its algorithms in text detection, character recognition, and image geometric correction to ensure that every word can be correctly identified. Geometric correction uses a series of mathematical models to balance the number of control points. The formula is: number of control points = (m+1)*(m+2) / 2, where m is the degree of the polynomial model.

[0037] In this embodiment, the specific operations of optimizing the image preprocessing formula are: Use Otsu's method to dynamically calculate the threshold instead of a fixed threshold: Otsu's is a classic algorithm for image binarization. Its core idea is to find the optimal threshold by calculating the inter-class variance, so that the pixels in the image can be divided into two categories with the maximum variance between the two categories.

[0038] Assume that the grayscale range of the image is [0, L-1], L: 8-bit image, corresponding to L=256, and the number of pixels with grayscale value i is n i , the total number of pixels is N=n0+n1+...+n L-1 , the probability of each gray value appearing is P i =n i / N. Assuming the threshold is t, the image is divided into two categories C0 (grayscale value range is [0,t]) and C1 (grayscale value range is [t+1,L-1]). The probability of C0 is w0=P0+P1+...+P t , the average gray value is u0=(P0*0+P1*1+...+P t *t) / w0. The probability of C1 is w1=P t+1 +P t+2 +...+P L-1 , the average gray value is u1=(P t+1 *(t+1)+P t+2 *(t+2)+...+P L-1*(L-1)) / w1. The overall average grayscale value of the image is u = w0*u0 + w1*u1. u0: The average grayscale value of the foreground (target); u1: The average grayscale value of the background. The inter-class variance g = w0*(u0-u)² + w1*(u1-u)². Otsu's algorithm aims to find the value t that maximizes g; this value t is the optimal threshold. Here, w0 represents the cumulative probability distribution of foreground pixels. w1 represents the cumulative probability distribution of background pixels.

[0039] The contrast enhancement operation is: Gamma correction is used to enhance contrast. The specific implementation formula is: Ienhanced(x,y)=I(x,y)γ in, x (horizontal axis): Indicates the horizontal position of the pixel in the image (column index). The value range is usually 0 ≤ x < image width.

[0040] y (Y-axis): Indicates the vertical position (row index) of a pixel in the image. The value range is typically 0 ≤ y < image height. Here, I(x,y) refers to the raw brightness value (grayscale image) or color channel value (color image) of the pixel at column x and row y in the original input image.

[0041] γ<1 enhances dark details, and γ>1 enhances bright details‌‌.

[0042] The electric energy meter installation specification intelligent identification system also includes a result output module, which is used to output the detection results, including whether the wiring is correct, the error type, and location information, and generate a detection report.

[0043] The wiring recognition module also includes a text detection unit, which adopts the EAST model. The EAST model directly predicts the text area in the image through a fully convolutional network and merges the final detection results through non-maximum suppression.

[0044] In this example, EAST (Efficient and Accurate Scene Text Detector) is a text detection model based on deep learning. Unlike traditional text detection methods, EAST eliminates complex intermediate steps such as candidate bounding box extraction and text formatting. Instead, it directly predicts text regions in images using a fully convolutional network (FCN) and combines the final detection results using non-maximum suppression (NMS).

[0045] The wiring recognition module also includes a character recognition unit, which uses a CRNN model. The CRNN model extracts local features of image data by using a convolutional layer, then reduces the spatial dimension of the features by a pooling layer, and finally performs classification or regression tasks by a fully connected layer. In this embodiment, the Convolutional Recurrent Neural Network (CRNN) is a deep learning model that combines a convolutional neural network (CNN) and a recurrent neural network (RNN). Its core is to use RNNs to perform temporal modeling of feature sequences extracted by CNNs, thereby solving image sequence recognition tasks (such as text recognition). CNNs use convolutional layers to extract local features from image data, then use pooling layers to reduce the spatial dimensionality of these features, and finally use fully connected layers to perform classification or regression tasks.

[0046] The CRNN model extracts features from the image through CNN and processes these feature sequences through RNN. It performs time series modeling based on these feature sequences and introduces an attention mechanism into the sequence modeling. The specific attention mechanism adopts the following formula:

[0047] Where Q is the query matrix, K is the key matrix, V is the value matrix, and d k The softmax function is used to calculate the similarity weight between the query Q and the key K, and then applied to the value V. : Calculate the similarity matrix between Query and Key.

[0048] The basic structure of the CRNN model is as follows: 1) Convolutional Layers: Convolutional layers are used to extract spatial features from images.

[0049] 2) Feature Map Serialization: The feature map output by the convolutional layer is converted into a sequence, which ensures that the position information of each feature map is compressed into a one-dimensional sequence.

[0050] 3) Recurrent Layers: These layers feed the sequenced feature maps into RNNs, such as LSTM (Long Short-Term Memory) or GRU (Gated Recurrent Unit). These RNN layers can capture temporal information in the sequence, such as the order of characters or words.

[0051] 4) Output Layer: The output of the RNN passes through a fully connected layer (FC) for classification or regression. For example, in an OCR task, the output layer might output a probability distribution for each character to identify text in an image.

[0052] The wiring recognition module also includes correcting the recognition results using the N-gram language model or BERT, and matching fixed formats using regular expressions.

[0053] In this embodiment, an N-gram language model or BERT is used to correct the recognition results, and a regular expression is used to match fixed formats (cased numbers), such as Ua, Ub, Uc... The loss function is used in the binarization process to evaluate the difference between the binarized result of the model output and the true label. The loss function can be used to measure the degree of model fit. By calculating the difference between the output predicted by the model and the actual label, a loss value can be obtained, and then this loss value is minimized. The formula used is:

[0054] is the function obtained after training, w and b are the two independent variables of the loss function, : The true label (target value) of the i-th sample, and two parameters obtained from function training. Momentum is used for optimization. This method introduces a momentum term based on SGD, accelerating convergence by accumulating previous gradients. This can speed up convergence and help escape local optima, similar to inertia.

[0055] In this embodiment, geometric correction refers to balancing the number of control points through a series of mathematical models to correct and eliminate geometric distortion caused by various factors during remote sensing image formation. The number of control points is primarily related to the degree of the correction polynomial, but also to the correction range and accuracy. The formula is: Number of Control Points = (m+1)*(m+2) / 2, where m is the degree of the polynomial model.

[0056] For example, the square requires 6 control points and the cube requires 10 control points. However, in the actual process of shooting wiring diagrams, due to the existence of slight concave and convex surfaces on the photographic material panel, the camera uses a wide-angle lens, which leads to certain distortion in the photos. If the traditional geometric correction formula is used, the effect is not very ideal. It is necessary to perform a weighted average algorithm on this basis. After multiple weightings, an ideal multi-power is obtained, and then the ideal number of control points in the current state is calculated to improve the accuracy and efficiency of the correction.

[0057] In the specific implementation of this embodiment, 1) an image of the meter connection area including the bushing field is captured; 2) Perform NLM denoising and gamma correction enhancement on the image to be identified to improve the subsequent recognition success rate; 3) Use Otsu's method to dynamically calculate the threshold, accurately detect the location of the casing in the meter image, and obtain the coordinates of each casing; 4) Use PaddleOCR to identify the casing fields and optimize its algorithms for text detection, character recognition, and image geometry correction to ensure that every word can be correctly recognized; The above description is only a preferred embodiment of the present invention and does not limit the scope of patent protection of the present invention. Any equivalent structural transformation made by using the description of the present invention and directly or indirectly applied to other related technical fields are also included in the scope of protection of the present invention.

Claims

1. An intelligent identification system for electric energy meter installation specifications, characterized in that: include: An image acquisition module, used for acquiring images of the electric energy meter terminal blocks; The image preprocessing module is used to preprocess the collected images, and the feature extraction module uses the image recognition algorithm to extract the feature information of the electricity meter terminals and wires; The wiring recognition module is used to identify the type and number of the electricity meter terminal blocks, as well as the connection relationship of the wires, based on the extracted feature information. The wiring recognition module includes a color recognition unit and a wiring accuracy recognition unit. The color recognition unit is used to identify the wire color through color space conversion and threshold segmentation algorithm and compare it with the preset color. The wiring accuracy recognition unit is used to determine whether the wiring of each color is correct and standardized based on the identified connection relationship between the terminal blocks and the wires. The wiring recognition module performs NLM denoising and gamma correction enhancement on the image to be identified, uses Otsu's method to dynamically calculate the threshold, detects the location of the casing in the meter image, obtains the coordinates of each casing, and uses PaddleOCR to identify the casing field. The specification comparison module is used to compare the identified wiring information with the preset wiring process and specifications. The wiring information includes horizontality, verticality, inclination, and curvature, and determine whether the wiring containing the above wiring information is correct and standardized.

2. The intelligent identification system for electric energy meter installation specifications according to claim 1, characterized in that: The preprocessing operations include image denoising, grayscale conversion, contrast enhancement, and binarization.

3. The intelligent identification system for electric energy meter installation specifications according to claim 2, characterized in that: The image denoising operation includes searching in the image block using the NLM algorithm, calculating the Euclidean distance between the sliding window and the specified window to determine the degree of similarity between them, and then determining the weighted average value for filtering operation; the specific NLM algorithm process uses the following formula: ; Among them, NL[v](i): the neighborhood aggregation result of node v in the i-th layer, w(i,j) represents the weighted average kernel value, i: the number of layers of the neural network, v(j): node v is on the target node; j∈i: the position index of neighbor pixel j belongs to the search window i, the kernel value is determined by the similarity between the two blocks, and the similarity between the two blocks is determined by calculating the Euclidean distance between them.

4. The intelligent identification system for electric energy meter installation specifications according to claim 2, characterized in that: The binarization operation is to find the optimal threshold by calculating the inter-class variance so that the pixels in the image can be divided into two categories and the variance between the two categories is maximized.

5. The intelligent identification system for electric energy meter installation specifications according to claim 2, characterized in that: The contrast enhancement operation is to use Gamma correction to enhance the contrast, using the following formula: Ienhanced(x,y)=I(x,y)γ; Where ‌x‌ represents the horizontal position of the pixel in the image, ‌y‌ represents the vertical position of the pixel in the image, γ<1 enhances dark details, and γ>1 enhances bright details.

6. An intelligent identification system for electric energy meter installation specifications according to claim 1, 2, 3, 4 or 5, characterized in that: The extracted feature information includes color, shape, and position.

7. An intelligent identification system for electric energy meter installation specifications according to claim 1, 2, 3, 4 or 5, characterized in that: It also includes a result output module, which is used to output the detection results, including whether the wiring is correct, the error type, and location information, and generate a detection report.

8. An intelligent identification system for electric energy meter installation specifications according to claim 1, 2, 3 or 4, characterized in that: The wiring recognition module also includes a text detection unit, which adopts the EAST model. The EAST model directly predicts the text area in the image through a fully convolutional network and merges the final detection results through non-maximum suppression.

9. An intelligent identification system for electric energy meter installation specifications according to claim 1, 2, 3 or 4, characterized in that: The wiring recognition module also includes a character recognition unit, which uses a CRNN model. The CRNN model extracts local features of image data by using a convolutional layer, then reduces the spatial dimension of the features by a pooling layer, and finally performs classification or regression tasks by a fully connected layer. The CRNN model extracts features from the image through CNN and processes these feature sequences through RNN. It performs time series modeling based on these feature sequences and introduces an attention mechanism into the sequence modeling. The specific attention mechanism adopts the following formula: ; where Q is the query matrix, K is the key matrix, V is the value matrix, and d k It is a dimension; Alternatively, the wiring recognition module also includes using an N-gram language model or BERT to correct the recognition results and using regular expressions to match fixed formats.

10. An intelligent identification system for electric energy meter installation specifications according to claim 1, 2, 3 or 4, characterized in that: The loss function is used in the binarization process to evaluate the difference between the binarized result of the model output and the true label. The loss function can be used to measure the degree of model fit. By calculating the difference between the output predicted by the model and the actual label, a loss value can be obtained, and then this loss value is minimized. The formula used is: is the function obtained after training, w and b are the two independent variables of the loss function, : The true label of the i-th sample, two parameters obtained by function training.

Images