Two-stage transformer substation pointer type meter non-calibration identification method and system

Abstract

The invention discloses a two-stage transformer substation pointer type meter non-calibration identification method and system, and the method specifically comprises the steps: collecting and making meter region detection data, and carrying out the expansion-cutting to construct a scale and pointer key point detection data set; training a meter area target detection model and a key point detection model by using the two data sets; sending an image of a meter to be detected into the detection model to obtain digital scales of the meter and key point information of a pointer; primary filtering is carried out, and digital scale key points are sequenced; carrying out secondary filtering on the sequenced digital scale key points by utilizing the characteristic information of the meter; finding out adjacent scale key points distributed on two sides of the pointer, and solving an intersection point of a straight line where the two adjacent scale key points are located and a straight line where the pointer is located; and reasoning the reading information of the pointer meter by using the information. Through the method, the reading information of the pointer meter can be quickly, accurately and automatically read, and the problems of high manpower inspection consumption, low efficiency, recording errors and the like are solved.

Description

Technical Field

[0001] This invention belongs to the technical field of intelligent inspection of substations, specifically relating to a method and system for identifying uncalibrated pointer-type meters in a two-stage substation. Background Technology

[0002] The continuous expansion of power systems places higher demands on the stable operation and security of the power grid. As a core component of the power system, the safe operation of substations is crucial for ensuring grid stability and power supply quality. In substations exposed to strong electromagnetic interference, digital / electronic instruments cannot operate stably, and pointer-type instruments remain widely used. Traditional manual meter reading methods are time-consuming, repetitive, prone to subjective errors, and risky in high-voltage environments, making them unsuitable for the operation and maintenance needs of large-scale substation clusters. While automatic meter reading identification methods based on manual calibration can reduce the workload of maintenance personnel to some extent, the algorithm requires extensive manual annotation, and complex situations such as on-site preset position offsets often lead to multiple rework sessions. Both accuracy and intelligence need improvement.

[0003] Therefore, there is an urgent need for an intelligent identification technology for substation pointer meters that balances efficiency and accuracy. Summary of the Invention

[0004] Purpose of the invention: In order to improve the reading efficiency and accuracy of pointer meters in substations, this invention provides a method and system for identifying pointer meters in substations. It uses a lightweight target detection network model to achieve target detection and key point detection, and processes the detected data to obtain the reading of the pointer meter.

[0005] Technical solution:

[0006] This invention proposes a method for identifying uncalibrated pointer-type meters in a two-stage substation, comprising:

[0007] Collect images of pointer meters in substations, perform image preprocessing and meter area annotation to form a meter type detection dataset, and perform key point annotation on images labeled as meter areas to form a key point dataset;

[0008] A target detection model is built based on yolov8-detect and trained using the meter type detection dataset to identify meter areas from substation pointer meter images; a key point detection model is built based on yolov8-pose and trained using a key point dataset to identify scale key points and pointer key points from meter area images.

[0009] The image of the pointer-type meter in the substation to be identified is acquired on site and serially fed into the trained target detection model and key point detection model. The model outputs several scale key points, one pointer key point and the corresponding confidence score of the image to be identified.

[0010] The model output is processed in the first stage to filter out duplicate scale key points. The filtered scale key points are sorted from smallest to largest to obtain a scale key point information table. All filtered scale key points are traversed, and scale key points with low confidence are deleted through the second stage of processing. The scale key point information table is then updated.

[0011] Based on the pointer key points output by the model, the straight line where the pointer is located is obtained. Connect the two adjacent scale key points on both sides of the pointer in the scale key point information table to obtain the intersection point of the key point line and the straight line where the pointer is located. Calculate the distance between the intersection point and the two adjacent scale key points to obtain the pointer reading.

[0012] Furthermore, the substation pointer meters include a sulfur hexafluoride pressure gauge, a leakage current meter, and an oil level gauge; the preprocessing includes data cleaning and data augmentation; the meter area labeling is to enclose the meters in the image of the substation pointer meters after data augmentation using a minimum bounding rectangle horizontal frame, the coordinate range of which does not exceed the image range.

[0013] Furthermore, the key point annotation includes:

[0014] The image labeled as the meter area is subjected to dilation-cropping processing. The pointer and scale numbers are labeled using the smallest bounding rectangle horizontal box. Two key points of the pointer and one key point of the scale are labeled using points. The horizontal box and its corresponding key points are associated by group_id. The two key points of the pointer are the pointer head position and the pointer tail position, and the key point of the scale is the endpoint position of the digital scale line on the dial.

[0015] Furthermore, the first stage of processing includes:

[0016] If the keypoint detection model outputs two or more identical scale keypoints, only the scale keypoint with the highest confidence is retained; if the output scale keypoints have corresponding horizontal boxes and the horizontal boxes corresponding to different scale keypoints overlap, only the scale keypoint with the highest confidence and its corresponding horizontal box are retained.

[0017] Furthermore, if the horizontal frames corresponding to the different numerical scale key points overlap, the determination is based on calculating the distance between the horizontal frames. The formula is:

[0018]

[0019] and This represents two horizontal boxes to be judged. If the value is greater than 0, there is overlap; otherwise, there is no overlap.

[0020] Furthermore, the sorted scale key point information table includes the dial numerals of the image to be identified and the coordinates of the scale key points corresponding to the numerals, arranged in ascending order of the dial numerals.

[0021] Furthermore, the second stage of processing includes:

[0022] Let any three adjacent scale keypoints in the keypoint information table be denoted as a group (P) in ascending order. i ,P j, P k The key point information table is traversed recursively, and each set of scale key points is judged as follows:

[0023] (1) Satisfies the condition of vector to vector It is a clockwise rotation relationship;

[0024] (2) Satisfies the vector The magnitude is less than the vector mold;

[0025] (3) Satisfies the vector to vector The included angle is less than 90°;

[0026] If any of the judgment rules are not met, delete the combination (P) from the key point information table. i ,P j, P k Update the list of key points by identifying the point with the lowest confidence level.

[0027] The process restarts the iteration based on the updated list of key points, until all combinations of key points satisfy the criteria.

[0028] Furthermore, obtaining the intersection point of the line connecting the key points and the straight line where the pointer is located includes:

[0029] Define the pointer head key point as P h The key point at the end of the pointer is P. t The intersection point is C. Calculate P. h The distance P is obtained by measuring the distance to all points in the key point information table. h The two nearest neighbors are key points;

[0030] Connect two nearest neighbor key points, connect P t and P hForm a pointer line. If the two nearest neighbor key points are distributed on both sides of the pointer, the intersection of the line connecting the two nearest neighbor key points and the pointer line is the required intersection point C.

[0031] If two nearest neighbor keypoints are located on the same side of the pointer, retain the distance P. h Recent key point P n ;

[0032] P n When the pointer is to the right, find the key point information table that corresponds to P. n The adjacent previous key point P n-1 P n-1 P n The intersection point of the line where the pointer is located and the line where the pointer is located is the required intersection point C; when P does not exist in the key point information table. n-1 At that time, there is no intersection point C, and the output pointer meter reading is less than the meter's minimum value;

[0033] P n When the pointer is to the left, find the key point information table that corresponds to P. n The adjacent next key point P n+1 P n P n+1 The intersection point of the line where the pointer is located and the line where the pointer is located is the required intersection point C; when P does not exist in the key point information table. n+1 At that time, there is no intersection point C, and the reading of the pointer meter is greater than the maximum value of the meter.

[0034] The present invention also proposes a calibration-free identification system for pointer-type meters in a two-stage substation, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of any of the aforementioned methods.

[0035] Furthermore, the computer program is applied to the substation remote intelligent inspection system based on the GRPC communication protocol interface, and exchanges information via the HTTP / 2 protocol.

[0036] Beneficial effects:

[0037] This invention proposes a strategy for jointly identifying pointer-type meter readings based on target detection and key point detection, which has the advantages of simple structure, fast recognition speed and high accuracy.

[0038] The present invention proposes a method for identifying uncalibrated pointer-type meters in substations. By designing a two-stage method, it fully utilizes the relationship features between meter scales, such as position, angle, and distance, to identify the model. The identification results of the meter, pointer, and dial scale are post-processed and filtered to achieve automatic reading of the readings of key meters in the substation.

[0039] This invention fully utilizes the uniqueness of the dial markings and the filtering and judgment based on the position, distance, and angle between the markings, overcoming the problem of reading inference errors caused by misdetection of markings. Compared with existing calibration and identification technologies, this invention eliminates the reliance on manual labeling of prior knowledge of the meter, achieving breakthrough improvements in deployment flexibility, environmental robustness, and generalization ability.

[0040] This invention utilizes the positional relationship between the pointer node of the meter and the key points of each scale in the key point information table to find the intersection point C, and then deduce the meter reading. This method is fast and reliable, greatly improves work efficiency, and significantly reduces long-term costs. Attached Figure Description

[0041] Figure 1 This is a flowchart of the method of the present invention;

[0042] Figure 2 This is an image of a pointer meter in this embodiment;

[0043] Figure 3 This is an image of a pointer meter that has been annotated in this embodiment;

[0044] Figure 4 This is a flowchart illustrating the image processing of a pointer meter. Detailed Implementation

[0045] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0046] This invention provides a two-stage method and system for identifying uncalibrated pointer-type meters in substations based on target detection and key point detection models. The process is as follows: Figure 1 As shown, it includes the following steps:

[0047] Step 1: Construct a meter detection dataset. Collect images of common types of pointer meters in substations, perform data cleaning, enhancement, and annotation to form a meter detection dataset, which is further divided into training, validation, and test sets.

[0048] High-resolution acquisition equipment was used to collect images of common types of pointer-type meters from multiple substations across China. The acquisition range covered climatic conditions such as extreme cold, high temperature, high humidity, and sandstorms, indoor and outdoor working environments with different electromagnetic interference intensities, and sulfur hexafluoride pressure gauges, leakage current gauges, oil level gauges, etc., with different service lives, including newly installed, normally operating, and old ones.

[0049] The acquired image data is filtered and processed to remove blurry, heavily noise-affected, or unlabelable images. Data augmentation is then applied to the cleaned data, including HSV color enhancement, rotation, translation, scaling, shearing, flipping, noise addition, and enhancement of special weather effects.

[0050] The specific steps are as follows: In the HSV color space, adjust the hue and saturation within the range of [-15°, 15°], with the adjustment ratios set to [0.8, 1.2] and the brightness within the range of [0.8, 1.2]. Rotate the pointer meter image at different angles, centering on any point in the image, and fill any blank areas with zero values. Randomly crop the meter image without cutting the main body of the meter. Randomly scale the image size. Add different types of noise to the training data, including Gaussian noise, salt-and-pepper noise, Poisson noise, speckle noise, uniform noise, multiplicative noise, and exponential noise. Enhance special weather effects by adding strong lighting, fogging, rain, snow, and snow cover effects to the pointer meter image. In practical application, select one or more of these methods in combination to expand the original single sample image into multiple images, significantly increasing the training load.

[0051] Using the Labelme tool, meter regions are labeled on the enhanced image data. A minimum bounding rectangle is used to enclose the corresponding meters, and the coordinates of this rectangle cannot exceed the image region. The labeled data is then divided into training, validation, and test sets in an 8:1:1 ratio. An example of a pointer meter image in this embodiment is shown below. Figure 2 As shown.

[0052] Specifically, the actual readings of the meters also need to be recorded for the test set data to calculate the evaluation indicators. In this embodiment, the Labelme open-source tool is used for annotation and recording. For example, if the reading of sulfur hexafluoride is 0.4 and the range is -0.1 to 0.9, it is marked as SF6&0.4&-0.1~0.9.

[0053] Step 2: Construct a dataset for detecting key points of the meter pointer and scale. Perform dilation-clipping processing on the meter area marked in Step 1, and mark the key points of the pointer and scale to form a dataset, which is divided into training set, validation set and test set in a ratio of 8:1:1.

[0054] Perform dilation-cropping processing on the meter area marked in step 1 to generate image data containing only the meter area. The dilation-cropping processing includes:

[0055] a. Obtain the horizontal bounding box of the meter area marked in step 1. Expand the four sides of the horizontal bounding box of the meter area outward by 10 pixels. If the expanded area exceeds the image boundary, the image boundary shall be the limit. Let the top left point of the marked horizontal bounding box of the meter area be... The pixel coordinates of the bottom right point are ,according to and Calculate the coordinates after dilation:

[0056]

[0057] in, In this embodiment, the number of pixels to be expanded is... w and h are the width and height of the image, respectively.

[0058] b. Crop the meter image along the boundary of the expanded frame to obtain a smaller image containing the meter area, as shown below:

[0059]

[0060] in, For the original image, The image is cropped, and x_min, x_max, y_min, and y_max are the dilated coordinate values.

[0061] Keypoint annotation is performed on the dilated-cropped image. A minimum bounding rectangle is used to annotate the pointer and scale numbers. Two keypoints on the pointer and one keypoint on the scale number are annotated using points. The horizontal bounding rectangle and its corresponding keypoint are associated using `group_id`. The two keypoints of the pointer are the pointer head position and the pointer tail position. The keypoint of the scale number is the endpoint position of the numerical scale line on the dial. The keypoint annotation in this embodiment is as follows: Figure 3 As shown in the figure, in this embodiment, the labeled data is divided into training set, validation set and test set in a ratio of 8:1:1.

[0062] Step 3: Build and train the detection model. Use yolov8-detect to build an object detection model, and train the object detection model using the meter detection dataset described in Step 1; use yolov8-pose to build a keypoint detection model, and train the pointer and scale keypoint detection model using the meter pointer and scale keypoint detection dataset described in Step 2.

[0063] The target detection model takes the original image containing the meter as input and outputs the bounding horizontal box of the meter area and the corresponding confidence score, thus achieving accurate identification of the meter type. The key point detection model takes the meter area image after dilation and cropping on the original image based on the target detection model results as input and outputs the position of the digital scale in the dial, the position of the key point of the scale, and the corresponding confidence score.

[0064] Step 4: Design a two-stage processing mechanism. Perform one-stage processing on the pointer and scale key point information detected by the model, and then sort and filter the processed scale key points in the second stage to obtain a scale key point information table.

[0065] The first stage of processing involves the pointer and tick mark keypoint information detected by the model, filtering out duplicate keypoints. This specifically includes:

[0066] (1) If the detection results contain two or more key points with the same numerical scale label, keep the one with the highest confidence and delete the other key points with the same numerical scale label;

[0067] (2) If a key point in the detection results has a corresponding horizontal frame, and the horizontal frames corresponding to the key points of each numerical scale overlap, retain the key point with the highest confidence level and its corresponding horizontal frame, and delete other key points and their corresponding horizontal frames that overlap with it. The criterion for determining whether horizontal frames overlap is the distance between the two horizontal frames. If the value is greater than 0, then there is overlap; otherwise, there is no overlap. The calculation method is as follows:

[0068]

[0069] and This indicates that it is necessary to determine whether two rectangles overlap. The sequence of tick marks after the first stage of processing is then sorted in ascending order according to the size of their corresponding numerical tick marks.

[0070] The sorted keypoint sequence is filtered a second time (location_filter) using the clockwise rotation relationship between scales, the length relationship between distances, and the angle relationship to obtain the scale keypoint information table.

[0071] The scale key point information table is the output of step 3 after the first stage of processing described above. It includes the digital scale of the dial of the image to be identified and the key points corresponding to the digital scale, arranged in ascending order, to support the second stage of processing and the inference of meter readings in subsequent steps.

[0072] For ease of explanation, let's denote three adjacent key points in the key point information table as P, in ascending order. i P j and P k The second-stage processing and filtering mechanism recursively evaluates three points sequentially. The evaluation specifically includes:

[0073] (1) Determining the clockwise rotation relationship of the scale:

[0074] Since the scales of common pointer-type meters in substations are arranged clockwise, the sequence of key points calculated and sorted through the model should also satisfy the vector... to vector It is a clockwise rotation relationship;

[0075] (2) Determining the relationship between the lengths of the graduations:

[0076] In the sorted keypoint sequence, the vector from the smallest tick to the middle tick should satisfy the condition that... Vectors whose magnitude is less than the distance from the smallest to the largest scale mold;

[0077] (3) Determination of the included angle relationship:

[0078] This invention's embodiments are for pointer-type meters with a relatively compact scale distribution. The sorted sequence of key points should satisfy the vector... to vector The included angle is less than 90°.

[0079] Three points are sequentially selected from the scale keypoint information table. A two-stage post-processing filtering mechanism is used. If any point does not meet any of the judgment rules, the point with the lowest model inference confidence is removed from the information list. The keypoint information list is then updated, and the recursive judgment is repeated in the updated order until all combinations of scale keypoints meet the judgment conditions. The overall process pseudocode is explained below:

[0080] Input: Assume the list of key point information is P={P0,P1,…,P} N};

[0081] for i = 1 ~ N-2:

[0082] j = i + 1, k = i + 2;

[0083] Judgment 1: arrive Does it conform to a clockwise rotation relationship?

[0084] Judgment 2: < ;

[0085] Judgment 3: arrive The included angle is less than 90°;

[0086] If none of the above conditions are met:

[0087] Delete the point with the lowest confidence in {Pi, Pj, Pk}; break;

[0088] After deleting the point with the lowest confidence in {Pi,Pj,Pk}, update the key point information list P';

[0089] Iterate through the relationships between points in P' until all combinations of points satisfy the determination conditions.

[0090] Output: A list of key points that satisfy all the criteria.

[0091] Specifically, assuming there are 6 points in the scale key point information table obtained from the model output in step 3, the following grouping is used for traversal and judgment: (1,2,3), (2,3,4), (3,4,5), (4,5,6); among them, the second group (2,3,4) does not meet the judgment condition, and the third key point has the lowest confidence, so point 3 is deleted, and the updated key point information table still has five points: 1, 2, 4, 5, 6; the above judgment is repeated for the updated key point information table, traversing the following grouping: (1,2,4), (2,4,5), (4,5,6), until all groups meet the three judgment conditions; at this time, the key point information table is finally used as the key point information table for reading inference and enters step 5 for calculation.

[0092] Step 5: Design the reading reasoning process. Using the key point information table obtained in Step 4, acquire the adjacent scale key points on both sides of the pointer, and find the intersection point of the line connecting two adjacent key points and the straight line where the pointer is located; using the positional distance relationship between this intersection point and the two key points, as well as the scale information, calculate the meter reading by "substituting a chord for an arc". The reading reasoning flowchart is as follows. Figure 4 As shown.

[0093] Find the intersection point C of the line connecting two adjacent key points and the line where the pointer is located. This includes the following sub-steps:

[0094] (1) Define the key point of the pointer head as P. h The key point at the end of the pointer is P. t Calculate P h Record the distance P to all points in the key point information table. h The two most recent key points are P3 and P4;

[0095] (2) Determine whether P3 and P4 are adjacent in the key point information table: If P3 and P4 are adjacent, determine whether P3 and P4 are distributed on both sides of the pointer; if they are not adjacent, report the error "There is a problem with the detection sequence" or "The pointer points between the maximum and minimum values";

[0096] (3) Connect P3 and P4 to form a straight line P3P4, and connect P t and P h Forming a straight line P t P h Find the lines P3P4 and P... t P h Determine whether the intersection point C is located between the key points P3 and P4 on the digital scale.

[0097] (4) If yes, then P3 and P4 are located on both sides of the pointer, and C is the intersection of the line connecting the two adjacent key points and the line where the pointer is located. Skip the subsequent steps (5)-(6). If no, then P3 and P4 are located on one side of the pointer, retaining the distance P.h Recent key points.

[0098] (5) Let P3 be the key point retained in step (4). Determine whether P3 is the previous key point of P4 in the key point information table.

[0099] If so, find the preceding key point adjacent to P3 in the key point information table, let's call it P2 here, and calculate the line containing P2 and P3 and the line containing P. t P h The intersection point C. If there is no preceding key point adjacent to P3 in the key point information table, that is, P3 is the first key point in the key point information table, it proves that P3 is to the right of the pointer, and the pointer points to a value less than the minimum value in the key point information table.

[0100] If not, find the next key point adjacent to P3, let's call it P5 in this case. Find the line containing P3 and P5 and the line containing P. t P h The intersection point C. If there is no next key point adjacent to P3 in the key point information table, that is, P3 is the tail key point in the key point information table, it proves that P3 is to the left of the pointer, and the pointer points to a value greater than the maximum value in the key point information table.

[0101] (6) Determine whether the intersection point C in (5) is between P2 and P3 or between P3 and P5. If so, then point C is the intersection point of the line connecting two adjacent key points and the line where the pointer is located. If not, then there is a problem with the error detection sequence.

[0102] After obtaining the intersection point C through the above steps (1)-(6), the meter reading is automatically read by using the positional distance relationship between the intersection point and the adjacent key points and the scale information, in the way of "using the chord to replace the arc".

[0103] Let P3 and P4 be the key points of the adjacent numerical scales at both ends of the pointer, and let C be the intersection of the line containing the pointer and the line P3P4. Let the scale numbers represented by P3 and P4 be denoted as... and The length of line segment P3C is denoted as The length of line segment P3P4 is denoted as meter reading for:

[0104]

[0105] Furthermore, to verify the effectiveness of the method in this embodiment, an evaluation method was designed: using the test set in step 1, the model effectiveness was comprehensively evaluated using three indicators: accuracy, precision, and average reading error.

[0106] The correct definition of meter reading inference is that the inference error is less than or equal to 2% of the meter's measurement range. Let the model inference reading be... The actual reading is The meter's measurement range is The indicator is calculated as shown in the formula:

[0107]

[0108] (1) Accuracy is expressed as: the number of meter readings correctly inferred by the model / the total number of meters. Let the number of meter readings correctly inferred by the model be... The total number of meters is Then the accuracy of reading reasoning It can be represented as:

[0109]

[0110] (2) The accuracy rate is expressed as: the number of meter readings correctly inferred by the model / the number of meter readings inferred by the model, as follows:

[0111] Suppose the number of meter readings correctly inferred by the model is... The number of readings on the table can be deduced to be... Then the reading accuracy It can be represented as:

[0112]

[0113] in, It contains two parts: the first part is the number of correctly deduced readings, denoted as . (Right now The other part consists of the number of inferred readings with an inference error greater than 2%, denoted as... , Then the reading accuracy It can also be expressed as:

[0114]

[0115] (3) Evaluate the model's effectiveness from the perspective of average reading error, assuming the total number of meters is... , No. The model inference reading of each meter is The actual reading is The range is Then the average reading error Represented as:

[0116]

[0117] Tests were conducted on 1000 actual meter images generated by the substation remote intelligent inspection system, and the model accuracy, reading precision, and average reading error were found to be 95.8%, 96.4%, and 3.2%, respectively. This demonstrates that the pointer meter identification method designed in this invention has a high accuracy rate.

[0118] Example 2

[0119] This invention also proposes a calibration-free identification system for pointer-type meters in substations, comprising a memory, a processor, and a computer program stored in the memory. The computer program loads the optimal model weights after training, forms an algorithm service based on the GPRC communication protocol interface, and applies it to a remote intelligent inspection system for substations. Running this computer program invokes the identification method of Example 1 to identify the readings of pointer-type meters.

[0120] The client device of the identification system is configured to send a request message encoded in Protocol Buffers format to the server via the HTTP / 2 protocol; the server device is configured to receive the above request message, execute the meter identification service method, generate the meter identification result, and return a response message encoded in Protocol Buffers format via the HTTP / 2 protocol, including information such as the automatically read meter reading, meter area location, and type.

[0121] The communication interface between the client device and the server device is defined through a predefined .proto file, which contains: at least one service definition, which contains one or more remotely invoked methods; each of the remotely invoked methods corresponds to at least one request message type and at least one response message type; the request message type and response message type use Protocol Buffers syntax to define the message structure.

Claims

1. A two-stage substation pointer-type meter no-calibration identification method, characterized in that, The application comprises the following steps: Collecting substation pointer meter image, image preprocessing and meter area labeling, forming meter type detection dataset, key point labeling for images labeled as meter area, forming key point dataset; Building a target detection model based on yolov8-detect, training using the meter type detection dataset, to identify meter area from substation pointer meter image; Building a key point detection model based on yolov8-pose, training using the key point dataset, to identify scale key points and pointer key points from images of meter area; On-site acquisition of substation pointer meter images to be identified, serially inputting the trained target detection model and key point detection model, and the model outputting a plurality of scale key points, a pointer key point and corresponding confidence of the image to be identified; First-stage processing of the model output to filter duplicate scale key points, sorting the filtered scale key points from small to large to obtain a scale key point information table; Traversing all filtered scale key points, deleting scale key points with low confidence through second-stage processing, and updating the scale key point information table; According to the pointer key point output by the model, obtaining a straight line where the pointer is located, connecting adjacent two scale key points on both sides of the pointer in the scale key point information table to obtain the intersection of the key point connecting line and the straight line where the pointer is located, calculating the distance between the intersection and the adjacent two scale key points to obtain the pointer meter reading.

2. The identification method according to claim 1, characterized in that, The substation pointer meter comprises a sulfur hexafluoride pressure gauge, a leakage current meter and an oil level gauge; the preprocessing comprises data cleaning and data enhancement; the meter area labeling is performed on the substation pointer meter image after data enhancement, and a minimum bounding rectangle horizontal box is used to enclose the meter in the image, and the coordinate range of the horizontal box does not exceed the image range.

3. The identification method according to claim 2, characterized in that, The key point labeling comprises: Performing inflation-cutting processing on the image labeled as the meter area, using a minimum bounding rectangle horizontal box to label the pointer and the scale number, using points to label two key points of the pointer and one key point of the scale, and associating the horizontal box and the corresponding key points through group_id; the two key points of the pointer are the pointer head position and the pointer tail position, and the key point of the scale is the endpoint position of the dial number scale line.

4. The identification method according to claim 3, characterized in that, The first-stage processing comprises: If the key point detection model outputs two or more same scale key points, only the scale key point with the highest confidence is retained; if the output scale key points have corresponding horizontal boxes and the horizontal boxes corresponding to different number scale key points overlap, only the scale key point with the highest confidence and the corresponding horizontal box are retained.

5. The identification method according to claim 4, characterized in that, The horizontal boxes corresponding to the different digital scale key points appear overlapping, and the judgment basis is to calculate the distance between the horizontal boxes , and the formula is: and denotes the two horizontal boxes to be judged, and is greater than 0, then there is an overlap, otherwise there is no overlap.

6. The identification method according to claim 5, characterized in that, The scale key point information table after sorting comprises dial number scales of the image to be identified, scale key point coordinates corresponding to the number scales, and is arranged in order from small to large according to the dial number scales.

7. The identification method according to claim 6, characterized in that, The second-stage processing comprises: Let any three adjacent scale key points in the key point information table be denoted as a combination (P i ,P j, P k ) in ascending order, and traverse the key point information table in a recursive manner to sequentially determine a group of scale key points as follows: (1) Satisfies the vector to the vector is a clockwise rotation relationship; (2) meet the vector modulus less than the vector modulus; (3) the vector to the vector makes an angle of less than 90°; If any of the decision rules are not satisfied, the point with the lowest confidence in the combination (P i ,P j, P k ) is deleted from the key point information table, and the key point information list is updated. Based on the updated key point information list, re-starting traversal until all scale key point combinations meet the determination condition.

8. The identification method according to claim 7, characterized in that, The intersection of the key point connecting line and the straight line where the pointer is located comprises: Define the pointer head key point as P h , the pointer tail key point as P t , the intersection point as C, calculate the distance between P h and all points in the key point information table, and obtain the two nearest neighbor key points closest to P h ; connecting two nearest neighbor key points, connecting P t and P h forming a pointer straight line, if two nearest neighbor key points are distributed on two sides of the pointer, the intersection point of the connecting line of the two nearest neighbor key points and the pointer straight line is the intersection point C; If two nearest neighbor key points are distributed on the same side of the pointer, keep the distance P h the nearest key point P n ; P n When the pointer is on the right side, find the previous key point P n adjacent to P n-1 in the key point information table, and the intersection of the line on which P n-1 P n is located and the line of the pointer is the intersection point C; when there is no P n-1 in the key point information table, there is no intersection point C, and the output of the pointer table meter reading is less than the minimum value of the meter. P n When the pointer is on the left side, find the key point P n adjacent to P n+1 in the key point information table, and the intersection of the line connecting P n and P n+1 and the line of the pointer is the intersection point C; when there is no P n+1 in the key point information table, there is no intersection point C, and the pointer meter reading is greater than the maximum value of the meter.

9. A two-stage substation pointer-type meter no-dialing identification system comprising a memory, a processor, and a computer program stored on the memory, wherein, The processor executes the computer program to implement the steps of any method of claims 1-8.

10. The identification system of claim 9, wherein, The computer program is applied to a transformer substation remote intelligent patrol system based on a GRPC communication protocol interface, and information interaction is performed through an HTTP / 2 protocol.

Images