Intelligent substation pressing plate state detection method and system based on image recognition

Abstract

The invention discloses an intelligent substation pressing plate state detection method and system based on image recognition, and the method comprises the steps: obtaining an image of a substation control cabinet, dividing a detection region according to the arrangement rule of pressing plates in the image, and constructing a region index containing the positions and switching states of the pressing plates; the current position and the initial position of the pressing plate are compared, the position offset is calculated, the detection areas with the position offset exceeding the preset range are screened and marked as abnormal areas, and abnormal feature data are formed; the abnormal feature data and the correction judgment rule are utilized, and an optimized dynamic judgment rule is generated in combination with a preset range of commissioning and retreating state judgment in the commissioning and retreating state first correction judgment rule of the abnormal area; applying the dynamic judgment rule to all the detection areas, judging whether the current position and the on-off state of the pressing plate conform to the judgment rule or not, marking the pressing plate which does not conform to the judgment rule, correcting the judgment rule for the second time in combination with the marked abnormal characteristics, forming an updated dynamic judgment rule, and outputting a detection report.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent monitoring of electric power equipment, and more specifically, to an intelligent substation pressure plate status detection method and system based on image recognition. Background Art

[0002] In the operation and maintenance of substations, accurate detection of the pressure plate status is crucial to the safe operation of the power system. The pressure plate's on / off status indicates the operating status of the relevant equipment or circuit, and its abnormal status may cause control command failure, equipment operation abnormality, or even system failure. However, the existing technology has the following problems in pressure plate status detection:

[0003] (1) Manual inspection is inefficient and error-prone

[0004] Currently, many substations still rely on manual inspections to record the status of the pressure plates. This method not only consumes a lot of manpower and time, but is also prone to errors in status recording due to human negligence, especially in scenarios where the pressure plates are arranged in a complex manner or in large numbers, and accuracy is difficult to guarantee.

[0005] (2) Limited image recognition accuracy

[0006] When detecting the status of the pressure plate, the existing image recognition technology is usually based on simple color recognition or position matching, which is difficult to handle complex situations such as uneven arrangement of pressure plates and blurred insertion and withdrawal marks. For example, due to long-term use, the pressure plate may have position shifts, changes in arrangement characteristics, or degradation of color marks, resulting in the inability of traditional methods to effectively identify abnormalities.

[0007] (3) Lack of dynamic adaptive rules

[0008] Traditional detection methods mostly set the platen status judgment criteria based on static rules, and it is difficult to dynamically optimize the judgment rules according to the real-time changes in the detection scene. For example, for the judgment of the platen's insertion and retraction status, a fixed threshold cannot cope with changes in ambient light or the diverse characteristics of abnormal platens. Summary of the invention

[0009] In order to solve the deficiencies in the prior art, the present invention provides a method and system for detecting the status of a pressure plate in an intelligent substation based on image recognition.

[0010] The present invention adopts the following technical solution.

[0011] A first aspect of the present invention provides a method for detecting a pressure plate state in an intelligent substation based on image recognition, comprising the following steps:

[0012] Step 1: Obtain an image of the substation control cabinet, and divide the image into multiple detection areas according to the arrangement rules of the pressure plates. Each detection area contains pressure plates that are physically adjacent or functionally related. The area index is constructed based on the initial position and the insertion and withdrawal status of the pressure plates in each detection area.

[0013] Step 2: Based on the area index, determine the position offset between the current position and the initial position of the pressing plate in each detection area, select the detection area whose position offset exceeds the preset range and mark it as an abnormal area, extract the boundary contour, arrangement characteristics and insertion and withdrawal status of the pressing plate in the abnormal area, and form abnormal feature data;

[0014] Step 3, setting the judgment rule by analyzing the abnormal feature data, wherein the position offset range in the judgment rule is set by using the position offset of the pressure plate in the abnormal area, the arrangement feature range in the judgment rule is set by the data distribution of the arrangement feature, and the range of the judgment of the insertion and withdrawal state is set in combination with the data of the insertion and withdrawal state, so as to generate a dynamic judgment rule;

[0015] Step 4, apply the dynamic judgment rules to all detection areas, determine whether the current position and insertion and retraction status of the pressure plate meet the judgment rules, mark the pressure plates that do not meet the judgment rules, extract the area index and abnormal features corresponding to the pressure plates that do not meet the judgment rules, modify the judgment rules based on the marked abnormal features, form an updated dynamic judgment rule, and output a detection report containing the pressure plate status, area index and abnormal features.

[0016] Preferably, the image is divided into a plurality of detection areas according to the arrangement rule of the pressure plates, each detection area contains pressure plates that are physically adjacent or functionally related, including:

[0017] Physically adjacent platens are divided as follows:

[0018] For one-dimensional linearly arranged pressure plates, the physical positions are determined to be adjacent based on whether the spacing between the center points of adjacent pressure plates is uniform, and the pressure plates whose continuous spacing is within the set range are divided into the same detection area;

[0019] For the platens arranged in a two-dimensional matrix, the detection areas are divided into rows or columns according to the alignment of the boundary lines in the horizontal or vertical direction and whether the spacing between the center points is uniform;

[0020] The function-related pressure plates are divided as follows:

[0021] According to the correlation of the pressure plates in the circuit control logic, the pressure plates involved in the same functional module are classified into the same detection area;

[0022] Combined with the substation control cabinet design specifications, the pressure plates are logically divided according to functional areas so that the pressure plates in the detection area are functionally consistent;

[0023] The division principles include:

[0024] If the number of platens in the detection area exceeds the maximum processing unit, it will be re-subdivided according to the physical location adjacent rule;

[0025] If there is a conflict between physical location and functional zoning, priority shall be given to dividing the inspection area by functionally related pressure plates.

[0026] Preferably, the construction of the regional index including the platen position and the insertion and withdrawal status includes:

[0027] The region index is stored as a two-dimensional array of key-value pairs:

[0028] The two-dimensional array includes rows and columns corresponding to the position of the pressure plate, and each unit records the pixel coordinates, the insertion and withdrawal status and the boundary information of the pressure plate;

[0029] The key-value pair includes the area number as the key, and stores the number, coordinates, insertion and retraction status and boundary information of the pressure plate in the area.

[0030] Preferably, the method of determining the position offset between the current position and the initial position of the pressure plate in each detection area based on the area index, screening the detection areas whose position offset exceeds a preset range and marking them as abnormal areas includes:

[0031] The edge detection is used to obtain the boundary contour of the platen. The geometric center of the platen in the image is calculated based on the contour pixels as the current position. The current position is compared with the center coordinates of the initial position, and the quantized offset is calculated as follows:

[0032]

[0033] In the formula, x, y, x i and i are the pixel coordinates of the boundary points respectively; N is the total number of pixels; the initial coordinates of the pressure plate (x0, y0);

[0034] The preset ranges are adjusted based on the deviation statistics of abnormal platens:

[0035] The distribution data of the platen offset in the detection area is statistically analyzed; the offset range is dynamically set according to the 95% quantile of the offset distribution.

[0036] Preferably, the step of extracting the boundary contour, arrangement characteristics and insertion and withdrawal status of the pressing plate in the abnormal area to form abnormal characteristic data includes:

[0037] The edge detection is used to obtain the boundary contour of the pressing plate, and the features of the boundary contour are quantified, including the boundary length, the number of corner points and the closure; the boundary length is the total number of pixels of the contour line segment, which is used to determine whether the pressing plate has defects; the number of corner points is obtained by the Harris corner point detection algorithm to obtain the number of feature points on the boundary, which is used to verify the integrity of the boundary; the closure is used to determine whether the contour forms a closed curve, and the distance between the first and last boundary points is calculated for detection;

[0038] Extracting arrangement features includes determining whether the arrangement complies with the rules by calculating the spacing between adjacent platens and alignment deviations; the spacing between platens is calculated as follows:

[0039]

[0040] In the formula, x i+1 and i+1 are the pixel coordinates of the boundary points respectively;

[0041] Fit the straight line of all the center points of the platens and calculate the average deviation from the point to the straight line. If the deviation exceeds the threshold, it is judged as misalignment;

[0042] Extracting the insertion and withdrawal status includes extracting the insertion and withdrawal status based on the color of the mark on the surface of the pressure plate, red for insertion and green for withdrawal; using color space segmentation to extract the insertion and withdrawal mark area, and the extracted mark color determines the insertion and withdrawal status by counting the proportion of the color components.

[0043] Preferably, the setting of determination rules by analyzing abnormal characteristic data includes:

[0044] The basis for the revision of the determination rules is as follows:

[0045] Based on the extracted abnormal feature data, including position offset abnormality, arrangement feature abnormality and investment and withdrawal status abnormality; dynamically adjust the rule parameters by analyzing the distribution statistics of the abnormal feature data, including the offset range is set by the 95% quantile of the abnormal area, the arrangement feature weight is adjusted according to the abnormal frequency, and the investment and withdrawal status threshold is dynamically corrected by the color component mean;

[0046] By counting the frequency of arrangement features of abnormal pressure plates, the weight of the arrangement features is dynamically adjusted when the abnormal proportion exceeds 30%, and the arrangement parameter range is modified at the same time.

[0047] Preferably, the combination of the investment and withdrawal status data to set the investment and withdrawal status judgment range and generate dynamic judgment rules includes:

[0048] Re-correct the judgment criteria of the investment and withdrawal status according to the color component and state switching frequency of the abnormal area, including dynamically adjusting the color threshold and setting the upper limit of the state switching frequency to include 10 times / minute;

[0049] Dynamically adjust the color threshold as follows:

[0050] T 红 =μ 红 -σ 红 ,T 绿 =μ 绿 -σ 绿

[0051] Where, T 红 and T 绿 are the thresholds for red and green respectively; μ 红 and μ 绿 are the means of the investment state and withdrawal state respectively; σ 红 and σ 绿 are the standard deviations of the investment state and withdrawal state respectively.

[0052] The state transition frequency is used to detect whether the number of state changes of the pressure plate within a certain period of time is abnormal.

[0053] Preferably, the dynamic determination rule is applied to all detection areas, judging whether the current position and the insertion and withdrawal state of the pressing plate meet the determination rule, and marking the pressing plate that does not meet the determination rule, including:

[0054] Position offset determination includes:

[0055] For each platen, calculate the offset between the current position and the initial position and compare it with the optimized offset range; the offset that exceeds the range is marked as abnormal;

[0056] Arrangement feature determination includes:

[0057] According to the arrangement characteristics of the platen, determine whether it meets the optimized arrangement rules; the alignment deviation exceeding the threshold is marked as abnormal;

[0058] The investment and withdrawal status determination includes:

[0059] The color and switching frequency of the pressure plate's insertion and withdrawal status are analyzed to determine whether it complies with the insertion and withdrawal status rules; abnormal colors and switching frequencies are marked as abnormal.

[0060] Preferably, the extracting of the region index and abnormal features corresponding to the pressing plate that does not meet the determination rule, and correcting the determination rule in combination with the marked abnormal features to form an updated dynamic determination rule includes:

[0061] In the process of rule determination, all platen features marked as abnormal are extracted, including the offset and direction beyond the offset range, which are recorded as position offset abnormalities; the arrangement center point beyond the spacing range and deviation is recorded as arrangement feature abnormalities; the color component and switching frequency that do not meet the throw-in and throw-out rules are recorded as throw-in and throw-out state abnormalities;

[0062] The extracted abnormal features are bound to the area index data of the corresponding detection area, including position binding and area binding; the position binding records the abnormal features through the index position of the pressure plate in the two-dimensional array; the area binding generates an abnormal feature list for each detection area, recording the area number and abnormal pressure plate information.

[0063] The second aspect of the present invention provides an intelligent substation pressure plate state detection system based on image recognition, comprising: an image acquisition module, a feature extraction module, a rule optimization module and an anomaly detection module;

[0064] The image acquisition module is used to obtain the image of the substation control cabinet, divide the detection area according to the arrangement rules of the pressure plates in the image, and record the initial position and the throw and retract status of the pressure plates in each detection area to construct a regional index containing the position and throw and retract status of the pressure plates;

[0065] A feature extraction module is used to compare the current position of the pressing plate with the initial position based on the area index, calculate the position offset, filter the detection area whose position offset exceeds the preset range and mark it as an abnormal area, and extract the boundary contour, arrangement characteristics and insertion and retraction status of the pressing plate in the abnormal area to form abnormal feature data;

[0066] A rule optimization module is used to use the abnormal feature data to correct the preset range of position offset and the priority of arrangement features in the judgment rule, and to first correct the preset range of the judgment of the throw and withdraw state in combination with the throw and withdraw state of the abnormal area to generate an optimized dynamic judgment rule;

[0067] The anomaly detection module is used to apply the optimized dynamic judgment rules to all detection areas, determine whether the current position and insertion and retraction status of the pressure plate conform to the rules, mark the pressure plate that does not conform to the rules, extract the area index and abnormal features corresponding to the pressure plate that does not conform to the rules, and revise the judgment rules for a second time in combination with the marked abnormal features to form an updated dynamic judgment rule, and output a detection report containing the pressure plate status, area index and abnormal features.

[0068] Compared with the prior art, the beneficial effects of the present invention include at least: the intelligent substation pressure plate status detection method and system based on image recognition provided by the present invention, by introducing a dynamic judgment rule optimization mechanism, can effectively solve the detection problems in complex scenarios such as pressure plate position offset, uneven arrangement and abnormal investment and withdrawal status, and realize the automation, efficiency and high-precision detection of pressure plate status; through two rule corrections and closed-loop optimization, the system has good dynamic adaptability and can adjust the judgment criteria in real time according to different detection environments; the final output includes a detailed detection report of pressure plate status, regional index and abnormal characteristics, which provides accurate and reliable data support for the operation and maintenance and system optimization of substation pressure plates, greatly reduces manual intervention, and improves detection efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a network architecture diagram of the ECA attention mechanism provided according to an embodiment of the present invention;

[0070] Figure 2 is a diagram of an angle cost calculation process provided according to an embodiment of the present invention;

[0071] Figure 3 is a distance cost calculation principle diagram provided according to an embodiment of the present invention;

[0072] Figure 4 It is a schematic diagram of the IoU cost calculation principle provided in accordance with an embodiment of the present invention;

[0073] Figure 5 YOLOE, YOLOE+ECA, YOLOE provided according to the embodiments of the present invention

[0074] +SIoU, detection effect diagram of the algorithm of the present invention;

[0075] Figure 6 It is an architecture diagram of a pressure plate status monitoring system for a smart substation provided according to an embodiment of the present invention;

[0076] Figure 7 This is a WeChat applet front-end interface diagram provided according to an embodiment of the present invention;

[0077] Figure 8 This is a diagram of a substation control cabinet management interface provided in accordance with an embodiment of the present invention;

[0078] Fig. 9 is a diagram of a pressure plate detection function provided according to an embodiment of the present invention;

[0079] Fig.10 is a visualization diagram of the detection results provided according to an embodiment of the present invention;

[0080] Fig.11 This is an example diagram of an image collected by a pressure plate of a substation control cabinet provided in accordance with an embodiment of the present invention;

[0081] Fig.12 is a diagram showing the noise reduction effect provided in accordance with an embodiment of the present invention;

[0082] Fig.13 It is a diagram of the insertion and retraction state of the pressing plate provided in accordance with an embodiment of the present invention;

[0083] Fig.14 This is an example diagram of the labeling of a pressure plate insertion and retraction status detection data set provided in accordance with an embodiment of the present invention. DETAILED DESCRIPTION

[0084] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

[0085] Example 1 of the present invention provides a method for detecting the state of a pressure plate in an intelligent substation based on image recognition, comprising the following steps:

[0086] Step 1: Obtain an image of the substation control cabinet, and divide the image into multiple detection areas according to the arrangement rules of the pressure plates. Each detection area contains pressure plates that are physically adjacent or functionally related. The area index is constructed based on the initial position and the insertion and withdrawal status of the pressure plates in each detection area.

[0087] Preferably, dividing the detection area according to the arrangement rule of the pressure plate in the image includes:

[0088] The pressure plates are distributed in the substation control cabinet according to physical arrangement rules, including one-dimensional linear arrangement and two-dimensional matrix arrangement;

[0089] In the image of the substation control cabinet, the matrix-arranged pressure plates are characterized by horizontal and vertical boundary line alignment and uniform center point spacing; the linearly arranged pressure plates are characterized by vertical boundary line alignment and uniform center point spacing;

[0090] The image is divided into multiple inspection areas based on the arrangement characteristics of the platens, including alignment and spacing uniformity.

[0091] Preferably, the construction of the regional index including the platen position and the insertion and withdrawal status includes:

[0092] The region index is stored in a two-dimensional array of key-value pairs:

[0093] The two-dimensional array includes rows and columns corresponding to the position of the pressure plate, and each unit records the pixel coordinates, the insertion and withdrawal status and the boundary information of the pressure plate;

[0094] The key-value pair includes the area number as the key, and stores the number, coordinates, insertion and retraction status and boundary information of the pressure plate in the area.

[0095] Step 2: Based on the area index, determine the position offset between the current position and the initial position of the pressing plate in each detection area, select the detection area whose position offset exceeds the preset range and mark it as an abnormal area, extract the boundary contour, arrangement characteristics and insertion and withdrawal status of the pressing plate in the abnormal area, and form abnormal feature data;

[0096] Preferably, the step of comparing the current position of the pressing plate with the initial position, calculating the position offset, screening the detection area where the position offset exceeds a preset range and marking it as an abnormal area comprises:

[0097] The edge detection is used to obtain the boundary contour of the platen. The geometric center of the platen in the image is calculated based on the contour pixels as the current position. The current position is compared with the center coordinates of the initial position, and the quantized offset is calculated as follows:

[0098]

[0099] In the formula, x, y, x i and i are the pixel coordinates of the boundary points respectively; N is the total number of pixels; the initial coordinates of the pressure plate (x0, y0);

[0100] The preset ranges are adjusted based on the deviation statistics of abnormal platens:

[0101] The distribution data of the platen offset in the detection area is statistically analyzed; the offset range is dynamically set according to the 95% quantile of the offset distribution.

[0102] Preferably, the step of extracting the boundary contour, arrangement characteristics and insertion and withdrawal status of the pressing plate in the abnormal area to form abnormal characteristic data includes:

[0103] The edge detection is used to obtain the boundary contour of the pressing plate, and the features of the boundary contour are quantified, including the boundary length, the number of corner points and the closure; the boundary length is the total number of pixels of the contour line segment, which is used to determine whether the pressing plate has defects; the number of corner points is obtained by the Harris corner point detection algorithm to obtain the number of feature points on the boundary, which is used to verify the integrity of the boundary; the closure is used to determine whether the contour forms a closed curve, and the distance between the first and last boundary points is calculated for detection;

[0104] Extracting arrangement features includes determining whether the arrangement complies with the rules by calculating the spacing between adjacent platens and alignment deviations; the spacing between platens is calculated as follows:

[0105]

[0106] In the formula, x i+1 and i+1 are the pixel coordinates of the boundary points respectively;

[0107] Fit the straight line of all the center points of the platens and calculate the average deviation from the point to the straight line. If the deviation exceeds the threshold, it is judged as misalignment;

[0108] Extracting the insertion and withdrawal status includes extracting the insertion and withdrawal status based on the color of the mark on the surface of the pressure plate, red for insertion and green for withdrawal; using color space segmentation to extract the insertion and withdrawal mark area, and the extracted mark color determines the insertion and withdrawal status by counting the proportion of the color components.

[0109] Step 3, setting the judgment rule by analyzing the abnormal feature data, wherein the position offset range in the judgment rule is set by using the position offset of the pressure plate in the abnormal area, the arrangement feature range in the judgment rule is set by the data distribution of the arrangement feature, and the range of the judgment of the insertion and withdrawal state is set in combination with the data of the insertion and withdrawal state, so as to generate a dynamic judgment rule;

[0110] Specifically, the determination rules are dynamically set, and the specific process is as follows:

[0111] Setting the position offset range

[0112] The abnormal feature data contains the position offset of the pressure plate in the abnormal area. The position offset is calculated by the coordinate difference between the current position and the initial position and quantified by the Euclidean distance. By statistically analyzing the distribution characteristics of these position offsets, taking the mean and 95% quantile to set the upper and lower limits of the position offset range, it is ensured that the range can cover most normal offsets. For example, when the offset distribution is 1 to 10 pixels, the position offset range can be set to [1,10] pixels.

[0113] Setting the range of arrangement features

[0114] The abnormal feature data also includes the arrangement features of the platens, including the arrangement spacing and alignment deviation. The arrangement spacing is obtained by calculating the distance distribution of the center points of adjacent platens, and the alignment deviation is determined by fitting a straight line of the center points and calculating its deviation from the straight line.

[0115] Arrangement spacing: Count the distribution range of the spacing between adjacent platens, and set the mean plus or minus the standard deviation as the allowable spacing range. For example, if the spacing is distributed between 15 and 25 pixels, set the arrangement spacing range to [15,25] pixels.

[0116] Alignment Deviation: Calculate the maximum deviation from the center point to the fitted line and set the allowed deviation threshold at the 95% quantile. For example, if the distribution range of alignment deviation is 0 to 3 pixels, set the threshold of alignment deviation to 3 pixels.

[0117] Setting of investment and withdrawal status judgment range

[0118] The throw and withdraw status information is represented by color components (such as the ratio of red component to green component), and the abnormal characteristic data includes the color component ratio distribution of each pressure plate. By counting the distribution range of the red component and the green component, the judgment threshold of the throw and withdraw status is dynamically set.

[0119] Red component threshold (projection status): For example, if the red component ratio is concentrated between 65% and 85%, the red component threshold is set to [65%, 85%].

[0120] Green component threshold (exit state): For example, if the green component ratio is concentrated between 70% and 90%, the green component threshold is set to [70%, 90%].

[0121] Generate dynamic decision rules

[0122] By integrating the position offset range, arrangement feature range and insertion and retraction status judgment range, a dynamic judgment rule is formed, which can comprehensively consider the position, arrangement and status characteristics, and make status judgments on the pressure plates in all detection areas, providing a unified judgment standard for subsequent detection and abnormal marking.

[0123] Preferably, the preset range of position offset and the priority of arrangement features in the correction determination rule include:

[0124] The basis for the revision of the determination rules is as follows:

[0125] Based on the extracted abnormal feature data, including position offset abnormality, arrangement feature abnormality and investment and withdrawal status abnormality; dynamically adjust the rule parameters by analyzing the distribution statistics of the abnormal feature data, including the offset range is set by the 95% quantile of the abnormal area, the arrangement feature weight is adjusted according to the abnormal frequency, and the investment and withdrawal status threshold is dynamically corrected by the color component mean;

[0126] The arrangement feature priority includes the spacing and alignment of the pressure plates; by counting the arrangement feature frequency of abnormal pressure plates, when the abnormal proportion exceeds 30%, the weight of the arrangement feature is dynamically adjusted, and the arrangement parameter range is modified at the same time.

[0127] Preferably, the preset range of investment and withdrawal state judgment in the first revised investment and withdrawal state judgment rule of the abnormal area is combined to generate an optimized dynamic judgment rule, including:

[0128] Re-correct the judgment criteria of the investment and withdrawal status according to the color component and state switching frequency of the abnormal area, including dynamically adjusting the color threshold and setting the upper limit of the state switching frequency to include 10 times / minute;

[0129] Dynamically adjust the color threshold as follows:

[0130] T 红 =μ 红 -σ 红 ,T 绿 =μ 绿 -σ 绿

[0131] Where, T 红 and T 绿 are the red and green thresholds respectively; μ 红 and μ 绿 are the means of the investment state and withdrawal state respectively; σ 红 and σ 绿are the standard deviations of the investment state and withdrawal state respectively.

[0132] The state transition frequency is used to detect whether the number of state changes of the pressure plate within a certain period of time is abnormal.

[0133] Step 4, apply the dynamic judgment rules to all detection areas, determine whether the current position and insertion and retraction status of the pressure plate meet the judgment rules, mark the pressure plates that do not meet the judgment rules, extract the area index and abnormal features corresponding to the pressure plates that do not meet the judgment rules, modify the judgment rules based on the marked abnormal features, form an updated dynamic judgment rule, and output a detection report containing the pressure plate status, area index and abnormal features.

[0134] Preferably, the optimized dynamic determination rule is applied to all detection areas, judging whether the current position and the insertion and withdrawal state of the pressing plate meet the determination rule, and marking the pressing plate that does not meet the determination rule, including:

[0135] Position offset determination:

[0136] For each platen, calculate the offset between the current position and the initial position and compare it with the optimized offset range; the offset that exceeds the range is marked as abnormal;

[0137] Arrangement feature determination:

[0138] According to the arrangement characteristics of the platens (such as spacing and alignment), determine whether they meet the optimized arrangement rules; alignment deviations exceeding the threshold are marked as abnormal;

[0139] Determination of investment and withdrawal status:

[0140] The color and switching frequency of the pressure plate's insertion and withdrawal status are analyzed to determine whether it complies with the insertion and withdrawal status rules; abnormal colors and switching frequencies are marked as abnormal.

[0141] Preferably, the extracting of the area index and abnormal features corresponding to the pressing plate that does not meet the determination rule, and the second correction of the determination rule in combination with the marked abnormal features to form an updated dynamic determination rule, includes:

[0142] In the process of rule determination, all platen features marked as abnormal are extracted, including the offset and direction beyond the offset range, which are recorded as position offset abnormalities; the arrangement center point beyond the spacing range and deviation is recorded as arrangement feature abnormalities; the color component and switching frequency that do not meet the throw-in and throw-out rules are recorded as throw-in and throw-out state abnormalities;

[0143] The extracted abnormal features are bound to the area index data of the corresponding detection area, including position binding and area binding; the position binding records the abnormal features through the index position of the pressure plate in the two-dimensional array; the area binding generates an abnormal feature list for each detection area, recording the area number and abnormal pressure plate information.

[0144] Example 2 of the present invention provides an intelligent substation pressure plate state detection system based on image recognition, including: an image acquisition module, a feature extraction module, a rule optimization module and an anomaly detection module;

[0145] The image acquisition module is used to obtain the image of the substation control cabinet, divide the detection area according to the arrangement rules of the pressure plates in the image, and record the initial position and the throw and retract status of the pressure plates in each detection area to construct a regional index containing the position and throw and retract status of the pressure plates;

[0146] A feature extraction module is used to compare the current position of the pressing plate with the initial position based on the area index, calculate the position offset, filter the detection area whose position offset exceeds the preset range and mark it as an abnormal area, and extract the boundary contour, arrangement characteristics and insertion and retraction status of the pressing plate in the abnormal area to form abnormal feature data;

[0147] A rule optimization module is used to use the abnormal feature data to correct the preset range of position offset and the priority of arrangement features in the judgment rule, and to first correct the preset range of the judgment of the throw and withdraw state in combination with the throw and withdraw state of the abnormal area to generate an optimized dynamic judgment rule;

[0148] The anomaly detection module is used to apply the optimized dynamic judgment rules to all detection areas, determine whether the current position and insertion and retraction status of the pressure plate conform to the rules, mark the pressure plate that does not conform to the rules, extract the area index and abnormal features corresponding to the pressure plate that does not conform to the rules, and revise the judgment rules for a second time in combination with the marked abnormal features to form an updated dynamic judgment rule, and output a detection report containing the pressure plate status, area index and abnormal features.

[0149] like Figure 1 As shown in the figure, the ECA attention mechanism performs global average pooling (GAP) on the input feature map χ, so that the dimension of the feature map changes from H×W×C to 1×1×C, and then processes the output feature vector with a sigmoid activation function, and then uses a one-dimensional convolution to weight the input importance to capture more important information in the input map, and obtain a feature map with channel attention. Among them, important features are assigned higher weights, and unimportant features are assigned lower weights.

[0150] The computational overhead of the ECA attention mechanism is relatively small because it only involves global average pooling and one-dimensional convolution operations without expensive matrix multiplication, which is highly efficient. This study introduces the ECA attention mechanism in the convolutional neural network to improve the accuracy and precision of the platen insertion and withdrawal state detection model.

[0151] In target detection, the loss function usually consists of three parts: classification loss, positioning loss, and confidence loss. These three parts measure the model's prediction accuracy of the target category, location, and existence, respectively. In YOLOE, both classification and confidence losses are calculated using the cross entropy loss function. The positioning loss is used to measure the model's prediction accuracy of the target box position. In YOLOE, the Smooth L1 loss function is usually used to calculate the positioning loss, which is used to measure the model's prediction accuracy of the target box position. In YOLOE, the Smooth L1 loss function is usually used to calculate the positioning loss, and its calculation formula is as follows:

[0152]

[0153] Among them, N obj is the number of objects in the image, t i,c is the real position information of the i-th target, including the center coordinates x, y and width and height w, h of the target box. is the position information of the i-th target predicted by the model. The Smooth L1 loss function is designed to balance the square loss and absolute loss. It uses the square loss when the error is small and the absolute loss when the error is large. However, the introduction of this smoothness will cause the loss function to be close to the threshold B. GT , C h Transitional changes sometimes make it difficult for the model to converge. At the same time, the performance of the Smooth L1 loss function may be affected by hyperparameters (such as thresholds), which usually need to be adjusted manually. Improperly selected hyperparameter values ​​may affect the training effect of the model.

[0154] In view of the shortcomings of the Smooth L1 loss function, the present invention introduces SIoU LOSS in the positioning loss calculation to replace the original Smooth L1 loss function. SIoU redefines the relevant loss function by introducing directionality into the loss function cost. The SIoU loss function consists of four cost functions: angle, distance, shape, and IoU.

[0155] (1) Angle cost

[0156] The SIoU loss function achieves this by adding an angle-aware component that minimizes the number of variables in the distance-related "wondering". Basically, the model will try to make predictions first on the X or Y axis (whichever is closest), and then continue to get closer along the relevant axis. To achieve this, the convergence process will first try to minimize α, if Otherwise minimize The angle cost calculation process is as follows Figure 2 shown.

[0157] The angle calculation formula is as follows:

[0158]

[0159] In the formula, certain specific values

[0160]

[0161] (2) Distance cost

[0162] The SIoU loss function redefines the distance cost by taking into account the angle cost defined above. Its calculation formula is as follows:

[0163]

[0164] where ρ x , y and γ are expressed as:

[0165]

[0166] It can be seen that when α→0, the contribution of distance cost is greatly reduced. On the contrary, when α is close to As α→0, the distance cost becomes normal. The distance cost is calculated as Figure 3 shown.

[0167] (3) Shape cost

[0168] The shape cost of the SIoU loss function is defined as

[0169]

[0170] The and θ values ​​define the shape cost, and its value is unique for each dataset. The θ value is a very important term in this equation, which controls how much attention is needed for the shape cost. If the θ value is set to 1, the shape will be optimized immediately, affecting the free movement of the shape.

[0171] (4) IoU cost

[0172] In the SIoU loss function, the IoU cost is usually used as part of the loss function to measure the accuracy of the model in locating the target. The IoU cost calculation principle is as follows: Figure 4 shown.

[0173] By calculating the above cost loss, the calculation formula of the loss function is finally defined as

[0174]

[0175] The SIoU loss function can be used to train the target detection model. By minimizing the IoU cost, the model can better learn the positioning accuracy of the platen target, thereby improving the accuracy of target detection.

[0176] The IoU cost of the SIoU loss function is a measure of the degree of overlap between two bounding boxes, which is used to evaluate the accuracy of the model in locating the target and has the characteristics of robustness, interpretability and training optimization.

[0177] In order to highlight the effectiveness of the present invention for the YOLOE model improvement method and to digitize the improvement effect, a comparative experiment was designed to prove the effect of the two research improvements of the present invention on the detection performance improvement of the YOLOE network model. The present invention proposes two improvement methods based on the YOLOvE model. The first method is to introduce the ECA attention mechanism, and the second method is to use the SIoU LOSS loss function to replace the Smooth L1 loss function. First, in order to highlight the improvement in detection performance of the improved YOLOE network, the model detection effect is listed in a table with the effects of various mainstream target detection models in the previous chapter for data comparison to reflect the improvement effect of the model. At the same time, a group ablation experiment was carried out, which is an important method for evaluating models or algorithms.

[0178] In order to gradually study the optimization of model detection performance by each improvement method and reflect the role and contribution of each improvement, we deeply analyze the internal mechanism of the model. The ECA attention mechanism and SIoU loss function are deployed to the initial YOLOE network model respectively, and the same self-built substation control cabinet pressure plate dataset is used for performance testing.

[0179] Table 1 Comparison of detection results between improved algorithm and mainstream algorithm

[0180]

[0181] In the improvement effect comparison experiment, the improved YOLOE model was trained and tested on the substation control cabinet pressure plate dataset. The detection effect was compared with the above mainstream algorithms. From the data in Table 1 above, it can be seen that the detection effect of the improved deep learning YOLOE network model has been improved, with an accuracy of 97.8%, a recall of 96.9%, and an average detection progress mAP of 90.4%, which is 0.9%, 1%, and 2.1% higher than the initial YOLOE target detection model, respectively. In addition to the improvement in accuracy and average precision, the most significant improvement is that the detection speed of the target detection model has been improved, reaching 29.86 frames per second. Compared with the original YOLOE target detection model, the detection speed has increased by 5..87 frames per second. Even compared with the faster detection speed of YOLOv7, the detection speed has still increased by 2.09 frames per second. It can be seen that the improved YOLOE has greatly improved in both detection accuracy and detection speed, which verifies the effectiveness of the improved YOLOE method studied in this design scheme. The improved YOLOE method studied in this design has achieved the detection accuracy required by the engineering project. Since the accuracy of manual inspection cannot be expressed in numerical form, this design did not compare the data with manual inspection.

[0182] In order to gradually study the optimization of model detection performance by each improvement method and reflect the role and contribution of each improvement. This design scheme deploys the ECA attention mechanism and SIoU loss function to the initial YOLOE network model respectively, and uses the same self-built substation control cabinet pressure plate dataset for performance testing. The ablation experiment data is shown in Table 2.

[0183] Table 2 Ablation experiment results

[0184]

[0185] After the horizontal comparison of the ablation experiment, it can be seen that whether it is the introduction of the ECA attention mechanism or the replacement of the SIoU loss function, the detection accuracy and detection speed of the YOLOE deep learning model can be improved. By comparing the detection accuracy of the two, it can be found that the introduction of the ECA attention mechanism has a significant improvement in the accuracy and average precision of the model detection, which are increased by 0.6 percentage points and 1.3 percentage points respectively. It can be seen that the ECA attention mechanism has a significant effect on improving the model detection performance. In contrast, although the replacement of the SIoU loss function does not bring a significant substitute in the model detection performance, it introduces directionality through the introduction, which makes the deep learning model converge faster and more accurately, and has achieved a qualitative leap in the detection speed. From the original 24.02 frames per second, it is increased to 28.18 frames per second. It can be seen that the two improved methods for the YOLOE detection model of the present invention have increased the accuracy and detection speed of the model detection in the result data comparison in many aspects, verifying the effectiveness of the improved method of this design scheme.

[0186] like Figure 5 The figure shows the detection effect of the model on the substation pressure plate sample. After careful observation, it is found that the average precision of the detection results of the YOLOE detection model is the lowest, and even the detection precision of individual pressure plates is only 0.85. The influence of personal photography on the picture is not ruled out, but the detection effect is far behind. The two improvements based on YOLOE studied in this design scheme have improved the detection effect of the model, and the final algorithm average precision has reached 0.91.

[0187] like Figure 6 As shown in the figure, this system uses a combination of front-end and back-end to complete the intelligent monitoring of the substation pressure plate status. The front-end uses WeChat as a carrier and designs a substation control cabinet management applet with a QR code scanning function to obtain the corresponding cabinet information, and take photos and upload the control cabinet pictures to the back-end monitoring.

[0188] Since the mini program and the backend are not in the same LAN, we rented an Alibaba Cloud server as an external server and used frp to penetrate the intranet. frp can expose the intranet host to the Internet through a server with a public IP, so that the intranet host can be directly accessed through the external network; frp has a server and a client. The server needs to be installed on a server with a public IP, and the client is installed on the intranet host. Through frp intranet penetration, the uploaded images and data requests used by the application mini program are passed to the local server, and then the reverse proxy settings of the server provided by Nignx are used to communicate with the mini program.

[0189] The applet communicates with the local server through the HTTP protocol. The HTTP protocol adopts a request / response model. The client sends a request message to the server, which contains the request method, URL, protocol version, request header, and request data. The server responds with a status line, which includes the protocol version, success or error code, server information, response header, and response data. Therefore, this system completes the communication between the front-end and back-end by pre-setting the response interface and data.

[0190] The software front end uses the JavaScipt framework and the WeChat developer tool platform to design and implement a smart substation management applet. WeChat applet has the advantages of low development cost, fast running speed and high user stickiness. Therefore, this system chooses WeChat applet as the front end for development.

[0191] The back-end server framework uses Python's Django application framework. The Django framework has a powerful database access component ORM, which facilitates database access and calls. At the same time, it has complete functional elements and is easy for developers to use.

[0192] The front-end user interface of the WeChat applet mainly includes the login interface, the substation control cabinet management interface, and the single inspection and patrol photo upload interface.

[0193] like Figure 7 As shown in the figure, in the user login interface, the operator enters the user name and password registered in advance on the backend server. If the information is correct, the current employee's authority level is verified to enter different area management and inspection single inspection pages. All the following interfaces are based on the highest administrator authority as an example.

[0194] By entering the correct user name and password, you can enter the homepage of the mini program, which is the function selection page. The station cabinet management button on the page corresponds to the substation control cabinet management interface, and the pressure plate detection button corresponds to the single inspection scan code photo upload interface.

[0195] Click the Station Cabinet Management button on the home page to enter the substation control cabinet management interface. On this page, you can add, delete, modify, and query substation control cabinet information. Take the substation site information interface and the control cabinet information addition interface as an example. Figure 8 shown.

[0196] As mentioned above Figures 1 to 8As shown in the figure, the complete substation control cabinet information management system includes the functions of adding, deleting, modifying and checking each site, area and electrical control cabinet of the substation. The "code" button in the site information interface generates a QR code containing the site information, which is used to provide the detected control cabinet information to the single inspection function, so as to facilitate the retrieval of the initial pressure plate throw and retraction status of the control cabinet. On the control cabinet information editing page, by entering the number of rows and columns of the control cabinet pressure plate, it is determined whether the number of pressure plate throw and retraction status in the task table exceeds the limit, so as to reduce the occurrence of errors in the initial throw and retraction status of the pressure plate entered by the staff.

[0197] like Fig. 9 As shown in the figure, the operator clicks the pressure plate detection button to enter the scanning and pressure plate image capture and upload interface. Scanning is to scan different area QR codes and control cabinet QR codes to realize the single inspection and patrol inspection functions in the demand analysis. The patrol inspection function is developed to detect the pressure plate status of all control cabinets in the entire area. Multiple pictures can be uploaded for processing at the same time. Taking the control cabinet single inspection as an example, the visualization effect of the inspection result is demonstrated.

[0198] Since this system uses manual mobile phone photography to obtain pictures of electrical control cabinets, in order to reduce the influence of external factors, this design adds two auxiliary lines to the photo upload interface to facilitate operators to take pictures and minimize the control cabinet from tilting too much when taking pictures, which will affect model recognition and result comparison, and cause incorrect sorting of the pressure plate positions. Therefore, two horizontal and vertical lines are added to facilitate the correction of the picture angle.

[0199] The inspector scans the QR code to obtain the information of the electrical control cabinet for single inspection or patrol inspection, obtains the cabinet number of the control cabinet, and retrieves the preset pressure plate throw-in and throw-out status data. Taking the control cabinet in the figure below as an example, the correct or incorrect initial throw-in and throw-out status of the pressure plate is set respectively, the current control cabinet is monitored, and the inspection results are visualized. The effect diagram is as follows Fig.10 shown.

[0200] The above visualization of the detection results shows that if the comparison result between the target detection model result and the preset pressure plate insertion and retraction status when creating the control cabinet is incorrect, the background server will use the location information to select the misplaced pressure plate in the original uploaded detection path and return the image to the WeChat applet front end to facilitate the staff to make timely changes.

[0201] First, the front end stores the QR code information, substation area information, and the set throw and unthrow status information of the pressure plate on each control cabinet in the area under normal operation. The QR code information includes the area number information of each substation area and the cabinet number information of each control cabinet in each substation area. When storing the throw and unthrow status information of the pressure plate on the control cabinet, extracting the throw and unthrow status includes based on the color of the pressure plate surface mark, red for throw and green for unthrow; using color space segmentation to extract the throw and unthrow mark area, and the extracted mark color determines the throw and unthrow status by counting the proportion of color components.

[0202] Then determine the need for single inspection or patrol inspection. When selecting the single inspection mode, scan the QR code of the current control cabinet and take a photo of the current control cabinet and upload it; when selecting the patrol inspection mode, scan the QR code of the current substation area and take photos of all control cabinets in order of cabinet numbers and upload them;

[0203] After uploading the photo of the control cabinet, the position information of each pressure plate in the current photo and the on / off status of each pressure plate are obtained, and the pressure plates are sorted according to the position information, and the pressure plates are sorted by calculating the Euclidean distance of the pressure plates; finally, the sorted pressure plate on / off status is taken out and marked with data to form a result array; the substation area where the control cabinet in the photo is located and the corresponding cabinet number are obtained through the QR code information in the current photo; the preset on / off status of the corresponding control cabinet is retrieved from the pre-stored information, and the pressure plate on / off status of the two are compared. If there is a difference in the comparison, the pressure plate status at different positions is marked in the photo and an error pressure plate annotation map is generated, and the error pressure plate annotation map is sent to the staff; if they are the same, it is prompted that the comparison is correct. The collected control cabinet pressure plate images are as follows: Fig.11 shown.

[0204] Before using the substation pressure plate dataset to test the performance of the target detection model, in order to reduce the difficulty and time of model training, it is necessary to preprocess the images in the pressure plate dataset to ensure the quality of the pressure plate images.

[0205] In the process of acquiring and transmitting the substation pressure plate image, some interference and changes caused by uncertain factors will occur. This interference and change is called image noise. These noises will affect the quality and visual effects of the acquired images. That is, noisy images will affect the training and detection of subsequent models, increasing the difficulty and time of training. Therefore, eliminating the noise contained in the image is an important task in the image preprocessing process.

[0206] Gaussian noise, salt and pepper noise, and mixed noise are three common types of noise in image processing. Gaussian noise is a type of noise caused by electromagnetic interference and other reasons during image acquisition and transmission. It will cause random fluctuations in the pixel values ​​in the image, making the image blurred and distorted. It is called Gaussian noise because the fluctuation intensity of this noise can always obey the Gaussian distribution. Salt and pepper noise will produce many random points on the image like pepper and salt. They can be white or black. They will be distributed in the image. The overall visual appearance is like the snow noise that appears on TV. This noise is sometimes also called impulse noise. Mixed noise, as the name suggests, is not a single noise, but a mixture of multiple noises. When this noise appears on the picture, due to the combined effect of multiple noises, these pictures will be more blurred than when there is only one noise distributed on them. Since these noises will affect the quality of the image and thus affect the detection of the target, it is necessary to remove the image noise. In order to remove the noise on the picture, there are usually three categories of methods, namely filter-based, model-based and learning-based methods. They each have their own characteristics and of course their own shortcomings. For pictures in different scenes, the most appropriate one should be selected based on the actual situation. Combined with the application environment of this design scheme, this design scheme needs to denoise the control cabinet dumb bar picture. The model-based method establishes a model to describe the noise and signal in the image, but its computational complexity is too high and does not conform to the actual application scenario. The learning-based method learns the statistical characteristics of noise by training a large number of noisy and noise-free images, thereby achieving noise suppression. The learning-based method can achieve better denoising effects in some cases, but it requires a lot of training data. Therefore, the filter-based method should have a better effect. Among them, median filtering, bilateral filtering and Gaussian filtering are all commonly used filtering algorithms, which can be used to filter image noise. This design scheme uses non-local mean denoising (NLMeans) to denoise the data set images.

[0207] NLMeans is an image denoising method based on local similarity. It reduces noise by finding the mean of similar regions in the image. It applies a non-local mean filter to each pixel in the image, calculates the similarity of the neighborhood around the pixel, and takes the mean of the similar regions as the new value of the central pixel. Fig.12 As shown:

[0208] By screening, sorting and denoising the collected pressure plate images, we obtained a self-built control cabinet pressure plate dataset, totaling 6,000 images, all of which have metal screens of substation electrical control cabinets as background.

[0209] Referring to the previous inspection habits of the substation, the pressure plate's activation and deactivation status is divided into three types: activated, not activated and removed.

[0210] like Fig.13 As shown, the pressure plate throw-in and throw-out state diagram. According to the above three pressure plate throw-in and throw-out states, the substation pressure plate data set is sorted and the data set is labeled. This unit selected 6,000 pictures as the data set. In order to use it for training, it is necessary to use the labeling tool labelme to manually label the data set to obtain the corresponding label for each picture. When labeling, you need to draw a frame around the detected object, and pay attention to the size and position. After the labeling is completed, you need to select the storage location of the data set pictures and labels to facilitate your own search and the call of the program. This design scheme selects the YOLO format when labeling, such as Fig.14 The following is a method for labeling a dataset. First, draw the labeling box, then select the label format. Then labelme will automatically generate a txt file with the labeled target information based on the drawn box. The txt file contains the target category and coordinate information.

[0211] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0212] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting the state of a pressure plate in an intelligent substation based on image recognition, characterized in that: The following steps are involved: Step 1: Obtain an image of the substation control cabinet, and divide the image into multiple detection areas according to the arrangement rules of the pressure plates. Each detection area contains pressure plates that are physically adjacent or functionally related. The area index is constructed based on the initial position and the insertion and withdrawal status of the pressure plates in each detection area. Step 2: Based on the area index, determine the position offset between the current position and the initial position of the pressing plate in each detection area, select the detection area whose position offset exceeds the preset range and mark it as an abnormal area, extract the boundary contour, arrangement characteristics and insertion and withdrawal status of the pressing plate in the abnormal area, and form abnormal feature data; Step 3, setting the judgment rule by analyzing the abnormal feature data, wherein the position offset range in the judgment rule is set by using the position offset of the pressure plate in the abnormal area, the arrangement feature range in the judgment rule is set by the data distribution of the arrangement feature, and the range of the judgment of the insertion and withdrawal state is set in combination with the data of the insertion and withdrawal state, so as to generate a dynamic judgment rule; Step 4, apply the dynamic judgment rules to all detection areas, determine whether the current position and insertion and retraction status of the pressure plate meet the judgment rules, mark the pressure plates that do not meet the judgment rules, extract the area index and abnormal features corresponding to the pressure plates that do not meet the judgment rules, modify the judgment rules based on the marked abnormal features, form an updated dynamic judgment rule, and output a detection report containing the pressure plate status, area index and abnormal features.

2. According to the method for detecting the state of a pressure plate in an intelligent substation based on image recognition according to claim 1, it is characterized in that: The image is divided into a plurality of detection areas according to the arrangement rules of the pressure plates, each detection area contains pressure plates that are physically adjacent or functionally related, including: Physically adjacent platens are divided as follows: For one-dimensional linearly arranged pressure plates, the physical positions are determined to be adjacent based on whether the spacing between the center points of adjacent pressure plates is uniform, and the pressure plates whose continuous spacing is within the set range are divided into the same detection area; For the platens arranged in a two-dimensional matrix, the detection areas are divided into rows or columns according to the alignment of the boundary lines in the horizontal or vertical direction and whether the spacing between the center points is uniform; The function-related pressure plates are divided as follows: According to the correlation of the pressure plates in the circuit control logic, the pressure plates involved in the same functional module are classified into the same detection area; Combined with the substation control cabinet design specifications, the pressure plates are logically divided according to functional areas so that the pressure plates in the detection area are functionally consistent; The division principles include: If the number of platens in the detection area exceeds the maximum processing unit, it will be re-subdivided according to the physical location adjacent rule; If there is a conflict between physical location and functional zoning, priority shall be given to dividing the inspection area by functionally related pressure plates.

3. According to the method for detecting the state of a pressure plate in an intelligent substation based on image recognition according to claim 1, it is characterized in that: The construction includes a regional index of the platen position and the insertion and withdrawal status, including: The region index is stored as a two-dimensional array of key-value pairs: The two-dimensional array includes rows and columns corresponding to the position of the pressure plate, and each unit records the pixel coordinates, the insertion and withdrawal status and the boundary information of the pressure plate; The key-value pair includes the area number as the key, and stores the number, coordinates, insertion and retraction status and boundary information of the pressure plate in the area.

4. According to the method for detecting the state of a pressure plate in an intelligent substation based on image recognition according to claim 1, it is characterized in that: The method of determining the position offset between the current position and the initial position of the pressure plate in each detection area based on the area index, screening the detection areas whose position offset exceeds the preset range and marking them as abnormal areas, includes: The edge detection is used to obtain the boundary contour of the platen. The geometric center of the platen in the image is calculated based on the contour pixels as the current position. The current position is compared with the center coordinates of the initial position, and the quantized offset is calculated as follows: In the formula, x, y, x i and i are the pixel coordinates of the boundary points respectively; N is the total number of pixels; the initial coordinates of the pressure plate (x0, y0); The preset ranges are adjusted based on the deviation statistics of abnormal platens: The distribution data of the platen offset in the detection area is statistically analyzed; the offset range is dynamically set according to the 95% quantile of the offset distribution.

5. The method for detecting the state of a pressure plate in an intelligent substation based on image recognition according to claim 1 or 4, characterized in that: The extraction of the boundary contour, arrangement characteristics and insertion and withdrawal status of the pressing plate in the abnormal area to form abnormal characteristic data includes: The edge detection is used to obtain the boundary contour of the pressing plate, and the features of the boundary contour are quantified, including the boundary length, the number of corner points and the closure; the boundary length is the total number of pixels of the contour line segment, which is used to determine whether the pressing plate has defects; the number of corner points is obtained by the Harris corner point detection algorithm to obtain the number of feature points on the boundary, which is used to verify the integrity of the boundary; the closure is used to determine whether the contour forms a closed curve, and the distance between the first and last boundary points is calculated for detection; Extracting arrangement features includes determining whether the arrangement complies with the rules by calculating the spacing between adjacent platens and alignment deviations; the spacing between platens is calculated as follows: In the formula, x i+1 and i+1 are the pixel coordinates of the boundary points respectively; Fit the straight line of all the center points of the platens and calculate the average deviation from the point to the straight line. If the deviation exceeds the threshold, it is judged as misalignment; Extracting the insertion and withdrawal status includes extracting the insertion and withdrawal status based on the color of the mark on the surface of the pressure plate, red for insertion and green for withdrawal; using color space segmentation to extract the insertion and withdrawal mark area, and the extracted mark color determines the insertion and withdrawal status by counting the proportion of the color components.

6. The method for detecting the state of a pressure plate in an intelligent substation based on image recognition according to claim 1 is characterized in that: The step of setting the determination rules by analyzing the abnormal characteristic data includes: The basis for the revision of the determination rules is as follows: Based on the extracted abnormal feature data, including position offset abnormality, arrangement feature abnormality and investment and withdrawal status abnormality; dynamically adjust the rule parameters by analyzing the distribution statistics of the abnormal feature data, including the offset range is set by the 95% quantile of the abnormal area, the arrangement feature weight is adjusted according to the abnormal frequency, and the investment and withdrawal status threshold is dynamically corrected by the color component mean; By counting the frequency of arrangement features of abnormal pressure plates, the weight of the arrangement features is dynamically adjusted when the abnormal proportion exceeds 30%, and the arrangement parameter range is modified at the same time.

7. The method for detecting the state of a pressure plate in an intelligent substation based on image recognition according to claim 1 is characterized in that: The combination of the investment and withdrawal status data to set the investment and withdrawal status judgment range and generate dynamic judgment rules includes: Re-correct the judgment criteria of the investment and withdrawal status according to the color component and state switching frequency of the abnormal area, including dynamically adjusting the color threshold and setting the upper limit of the state switching frequency to include 10 times / minute; Dynamically adjust the color threshold as follows: T 红 =μ 红 -s 红 ,T 绿 =μ 绿 -s 绿 Where, T 红 and T 绿 are the thresholds for red and green respectively; μ 红 and μ 绿 are the means of the investment state and withdrawal state respectively; σ 红 and σ 绿 are the standard deviations of the investment state and withdrawal state respectively. The state transition frequency is used to detect whether the number of state changes of the pressure plate within a certain period of time is abnormal.

8. The method for detecting the state of a pressure plate in an intelligent substation based on image recognition according to claim 1 is characterized in that: The dynamic determination rule is applied to all detection areas to determine whether the current position and the insertion and retraction state of the pressure plate meet the determination rule, and the pressure plate that does not meet the determination rule is marked, including: Position offset determination includes: For each platen, calculate the offset between the current position and the initial position and compare it with the optimized offset range; the offset that exceeds the range is marked as abnormal; Arrangement feature determination includes: According to the arrangement characteristics of the platen, determine whether it meets the optimized arrangement rules; the alignment deviation exceeding the threshold is marked as abnormal; The investment and withdrawal status determination includes: The color and switching frequency of the pressure plate's insertion and withdrawal status are analyzed to determine whether it complies with the insertion and withdrawal status rules; abnormal colors and switching frequencies are marked as abnormal.

9. The method for detecting the state of a pressure plate in an intelligent substation based on image recognition according to claim 1 is characterized in that: The extracting of the region index and abnormal features corresponding to the pressing plate that does not meet the determination rule, and correcting the determination rule in combination with the marked abnormal features to form an updated dynamic determination rule includes: In the process of rule determination, all platen features marked as abnormal are extracted, including the offset and direction beyond the offset range, which are recorded as position offset abnormalities; the arrangement center point beyond the spacing range and deviation is recorded as arrangement feature abnormalities; the color component and switching frequency that do not meet the throw-in and throw-out rules are recorded as throw-in and throw-out state abnormalities; The extracted abnormal features are bound to the area index data of the corresponding detection area, including position binding and area binding; the position binding records the abnormal features through the index position of the pressure plate in the two-dimensional array; the area binding generates an abnormal feature list for each detection area, recording the area number and abnormal pressure plate information.

10. Intelligent substation plate status detection system based on image recognition, including: Image acquisition module, feature extraction module, rule optimization module and anomaly detection module; characterized by: The image acquisition module is used to obtain the image of the substation control cabinet, divide the detection area according to the arrangement rules of the pressure plates in the image, and record the initial position and the throw and retract status of the pressure plates in each detection area to construct a regional index containing the position and throw and retract status of the pressure plates; A feature extraction module is used to compare the current position of the pressing plate with the initial position based on the area index, calculate the position offset, filter the detection area whose position offset exceeds the preset range and mark it as an abnormal area, and extract the boundary contour, arrangement characteristics and insertion and retraction status of the pressing plate in the abnormal area to form abnormal feature data; A rule optimization module is used to use the abnormal feature data to correct the preset range of position offset and the priority of arrangement features in the judgment rule, and to first correct the preset range of the judgment of the throw and withdraw state in combination with the throw and withdraw state of the abnormal area to generate an optimized dynamic judgment rule; The anomaly detection module is used to apply the optimized dynamic judgment rules to all detection areas, determine whether the current position and insertion and retraction status of the pressure plate conform to the rules, mark the pressure plate that does not conform to the rules, extract the area index and abnormal features corresponding to the pressure plate that does not conform to the rules, and revise the judgment rules for a second time in combination with the marked abnormal features to form an updated dynamic judgment rule, and output a detection report containing the pressure plate status, area index and abnormal features.

Images