A method for measuring the proximity of a worker to electricity suitable for complex work scenarios

Abstract

This invention discloses a method for measuring the proximity induction of workers in complex work scenarios, comprising the following steps: S1, establishing a complex work scenario recognizer based on MR-CNN; S2, acquiring images of complex work scenarios on-site and inputting them into the complex work scenario recognizer obtained in step S1 for scene feature recognition, obtaining scene feature recognition results; S3, calculating the electric field strength based on the image charge method according to the scene feature recognition results obtained in step S2, obtaining the electric field strength, and calculating the electric field strength error and the time consumed in the electric field strength calculation, respectively, obtaining the electric field strength error and the time consumed in the electric field strength calculation; S4, optimizing the electric field strength, electric field strength error, and electric field strength calculation time obtained in step S3 based on the simulated annealing algorithm, thereby obtaining the proximity induction measurement results of the electric field strength.

Description

Technical Field

[0001] This invention belongs to the field of power grid maintenance technology, specifically relating to a method for measuring the proximity induction of workers in complex work scenarios. Background Technology

[0002] With the rapid development of modern society, economy, and technology, all industries are becoming increasingly reliant on electricity, leading to a surge in demand for power and a continuous expansion of the power grid. Simultaneously, the workload of operation and maintenance is increasing daily. At power operation and maintenance sites, workers frequently need to approach live equipment. Without appropriate safety precautions, there is a risk of electric shock, potentially causing serious consequences such as injury or death. Therefore, the proximity sensing devices worn by workers need to accurately and quickly measure the ambient electric field.

[0003] However, power operation and maintenance sites are complex, featuring diverse terrains such as hills, valleys, plains, and cities, as well as varying types and numbers of charged objects, posing a significant challenge to accurate measurement of proximity induction. Numerous studies and experiments have shown that complex terrain significantly impacts the distribution of power frequency electric fields. Traditional proximity induction measurement methods primarily target ideal flat terrain, neglecting the influence of complex terrain on the power frequency electric field. Applying these methods to power frequency electric field calculations in complex terrain environments makes it difficult to accurately identify the terrain features and the number, type, and spatial location of charged objects, leading to substantial errors and unreliable calculation results. Furthermore, traditional electric field distribution characteristic analysis based on the image charge method suffers from low analysis speed and susceptibility to local optima during optimization, resulting in low proximity induction measurement accuracy. This fails to meet the high-precision proximity induction measurement requirements of complex operation scenarios and compromises the safety of operation and maintenance personnel. Therefore, there is an urgent need to design an accurate proximity induction measurement method for personnel in complex operation scenarios. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for measuring the proximity induction of workers in complex work scenarios.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for measuring the proximity induction of workers in complex work scenarios includes the following steps:

[0007] S1. Establish a complex operation scene recognizer based on MR-CNN;

[0008] S2. Collect images of complex work scenarios on site and input them into the complex work scenario recognizer obtained in step S1 for scene feature recognition to obtain scene feature recognition results;

[0009] S3. Based on the scene feature recognition results obtained in step S2, establish a three-dimensional model of the complex operation scene, obtain the number of discrete point charges and discrete line charges, calculate the electric field strength based on the number of discrete point charges and discrete line charges, and calculate the electric field strength error and the time consumed by the electric field strength calculation respectively.

[0010] S4. Based on the simulated annealing algorithm, optimize the number of discrete point charges and discrete line charges obtained in step S3, so as to minimize the weighted sum of the electric field strength error and the electric field strength calculation time obtained in step S3, and finally obtain the optimal electric field strength near-electro-induction measurement result.

[0011] Preferably, step S1 specifically includes the following steps:

[0012] S11, Image Feature Extraction

[0013] Feature maps are obtained by processing image data of complex work scenarios using the MR-CNN algorithm;

[0014] S12, Target Box Regression

[0015] The task scene part in the feature map obtained in step S11 is manually annotated to obtain the target box. The target box regression method is used to obtain the candidate box that is close to the target box. The candidate box results are mapped onto the feature map, and the area inside the candidate box is the region of interest.

[0016] S13, Classification Training

[0017] The region of interest obtained in step S12 is fed into the fully connected layer, and classification training is performed using preset labels. The function between the feature map and the image category is fitted, and the target box regression is performed to obtain a complex task scene recognizer based on MR-CNN.

[0018] Preferably, step S3 specifically includes the following steps:

[0019] S31. Based on the scene feature recognition results obtained in step S2, establish a three-dimensional model of the complex operation scene;

[0020] S32. Divide the charged wires in the three-dimensional model obtained in step S31 into multiple line segments that meet the conditions according to the cross-sectional area and charge of the wires. Use the multiple line segments as discrete line charge units, determine the number of discrete line charges, and set discrete point charges in a regular hexagonal honeycomb pattern in the non-mirror ground. Determine the number of discrete point charges. Establish the discrete equation of the three-dimensional electric field integral mathematical model based on the number of discrete line charges and the number of discrete point charges. Use the discrete equation to calculate the electric field intensity at a certain point P, and obtain the electric field intensity at a certain point P.

[0021] S33. The electric field strength error is calculated by solving the sum of the known field strength and the calculated field strength difference at all matching points on the field boundary. The electric field strength error is obtained, and the time consumed by the electric field strength calculation is calculated to obtain the electric field strength calculation time.

[0022] Preferably, step S4 specifically includes the following steps:

[0023] S41. Let T = T0, where T0 represents the initial temperature at which annealing begins, to generate an initial solution, which is the number of discrete line charges obtained in step S3. and the number of discrete point charges according to and The electric field strength calculation error Er and the electric field strength calculation time T obtained in step S3 are obtained. sum Then, the electric field strength calculation error Er and the electric field strength calculation time T are obtained. sum The weighted sum yields the cost function.

[0024]

[0025] S42. Let T = kT0, where k∈[0,1] is the rate of temperature decrease;

[0026] S43. For the current solution and Applying a random perturbation, i.e., increasing or decreasing the number of line and point charges, produces a new solution in its neighborhood. and And calculate the new cost function value.

[0027] S44, Calculation If ΔO K,H >0, accept the new solution as the current solution; otherwise, follow the probability e. -ΔE / kT Determine whether to accept the new solution;

[0028] S45. Repeat the perturbation and reception process L times at temperature T, that is, repeat steps S43 and S44.

[0029] S46. Determine whether the temperature has reached the termination temperature level and satisfies the maximum constraint on the field strength calculation error and the maximum constraint on the calculation time. If so, terminate the algorithm. At this time, the electric field strength error and the electric field strength calculation time obtained based on the temperature are the optimal electric field strength near-induction measurement results. Otherwise, return to step S42.

[0030] Compared with the prior art, the advantages of this invention are as follows:

[0031] (1) The design of the complex operation scene recognizer based on MR-CNN in this invention can perform image feature extraction, target box regression and classification training. It can use the trained mask convolutional neural network to achieve accurate recognition of the terrain and landforms, as well as the number, type and spatial location of charged bodies in complex operation scenes, thereby improving the accuracy of near-field induction measurement.

[0032] (2) Based on the obtained three-dimensional model of the complex operation scenario, the present invention uses the mirror charge method based on the simulated annealing algorithm to analyze the electric field distribution characteristics, and uses the simulated annealing algorithm to optimize the number of charges in the mirror charge method, so that the overhead function converges to the minimum and the prediction error is lower than the threshold. This can solve the problem that the traditional algorithm has a slow convergence speed and is prone to getting trapped in local optimal solutions, realize high-precision real-time calculation of near-field induction, and improve the practicality of near-field induction measurement for operators in complex operation scenarios. Attached Figure Description

[0033] Figure 1 A schematic diagram illustrating the proximity induction measurement method for workers in complex work scenarios provided by this invention;

[0034] Figure 2 This is a flowchart illustrating the optimization of proximity induction measurement results using the mirror charge method based on the simulated annealing algorithm in this invention. Detailed Implementation

[0035] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0036] This invention provides a method for measuring the proximity induction of workers in complex work scenarios, specifically including the following steps:

[0037] S1. Establish a complex task scene recognizer based on MR-CNN, which includes the following steps:

[0038] S11, Image Feature Extraction

[0039] The MR-CNN algorithm is used to process image data of complex work scenes to obtain feature maps, specifically:

[0040] Complex work scene image data is input into a neural network. The MR-CNN algorithm divides it into three channels: red, green, and blue. A pre-sized convolutional kernel slides across each channel image with a stride S and performs convolution processing. The convolution process is Q... j =b+∑W i Xi,j Q j W represents the value of the j-th pixel in the feature map corresponding to a channel. i X represents the weight of the i-th element in the convolution kernel. i,j is the value of the i-th pixel after the j-th sliding of the convolution kernel in one channel of the original image; b is the bias, with one bias corresponding to one convolution kernel.

[0041] During convolution, the kernel size may not match the image region size. Therefore, rows and columns with zero pixels need to be added around the original image to enlarge the image size. This operation is called zero-padding, and the output is a primary feature map. The size of the output primary feature map is calculated as follows:

[0042]

[0043] Where: W2, H2, and D2 are the width, height, and depth of the output feature map, respectively; W1 and H1 are the width and height of the original input map, respectively; F is the kernel size; C is the number of rows and columns with zero padding; and S is the number of kernels.

[0044] The primary feature map obtained above is activated by a nonlinear activation function, typically a modified linear unit (MRL), expressed as R(x) = max(0,x), where x is the pixel value of the primary feature map. Activation reduces poorly correlated features. After activation, pooling is performed on the feature map, i.e., non-maximum suppression is applied to a specified region in the feature map, gradually reducing the spatial size of the feature map, decreasing the number of network parameters, effectively controlling overfitting, and thus obtaining a feature map that meets the requirements. Generally, the deeper the neural network, the better the fit between the input and output. However, traditional CNNs suffer from gradient vanishing at greater depths, leading to deterioration in network performance. MR-CNN effectively solves this gradient vanishing problem. Its main idea is to use the identity mapping F(x) = x to ensure that the fitting effect of deeper layers is at least no weaker than that of shallower layers. The output of the original network is changed from F(x) to H(x) = F(x) + x, ensuring that the output is not weaker than the input, allowing the network to learn new features based on the input features.

[0045] Taking a wire image as an example, the shallowest layer is the original image containing the wire. In traditional feature extraction, the network performs convolutions from shallow to deep, using only the deepest feature map for prediction. This method, which only focuses on high-dimensional information, cannot detect small objects, i.e., it cannot accurately identify scenes in complex operational systems. MR-CNN incorporates a deep-to-shallow sampling process and lateral connections, fusing multi-scale features, which is more conducive to mining multi-scale information in complex operational scenarios.

[0046] S12, Target Box Regression

[0047] The task scene part in the feature map obtained in step S11 is compared with the complex task scene annotated by humans to obtain the target box. The target box regression method is used to obtain the candidate box that is close to the target box. The candidate box results are mapped onto the feature map, and the area inside the candidate box is the region of interest.

[0048] Target bounding box regression aims to find a transformation f that transforms the original anchor box into a candidate box that is closer to the manually labeled target box containing the work scene, i.e., f(P). x ,P y ,P w ,P h )=(G x ',G y ',G' w G h '), where P x and P y These are the x and y coordinates of the center of the anchor frame before the transformation, respectively, P w and P h To change the width and height of the anchor frame, G' x and G' y These are the x and y coordinates of the center of the anchor frame after transformation, respectively, G' w and G' h The width and height of the transformed anchor box are defined as follows. A linear transformation is applied when the original anchor box approaches the manually annotated target box containing the work scene. The transformed box becomes a candidate box. Specifically, it is first translated and then scaled, with the offset and scaling amounts as follows:

[0049]

[0050] Among them, E x and E y d is the offset. x (P) and d y (P) represents the translation scale, E w and E h d is the scaling factor. w (P) and d h (P) represents the scaling factor. MR-CNN uses gradient descent to converge the target box regression loss function, thereby achieving target box regression and mapping the result onto the feature map. The area inside the candidate box is the region of interest.

[0051] S13, Classification Training

[0052] In MR-CNN, the candidate boxes are only divided into charged bodies and terrain features, without further subdivision into which type of charged body or terrain feature. Therefore, the normalized region of interest obtained in step S12 is fed into the fully connected layer and classified and trained using preset labels (including different scenes such as transformers, conductors and other charged bodies, transformer stations and substations). The function between the feature map and the image category is fitted in the same way as in step S11, and the target box regression is performed in the same way as in step S12 to obtain a complex operation scene recognizer based on MR-CNN. This complex operation scene recognizer can predict the image classification and image target boxes, and finally generate a target mask, i.e. the contour edge of the target.

[0053] S2. Collect images of complex work scenarios on site using the miniature camera on the safety helmet, and input them into the complex work scenario recognizer obtained in step S1 for scene feature recognition, and obtain scene feature recognition results. The scene feature recognition results specifically include features such as the spatial location, type and quantity of charged bodies, as well as topographical data such as hills, valleys, plains and cities.

[0054] S3. Based on the scene feature recognition results obtained in step S2, calculate the electric field strength using the image charge method to obtain the electric field strength. Then, calculate the electric field strength error and the time consumed in the calculation, respectively. This includes the following steps:

[0055] S31. Simplify the scene feature recognition results obtained in step S2. The principle is to model complex scene objects as a combination of multiple planes with certain angles to represent them. On this basis, establish a three-dimensional model of the complex operation scene.

[0056] S32. Divide the charged wires in the 3D model obtained in step S31 into multiple sufficiently small line segments based on the cross-sectional area and charge of the wires. Use these sufficiently small line segments as discrete line charge units to determine the number of discrete line charges. Set discrete point charges in a regular hexagonal honeycomb pattern within the non-mirror ground and determine the number of discrete point charges. Based on the number of discrete line charges and discrete point charges, establish the discrete equations of the 3D electric field integral mathematical model. The field sources include wire charges and charges inside the non-mirror ground. The equations also consider the influence of terrain and the spatial position of charged bodies. Calculate the electric field intensity at a point P(x,y,z) using the discrete equations. The calculation formula is shown below:

[0057]

[0058] Where H and K are the number of discrete point charges and discrete line charges, respectively; L is the length of the discrete line charge; l is the region where the line charge exists; (x, y, z) are the coordinates of the point to be determined; (x...y...z) are the coordinates of the point to be determined. i,y i ,z i (x) represents the coordinates of the line source, and (x) represents the coordinates of the line source. j ,y j ,z j Let P be the coordinates of the point source, and q be the charge of the discrete point charge. Then the effective value of the electric field intensity at point P is:

[0059]

[0060] S33. The electric field strength error is calculated by summing the known and calculated field strength differences at all matching points on the field boundary, resulting in the electric field strength error Er. The specific formula for calculating the electric field strength error Er is as follows:

[0061]

[0062] Among them, E i Let E be the electric field of all charges at the i-th matching point. i0 The known electric field strength at the i-th matching point;

[0063] The time consumed in calculating the electric field strength is then calculated to obtain the total time T for calculating the electric field strength. sum The calculation formula is:

[0064]

[0065] Among them, T i and T il These represent the unit time spent calculating the effect of discrete point charges on the target point and the unit time spent calculating the effect of discrete line charges on the target point, respectively.

[0066] S4. The optimization objective of this invention is to minimize the measurement overhead function O by optimizing the number of discrete point charges and discrete line charges. K,H =Er+ωT sum Where ω∈[0,1] are weighting factors, the optimization problem needs to satisfy the following constraints:

[0067] (1) Maximum field strength error constraint, that is, the field strength error caused by the number of discrete point charges and discrete line charges selected must be less than the maximum field strength error in order to ensure the accuracy of the calculation results.

[0068] (2) Maximum computation time constraint, that is, the total computation time must be less than the maximum computation time to ensure the real-time performance of the computation results.

[0069] To solve this problem, this invention optimizes the electric field strength, electric field strength error, and electric field strength calculation time obtained in step S3 based on the simulated annealing algorithm, thereby obtaining the near-induction measurement result of the electric field strength. Specifically, it includes the following steps:

[0070] S41. Let T = T0, where T0 represents the initial temperature at which annealing begins, to generate an initial solution, which is the number of discrete line charges obtained in step S3. and the number of discrete point charges according to and The electric field strength calculation error Er and the electric field strength calculation time T obtained in step S3 are obtained. sum Then, the electric field strength calculation error Er and the electric field strength calculation time T are obtained. sum The weighted sum yields the cost function.

[0071]

[0072] S42. Let T = kT0, where k∈[0,1] is the rate of temperature decrease;

[0073] S43. For the current solution and Applying a random perturbation, i.e., increasing or decreasing the number of line and point charges, produces a new solution in its neighborhood. and And calculate the new cost function value.

[0074] S44, Calculation If ΔO K,H >0, accept the new solution as the current solution; otherwise, follow the probability e. -ΔE / kT Determine whether to accept the new solution;

[0075] S45. Repeat the perturbation and reception process L times at temperature T, that is, repeat steps S43 and S44.

[0076] S46. Determine whether the temperature has reached the termination temperature level and satisfies the maximum constraint on the electric field strength calculation error and the maximum constraint on the calculation time. If so, terminate the algorithm. At this time, the electric field strength error and the electric field strength calculation time obtained based on the temperature are the optimal electric field strength near-induction measurement results. Otherwise, return to step S42. The maximum constraint on the electric field strength calculation error and the maximum constraint on the calculation time are in accordance with the maximum constraint on the electric field strength calculation error and the maximum constraint on the calculation time of the corresponding voltage level in the GB 26859-2011 standard.

[0077] In summary, the proximity induction measurement method for workers in complex work scenarios provided by the embodiments of the present invention can solve the problems of slow convergence speed and easy getting trapped in local optima of traditional algorithms, realize high-precision real-time calculation of proximity induction, and improve the practicality of proximity induction measurement for workers in complex work scenarios.

[0078] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for measuring the proximity induction of workers in complex work scenarios, characterized in that, Includes the following steps: S1. Establish a complex task scene recognizer based on MR-CNN. S2. Collect images of complex work scenarios on site and input them into the complex work scenario recognizer obtained in step S1 for scene feature recognition to obtain scene feature recognition results; S3. Based on the scene feature recognition results obtained in step S2, establish a three-dimensional model of the complex operation scene, obtain the number of discrete point charges and discrete line charges, calculate the electric field strength based on the number of discrete point charges and discrete line charges, and calculate the electric field strength error and the time consumed by the electric field strength calculation respectively. S4. Based on the simulated annealing algorithm, optimize the number of discrete point charges and discrete line charges obtained in step S3, so as to minimize the weighted sum of the electric field strength error and the electric field strength calculation time obtained in step S3, and finally obtain the optimal electric field strength near-electro-induction measurement result.

2. The method for measuring the proximity induction of workers in complex work scenarios according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11, Image Feature Extraction Feature maps are obtained by processing image data of complex work scenarios using the MR-CNN algorithm; S12, Target Box Regression The task scene part in the feature map obtained in step S11 is manually annotated to obtain the target box. The target box regression method is used to obtain the candidate box that is close to the target box. The candidate box results are mapped onto the feature map, and the area inside the candidate box is the region of interest. S13, Classification Training The region of interest obtained in step S12 is fed into the fully connected layer, and classification training is performed using preset labels. The function between the feature map and the image category is fitted, and the target box regression is performed to obtain a complex task scene recognizer based on MR-CNN.

3. The method for measuring the proximity induction of workers in complex work scenarios according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31. Based on the scene feature recognition results obtained in step S2, establish a three-dimensional model of the complex operation scene; S32. Divide the charged wires in the three-dimensional model obtained in step S31 into multiple line segments that meet the conditions according to the cross-sectional area and charge of the wires. Use the multiple line segments as discrete line charge units, determine the number of discrete line charges, and set discrete point charges in a regular hexagonal honeycomb pattern in the non-mirror ground. Determine the number of discrete point charges. Establish the discrete equation of the three-dimensional electric field integral mathematical model based on the number of discrete line charges and the number of discrete point charges. Use the discrete equation to calculate the electric field intensity at a certain point P, and obtain the electric field intensity at a certain point P. S33. The electric field strength error is calculated by solving the sum of the known field strength and the calculated field strength difference at all matching points on the field boundary. The electric field strength error is obtained, and the time consumed by the electric field strength calculation is calculated to obtain the electric field strength calculation time.

4. The method for measuring the proximity induction of workers in complex work scenarios according to claim 1, characterized in that, Step S4 specifically includes the following steps: S41. Let T = T0, where T0 represents the initial temperature at which annealing begins, to generate an initial solution, which is the number of discrete line charges obtained in step S3. and the number of discrete point charges according to and The electric field strength calculation error Er and the electric field strength calculation time T obtained in step S3 are obtained. sum Then, the electric field strength calculation error Er and the electric field strength calculation time T are obtained. sum The weighted sum yields the cost function. S42. Let T = kT0, where k∈[0,1] is the rate of temperature decrease; S43. For the current solution and Applying a random perturbation, i.e., increasing or decreasing the number of line and point charges, produces a new solution in its neighborhood. and And calculate the new cost function value. S44, Calculation If ΔO K,H >0, accept the new solution as the current solution; otherwise, follow the probability e. -ΔE / kT Determine whether to accept the new solution; S45. Repeat the perturbation and reception process L times at temperature T, that is, repeat steps S43 and S44. S46. Determine whether the temperature has reached the termination temperature level and satisfies the maximum constraint on the electric field strength calculation error and the maximum constraint on the calculation time. If so, terminate the algorithm. At this time, the electric field strength error and the electric field strength calculation time obtained based on the temperature are the optimal electric field strength near-induction measurement results. Otherwise, return to step S42.

Images