Intelligent substation inspection method based on intelligent identification and edge calculation

Abstract

The invention relates to an intelligent substation inspection method based on intelligent identification and edge calculation. The method comprises the following steps: step 1, preprocessing a substation scene image; 2, transformer substation intelligent robot inspection under scene recognition: inputting scene features into multiple reverse convolutional layers in a convolutional neural network, obtaining an inspection target activation function, and completing inspection; 3, a relocation method based on a database; 4, transformer substation field operation risk perception based on edge calculation: acquiring various data of transformer substation field operation, obtaining risk data, performing risk assessment, and perceiving the transformer substation field operation risk in real time; step 5, transformer substation inspection period calculation based on reliability analysis: obtaining an inspection period of a possible fault point through a function relationship between task reliability and inspection time; the method has the advantages of realizing rapid recovery positioning and real-time risk perception based on scene accurate identification and database repositioning.

Description

Technical Field

[0001] The present invention belongs to the technical field of substations, and specifically relates to a substation intelligent inspection method based on intelligent identification and edge computing. Background Art

[0002] Substations are an important part of the power grid. Their operating status determines whether the power grid can supply power normally and safely. At present, large-scale smart grid power transportation systems follow the traditional manual inspection mode, which has limitations such as low work efficiency, strong subjectivity, and environmental factors that have a great impact on the inspection quality. Therefore, substations have gradually used intelligent robots to inspect equipment functions. Since the intelligent robots are always moving during the inspection, the dynamic sequence or continuous images collected are affected by external light and weather, and the inspection targets are difficult to identify. Image preprocessing is required to accurately complete the inspection work. However, the current robot inspection cannot intelligently adjust the devices that need to be focused on according to the scene, resulting in missed inspections; In addition, due to the complex environment of substations, current inspection robots are unable to quickly locate themselves and are unable to solve the problem of global positioning loss and recovery. During the robot inspection process, it is necessary to effectively perceive the risks of on-site operations in substations, but the current method has low data processing efficiency, and when the amount of data increases rapidly, it takes too long and cannot be solved quickly. Moreover, regular inspections of substations are generally determined by each unit based on actual conditions and historical experience, which is theoretically insufficient and cannot guarantee the accuracy and efficiency of inspections. Therefore, it is very necessary to provide a substation intelligent inspection method based on intelligent identification and edge computing that is based on scene precise identification, database relocation, rapid recovery of positioning, and real-time perception of risks. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the prior art and to provide a substation intelligent inspection method based on intelligent identification and edge computing, which is based on accurate scene identification, database relocation, rapid recovery positioning, and real-time risk perception.

[0004] The object of the present invention is achieved by: a substation intelligent inspection method based on intelligent identification and edge computing, the method comprising the following steps:

[0005] Step 1: Substation scene image preprocessing: grayscale, brightness balance, image contrast improvement, and redundant noise removal are used to solve the problems of unclear and distorted images and improve the quality of the image to be identified;

[0006] Step 2: Substation intelligent robot inspection under scene recognition: Input the scene features into the multi-layer reverse convolution layer in the convolutional neural network, obtain the inspection target activation function, and complete the inspection;

[0007] Step 3: Relocation method based on database: Use the database to store the positioning values. When the positioning fails to match, obtain the positioning values stored in the database to initialize the particles, so as to achieve rapid recovery of positioning;

[0008] Step 4: Substation on-site operation risk perception based on edge computing: Collect various data of substation on-site operations, obtain risk data, conduct risk assessment, and perceive the risks of substation on-site operations in real time;

[0009] Step 5: Calculation of substation inspection cycle based on reliability analysis: Obtain the inspection cycle of possible fault points through the functional relationship between task reliability and inspection time.

[0010] The preprocessing of the substation scene image in Step 1 is specifically as follows:

[0011] Step 1.1: Describe the color information in the robot patrol image in RGB, and use weighted average to quantify the color information into pixel gray values: , where is the gray value after the weighted average processing of the image; is the pixel point in the image;

[0012] Step 1.2: Use the histogram equalization method to solve the problems of uneven image color, overexposure or too dark color. First, determine the gray level of the original image, and calculate the corresponding pixel points on this basis to obtain the gray histogram as: , where is the number of pixels with different gray levels in the image; is the gray level corresponding to the pixel;

[0013] Step 1.3: Use the transformation relationship to correct the image gray value, and statistically make the image brightness uniform in terms of the number of pixels; then enhance the image contrast through differentiation, use the gradient algorithm to achieve the differential transformation, make the target feature elements clearer, and set the image function as , then The gradient vector in the image coordinates is: G [ f ( u , v ) ] = [ ∂ f ∂ u , ( ∂ v ∂ u ) 2 ] ( 3 ) , from the calculation result of the above formula, it can be seen that The amplitude of the change in the gradient directly affects the size of the amplitude unit of the image function. If it is a digital image, the differential calculation can be used to obtain: G [ f ( u , v ) ] = | f ( u , v ) − f ( u + 1 , v ) | ( 4 ) ;

[0014] Step 1.4: Use the linear filtering algorithm to remove redundant noise. Assume that contains pixels, and the filtered image is , that is: , where takes a positive integer value; is the set of pixel points in the set.

[0015] The substation intelligent robot patrol under scene recognition in step 2 includes the following steps:

[0016] Step 2.1: Scene feature extraction: The processed image is decomposed by wavelet transform method at multiple scales. The decomposed image is classified into two parts: high frequency and low frequency. The mean difference and variance of contrast are used to describe the average intensity contrast and gray fluctuation difference between the target area and the background area, clarify the details of the image in the vertical and horizontal directions, and suppress background noise.

[0017] Step 2.2: Fault recognition within the scene based on convolutional neural network: Use the convolutional neural network for feature classification and recognition. When the convolutional neural network is running, a convolutional layer needs to be connected in front of the input layer to remove the image that does not contain target feature information, and the differentiation structure is used to obtain the final effective image to be recognized.

[0018] When extracting the target features in the scene feature extraction in step 2.1, the directional characteristics in the horizontal direction of the scene and the directional characteristics in the vertical direction of the scene are constructed into a rectangular sliding window with spatial significance; is the target area defined on the current pixel : T = { ( x 0 + i , y 0 + j ) | i ∈ [ − w , w ] , j ∈ [ − h , h ] } ( 6 ) , according to the above formula, the size of the target area is , in this area, and are set to values greater than the size of the detection target, and are the background contrasts of the current image pixels in two directions, and we can get: B ( H ) = { ( x 0 + i , y 0 + j ) | i ∈ [ − b , − w ] ∪ [ w , b ] , j ∈ [ − h , h ] } ( 7 ) , B ( V ) = { ( x 0 + i , y 0 + j ) | i ∈ [ − w , − w ] , j ∈ [ − b , − h ] ∪ [ h , b ] } ( 8 ) , where the size of the background area is twice that of the target background; according to the value results of the above formula, after defining the rectangular sliding window in the horizontal and vertical directions, the mean difference and variance features of the target contrast in the two directions can be extracted respectively to realize the feature extraction of the target image.

[0019] The fault recognition within the scene based on convolutional neural network in step 2.2 is specifically as follows:

[0020] Step 2.21: Divide the convolution operation into 2 types, namely continuous type and discrete type. The continuous type convolution calculation formula is: , and the discrete type convolution calculation formula is: , where is the height of the input inspection image; is the number of continuous convolutional and discrete neurons; is the weight of the next neuron; is the number of channels occupied by the image;

[0021] Step 2.22: The discrete convolutional network uses linear operations, and the corresponding convolutional kernel can also be called a filter. There are multiple feature variables in the previous layer of the network, and the similarity between different feature variables is comprehensively calculated , and the input value of the activation function can be obtained, and the output map can be obtained. There are many convolutional feature quantities in any map, and the expression is: , where in the formula, is at the layer and the th feature map; and are the activation function and the input image set in this formula respectively;

[0022] Step 2.23: If the next layer after the convolutional layer is named , then according to the error backpropagation algorithm, the neuron signal value can be known. During the training process, the error signal is used for upsampling operation to facilitate obtaining a more accurate error value: , where in the formula, is the weight gain coefficient; is the mapping deviation of the layer and the th feature map; is the weight result in the upsampling layer;

[0023] Step 2.24: Any layer in the network can be connected, so the Kronecker product can be used to obtain: , where in the formula, is the upsampling;

[0024] Step 2.25: Using the error signal image, sum to calculate all target information in the scene image, and the deviation gradient is obtained as: , where in the formula, is the output and the training sample, and their training iteration combination is ;

[0025] Step 2.26: Calculate the difference in the kernel function weights in the network model through the backpropagation algorithm, and the sum of the gradients of all weight values is: .

[0026] The method for relocating based on the database in step 3 is specifically as follows: The robot loses its position in state , and the coordinate value is , due to the frequency limit of the positioning status information release, the status of positioning loss cannot be obtained in a timely manner, and the robot will continue to move forward along the angle until the navigation module obtains the positioning loss status. At this time, the robot has traveled to the coordinate values , and satisfy: , where is the linear velocity value when the robot's positioning is lost; is the time from when the robot's positioning is lost to when the navigation module receives the positioning loss status; Let the coordinate value after speed compensation be , which satisfies the following formula: , so that then the positioning can be restored.

[0027] The risk perception of substation on-site operations based on edge computing in step 4 is specifically as follows:

[0028] Step 4.1: Build a risk perception architecture for substation on-site operations based on edge computing: The perception layer of the inspection robot uses instruments such as electronic theodolites, digital rebound hammers, and laser rangefinders to collect substation operation data in real time. The sensing devices at the substation site are connected to the edge computing platform through wireless conversion and communication aggregation to realize the transmission of on-site operation data. The edge device identifies the on-site operation data of the substation, builds an edge computing model, solves the risk data in the on-site operation data of the substation, and transmits it to the cloud platform;

[0029] Step 4.2: Obtain risk data of substation on-site operations based on the edge computing model;

[0030] Step 4.3: Risk perception of substation on-site operations: Establish a risk assessment index system for substation on-site operations and conduct risk assessment of substation on-site operations.

[0031] The obtaining of risk data of substation on-site operations based on the edge computing model in step 4.2 is specifically as follows: There are several network nodes in the edge computing network, and use to represent the risk data information vector of substation on-site operations: , where represents the activation program for starting the operation of the risk data information of substation on-site operations; are the first, second, and third regression parameters respectively;

[0032] The risk data model of the substation on-site operation solved by the server is: , where is the occurrence time of various risk data during substation on-site operations; are respectively at the substation The edge computing devices and cloud computing devices on a network branch communicate with the cloud computing center The th latency; is the time for data fusion completion between the edge device and the cloud server; is the time required for data fusion; the edge computing network contains edge devices;

[0033] The duration of risk data generated during on-site operations at the substation is: , where is the time for the cloud calculator and network nodes to perform device settings; The th edge device's data reception time; is the time for the th edge computing device on the th network branch at the substation to perform device settings.

[0034] The calculation of the substation inspection cycle based on reliability analysis in step 5 is specifically as follows:

[0035] Step 5.1: Intelligent substation reliability FTA analysis: The minimal cut sets obtained through the fault tree model are process bus fault, line protection fault, transformer protection fault, bus protection fault, circuit breaker fault, current transformer fault, voltage transformer fault, communication link fault, line fault, bus fault, transformer fault, and merging unit fault;

[0036] Step 5.2: Reliability analysis of possible fault points;

[0037] Step 5.3: Inspection cycle calculation: Calculate the task reliability of possible fault points after assigning weights : , where represents the availability; represents the unavailability; represents the failure rate; Considering the influence of main equipment such as transformers and circuit breakers on the inspection cycle and based on the operation data of unmanned intelligent substations in a certain area, a substation task reliability function is proposed: .

[0038] The reliability analysis of possible fault points in step 5.2 includes the following steps:

[0039] Step 5.21: Calculation of reliability data for possible fault points: Since the failure of any device connected to a possible fault point will cause a fault to occur, the available rate calculation formula for possible fault points can be obtained: , where represents the availability of possible fault points; represents the th device connected to the possible fault point; The calculation formula for the unavailability of possible fault points is: , where represents the unavailability of possible fault points; represents the th device connected to the possible fault point; The calculation formula for the failure rate of possible fault points is: , where represents the failure rate of possible fault points; represents the th device connected to the possible fault point;

[0040] Step 5.22: Determine the weights of reliability data by the analytic hierarchy process: The analytic hierarchy process is used to determine the weights of availability, unavailability, and failure rate. After weighted calculation, reliability data is obtained; The expert group judges and scores according to the nine-point scale method to obtain the elements in the judgment matrix. The judgment matrices of the three reliability indicators of availability, unavailability, and failure rate are shown as follows: I = [ 1 9 8 1 / 9 1 1 / 2 1 / 8 2 1 ] ( 43 ) .

[0041] Advantages of the present invention: The present invention is a substation intelligent inspection method based on intelligent recognition and edge computing. In use, the substation intelligent robot inspection based on scene recognition in the present invention grayscales, equalizes brightness, enhances contrast, and denoises the collected images, making the features of the targets to be recognized in the images in the current scene clearer. The wavelet transform method is used to decompose the images to extract target features, and finally the features are input into the convolutional neural network mathematical model for training to obtain the final inspection results; The repositioning method based on the database in the present invention uses the database to store positioning values. When the positioning fails to match, the positioning values stored in the database are obtained to initialize the particles, thereby achieving fast recovery of positioning; The substation on-site operation risk perception method based on edge computing in the present invention obtains on-site operation data through the edge computing platform, finds out the risk data among them, and uses the risk data to construct an on-site operation risk evaluation index system, improving the accuracy of substation on-site operation risk perception and ensuring the safe and stable progress of substation on-site operations; The substation inspection cycle calculation method based on reliability analysis in the present invention obtains the inspection cycle of possible fault points through the functional relationship between reliability and inspection time, meeting the actual needs of inspection cycle calculation and having important significance for ensuring the safe and stable operation of intelligent substations; The present invention has the advantages of accurate scene recognition, repositioning based on the database, rapid recovery of positioning, and real-time risk perception. Description of the Drawings

[0042] Figure 1 This is the convolutional neural network framework diagram of the present invention.

[0043] Figure 2 This is the training flow chart of the present invention.

[0044] Figure 3 This is the flowchart of the relocalization algorithm of the present invention.

[0045] Figure 4 This is the schematic diagram of positioning loss of the present invention.

[0046] Figure 5 This is the schematic diagram of the architecture principle of the risk perception method of the present invention.

[0047] Figure 6 This is the architecture diagram of the edge computing network of the present invention.

[0048] Figure 7 This is the schematic diagram of the risk assessment index system for on-site operations in a substation of the present invention.

[0049] Figure 8 This is the schematic diagram of the fault tree of the intelligent substation of the present invention. Detailed implementation manners

[0050] The present invention will be further described below with reference to the accompanying drawings.

[0051] Embodiment 1

[0052] As Figure 1-8 shown, a substation intelligent inspection method based on intelligent recognition and edge computing, the method comprising the following steps:

[0053] Step 1: Substation scene image preprocessing: Using grayscale conversion, brightness equalization, image contrast enhancement, and removal of redundant noise to solve problems such as unclear and distorted parts of the image, and improve the quality of the image to be recognized;

[0054] In the present invention, the color information in the robot patrol image is described by RGB (red, green, blue), and the color information is quantified into pixel grayscale values using weighted average, and the expression is: , where is the grayscale value of the image after weighted average processing; is the pixel point in the image.

[0055] Due to factors such as the external light environment and camera exposure, the image color is uneven, overexposed or too dark in color, increasing the difficulty of target recognition; for such problems, the present invention adopts the histogram equalization method. First, determine the grayscale level of the original image, and on this basis, calculate the corresponding pixel points, and the grayscale histogram can be obtained as: , where is the number of pixels with different gray levels in the image; is the gray level corresponding to the pixel; for the convenience of calculation, the gray value of the image is corrected using the transformation relationship, and the pixel quantity is statistically counted to equalize the image brightness; subsequently, the image contrast is enhanced by differentiation. The present invention uses a gradient algorithm to achieve the differential transformation, making the target feature elements clearer.

[0056] Set the image function to , then In the image coordinates the gradient vector is: G [ f ( u , v ) ] = [ ∂ f ∂ u , ( ∂ v ∂ u ) 2 ] ( 3 ) , it can be seen from the calculation result of the above formula that the amplitude change on the gradient directly affects the size of the amplitude unit of the image function. If it is a digital image, the difference calculation can be used to obtain: G [ f ( u , v ) ] = | f ( u , v ) − f ( u + 1 , v ) | ( 4 ) .

[0057] Use the linear filtering algorithm to remove redundant noise. Assume that there are pixels in it, and the filtered image is , that is: , where takes positive integer values; is the set of pixel points in the set, but the set does not contain the central target point .

[0058] Step 2: Substation intelligent robot patrol under scene recognition: Input the scene features into the multi-layer deconvolution layer in the convolutional neural network to obtain the inspection target activation function and complete the patrol;

[0059] In the present invention, ① Scene feature extraction: The processed image is decomposed by the wavelet transform method at multiple scales. The decomposed image is classified into two parts: high frequency and low frequency. The mean difference and variance of the contrast are used to describe the average intensity contrast and the gray scale fluctuation difference between the target area and the background area, clarify the details of the image in the vertical and horizontal directions, and suppress the background noise.

[0060] When extracting its target features, in order to provide the most reliable decision-making basis for the patrol work and avoid major failures, it is necessary to construct the directional characteristics in the horizontal direction of the scene and the directional characteristics in the vertical direction of the scene into a rectangular sliding window with spatial significance; is the target area defined on the current pixel , and the expression is: T = { ( x 0 + i , y 0 + j ) | i ∈ [ − w , w ] , j ∈ [ − h , h ] } ( 6 ) , it can be known from the above formula that the size of the target area is , in this area, and are set to values greater than the size of the detection target, and are the background contrasts of the current image pixel in two directions, and we can get: B ( H ) = { ( x 0 + i , y 0 + j ) | i ∈ [ − b , − w ] ∪ [ w , b ] , j ∈ [ − h , h ] } ( 7 ) , B ( V ) = { ( x 0 + i , y 0 + j ) | i ∈ [ − w , − w ] , j ∈ [ − b , − h ] ∪ [ h , b ] } ( 8 ) , where the size of the background area is twice that of the target background; according to the value results of the above formula, after defining rectangular sliding windows in the horizontal and vertical directions, the mean difference and variance features of the target contrast in the two directions can be extracted respectively to achieve feature extraction of the target image.

[0061] ② Fault recognition in the scene based on convolutional neural network: Use a convolutional neural network for classification and recognition of features. When the convolutional neural network is running, a convolutional layer needs to be connected in front of the input layer to remove the image that does not contain target feature information, and the final effective image to be recognized is obtained through the differentiation structure. The network processing structure and process are as Figure 1 shown; the convolutional operation can be divided into two types, namely continuous type and discrete type. The continuous type convolutional calculation formula is: , and the discrete type convolutional calculation formula is: , where is the height of the input inspection image; is the number of neurons in the continuous convolution and discrete; is the weight of the next neuron; is the number of channels occupied by the image.

[0062] The convolutional network type of the present invention belongs to the discrete type and uses linear operations. The corresponding convolutional kernel can also be called a filter; the convolutional kernel has the ability to determine the target area size and convolutional recognition, is a linear change process; is the activation function; there are multiple feature variables in the previous layer of the network, and by comprehensively calculating the similarity between different feature variables , the input value of the activation function can be obtained, and the output map can be obtained. There are many convolutional feature quantities in any image, and the expression is: , where is the th feature map in the and are the activation function and the input image set in this formula respectively.

[0063] If the next layer after the convolutional layer is named , then according to the backpropagation algorithm, the neuron signal values can be known. During the training process, upsampling operations are performed using the error signal to facilitate obtaining more accurate error values: , where is the weight gain coefficient; is at the layer and the th feature map mapping deviation; is the weight result in the upsampling layer; from the calculation result of the above formula, it can be known that any layer in the network can be connected. Then, using the Kronecker product, we can get: , where is upsampling; from the above formula, it can be known that by using the error signal image, all target information in the scene image is summed and calculated to obtain the deviation gradient, and the expression is: , where is the output and the training sample, and their training iteration combination is ; by calculating the difference in the kernel function weights in the network model through the backpropagation algorithm, the sum of the gradients of all weight values is: , and all information in the scene of the robot patrol image is calculated by the above formula, and the positions of dangerous instruments exceeding the safe range are marked, providing an effective data basis for subsequent decision-making and maintenance.

[0064] The present invention uses a convolutional neural network to more accurately identify and judge the possible fault risks in a substation. The specific model training and identification classification are as Figure 2 shown, and the detailed process is as follows:

[0065] a. Preprocess the scene image, extract the target features contained in the image through grayscale conversion, brightness equalization, contrast enhancement, and denoising processing;

[0066] b. Input the images in the historical database and the extracted features into the neural network model for training;

[0067] c. Input the image containing the target features into the convolutional layer to further identify whether there is "abnormality" in the image;

[0068] d. If there is an abnormality, an alarm is issued; otherwise, continue the patrol.

[0069] In summary, the present invention uses grayscale conversion, brightness equalization, contrast enhancement, and denoising techniques to make the information features in the image more prominent; uses wavelet transform for multi-scale decomposition to extract the substation equipment features in turn; finally, through linear operations, the deviation gradient is obtained using the error signal image, and the backpropagation algorithm is used to calculate the weight difference of the kernel function in the network model to complete the patrol.

[0070] Step 3: Database-based relocalization method: The database stores the localization values. When the localization fails to match, the localization values stored in the database are obtained to initialize the particles, thus achieving rapid recovery of localization;

[0071] In the present invention, ① AMCL algorithm: The motion state equation and observation equation of the mobile robot system can be expressed as: , , where, represents the motion state of the system at time; represents the observation information of the system at time; represent the motion state transition function and the observation function of the system respectively; represent the control noise and the observation noise of the system respectively.

[0072] The mobile robot localization problem can be described as estimating the posterior probability density of the robot motion state by obtaining the observation information of the sensor . The MCL algorithm completes the estimation of the robot motion state through the following 5 steps.

[0073] (1) Prediction: According to the motion model of the inspection robot (Equation (16)) and the probability density at time, estimate the probability of the occurrence of the robot motion state at time:

[0074] (2) Update: Use the observation data of the lidar sensor at time to correct to obtain the posterior probability : , where the normalization constant: .

[0075] (3) Importance sampling: Both Equation (18) - Equation (20) contain integral terms. For the inspection robot system, it is difficult to obtain the analytical solution of the posterior probability. Monte Carlo sampling can be used to replace the calculation of the posterior probability; from the proposal distribution , collect particles to form a particle set , and then recursively calculate the weight of each particle according to the following formula: , where the particle weight satisfies the normalization condition: , where, is the total number of particles contained in the particle set.

[0076] (4) Resampling: Importance sampling can lead to the problem of particle degeneracy. After multiple iterations, the effective particles in the state space decrease sharply, resulting in a decline in estimation performance. Resampling can be used to suppress particle degeneracy. The idea of resampling is to suppress or eliminate particles with small weights and replicate particles with large weights according to the weight size. By resampling the particle set, a new particle set is generated. .

[0077] (5) State estimation: The state estimation of the robot is obtained by solving the expected value of the current state: E [ f ( x t ) ] ≈ ∑ i = 1 N W t ( x t ( i ) ) f ( x t ( i ) ) ∑ i = 1 N W t ( x t ( i ) ) ( 23 ) , where , ; On the basis of the MCL algorithm, the Adaptive Monte Carlo Localization (AMCL) algorithm introduces random sampling and KLD sampling in the resampling process. During the resampling process, random sampling adds particles with a probability, which can be obtained by the following formula: , where are the average decay rates of the exponential filters for estimating the long term and short term respectively.

[0078] ② Problems of the AMCL algorithm: (1) The AMCL algorithm accepts lidar and odometer data. If the change in odometer is greater than the threshold , the localization value is updated; When differs greatly from the actual position change of the robot , not updating will lead to a large error in localization; (2) During the resampling process, when the particles converge locally, if is large at this time, it will lead to mutation, and the algorithm will no longer be able to converge to the optimal value; (3) When the robot's inspection speed is too fast, the update speed of the localization state cannot match the running speed, resulting in the loss of localization.

[0079] ③ Improvements to the AMCL algorithm: For problems (1) and (2), the following improvements are made: Regularly perform dynamic particle scattering, that is, when the time change is equal to the set period , obtain , , such that: E [ f ( x t ) ] = ∑ i = 1 N x t ( i ) w t ( i ) ( 25 ) , where E [ f ( x t ) ] is the expected value of the pose of the particle; Update the particle pose information, such that: E [ f ( x t ) ] = ∑ i = 1 N x t ( i ) w t ( i ) w t ( i ) = 1 2 π exp ( − { x t ( i ) − E [ f ( x t ) ] } 2 2 ) ( 26 ) ; For problem (3), the solution is: according to the actual running speed of the robot and the confidence of laser matching configuration , determine the particle scattering mode: when or (positioning is not reliable), adopt the above dynamic particle scattering method; when or , obtain particles of and , calculate E [ f ( x t ) ] , and update the pose and weight information of these particles.

[0080] ④ Re-localization algorithm: The re-localization algorithm based on the database, as shown in Figure 4 , can quickly retrieve the location after the robot loses its location, improving the running efficiency of the robot; the re-localization algorithm uses the lightweight database sqlite to store the location values. When the re-localization algorithm is run for the first time, a new database and data table need to be created. During subsequent runs, there is no need to create the database again. The data table consists of four columns of data, namely , where is used to record the time when the value of this row is added to the database; is used to represent the pose of the robot; the re-localization algorithm judges whether the positioning is accurate according to the positioning confidence given by the AMCL algorithm. When the positioning state is accurate, in order to reduce the space overhead occupied by the database, the number of stored in the database is set to 10 through parameters. When the number of values in the database is greater than 10, according to size, delete the row with the smallest value, and then store the coordinate value given by AMCL into the database.

[0081] As shown in Figure 4 , the robot loses its location in state , and the coordinate value is . Due to the frequency limit of the positioning state information release, the state of losing the location cannot be obtained in time, and the robot will continue to move forward along angle until the navigation module obtains the state of losing the location. At this time, the robot has moved to the coordinate value , and satisfy: , where, is the linear velocity value when the robot loses its location; is the time from when the robot loses its location to when the navigation module receives the state of losing the location; let the coordinate value after speed compensation be , which satisfies the following formula: , so that Then the positioning can be restored. The speed values in formulas (27) and (28) are the same, while the time cannot be accurately obtained, so it is impossible to make , and an approximate method is adopted to make , make , because the improved AMCL algorithm can dynamically scatter particles and has a certain adjustment ability. Therefore, when and , the AMCL algorithm can also restore positioning.

[0082] Step 4: Substation on-site operation risk perception based on edge computing: Collect various data of substation on-site operations, obtain risk data, conduct risk assessment, and perceive substation on-site operation risks in real time;

[0083] Step 5: Calculation of substation inspection cycle based on reliability analysis: Obtain the inspection cycle of possible fault points through the functional relationship between task reliability and inspection time.

[0084] The present invention is a substation intelligent inspection method based on intelligent recognition and edge computing. In use, the present invention adjusts the image gray value by the weighted average method, equalizes the histogram brightness, and uses the linear filtering algorithm to remove redundant noise, improves the resolution of the original image scene, extracts multiple features of the target on each sub-band image based on the comparison results of the image mean difference and variance, divides the undulation difference between the target and the background area, inputs the scene features into the multi-layer deconvolution layer in the convolutional neural network, and uses the discrete linear operation mode to obtain the activation function of the inspection target, completing the substation inspection; the repositioning algorithm based on the database of the present invention, the repositioning method based on the database can enable the inspection robot to quickly restore positioning after the positioning is lost and can work effectively outdoors; the present invention has the advantages of accurate scene recognition, repositioning based on the database, rapid restoration of positioning, and real-time risk perception.

[0085] Embodiment 2

[0086] As Figure 1-8 shown, a substation intelligent inspection method based on intelligent recognition and edge computing, the method includes the following steps:

[0087] Step 1: Preprocessing of substation scene images: Use grayscale conversion, brightness equalization, improvement of image contrast, and removal of redundant noise to solve the problems of unclear and distorted parts of some images, and improve the quality of the images to be recognized;

[0088] Step 2: Inspection of substation intelligent robots under scene recognition: Input the scene features into the multi-layer deconvolution layer in the convolutional neural network, obtain the activation function of the inspection target, and complete the inspection;

[0089] Step 3: Database-based relocation method: The positioning values are stored in a database. When the positioning fails to match, the positioning values stored in the database are obtained to initialize the particles, thereby achieving rapid recovery of positioning.

[0090] Step 4: Substation on-site operation risk perception based on edge computing: Collect various data of substation on-site operations, obtain risk data, conduct risk assessment, and perceive the risks of substation on-site operations in real time.

[0091] In the present invention, ① Architecture of the substation on-site operation risk perception method based on edge computing: To perceive the risks of substation on-site operations in real time, the present invention proposes a substation on-site operation risk perception method based on edge computing. Figure 5 It is the schematic diagram of the method architecture.

[0092] The perception layer uses instruments such as electronic theodolites, digital rebound hammers, and laser rangefinders to collect substation operation data in real time. The sensing devices at the substation site are connected to the edge computing platform through wireless conversion and communication aggregation to realize the transmission of on-site operation data. The underlying hardware in the edge computing platform includes computing, communication, and storage sub-modules, which serve as the hardware infrastructure for the operation of the edge computing platform; identify the substation on-site operation data through edge devices, build an edge computing model, solve the risk data in the substation on-site operation data, and transmit it to the cloud platform.

[0093] ② Obtaining substation on-site operation risk data based on the edge computing model: 1) Edge computing network model architecture: After edge computing, the data is transmitted through the interaction network and temporarily stored in the database of the cloud server to achieve efficient data transmission and real-time update of on-site operation data; edge computing needs to collect substation on-site operation risk data through several types of sensors. Therefore, in actual application, a variety of sensor devices are placed at the collection ports of substation on-site operation risk data. Figure 6 The shown is the edge computing network architecture.

[0094] 2) Calculating operation risk data based on the edge computing model: There are several network nodes in the edge computing network, and use to represent the substation on-site operation risk data information vector: , where represents the activation program for starting the operation of the substation on-site operation risk data information; are the first, second, and third regression parameters respectively.

[0095] The solution for the substation on-site operation risk data model of the server is: , where is the occurrence time of various risk data during substation on-site operations; are respectively at the substation Edge computing devices and cloud computing devices on a network branch communicate with the cloud computing center The th latency; is the time when data fusion for both the edge device and the cloud server is completed; is the time required for data fusion; The edge computing network contains edge devices. Taking the th network branch of a substation as an example, connect the th edge device to the cloud computing center. After data fusion, retrieve the occurrence time of the on-site operation risk data of the substation.

[0096] The duration of the risk data generated during on-site operations at the substation is: , where is the time for the cloud calculator and network nodes to perform device settings; is the th edge device's data reception time; is the time for the rd edge computing device on the th network branch of the substation to perform device settings.

[0097] So far, edge computing is used to solve the on-site operation risk data of the substation, so as to establish an on-site operation risk assessment index system for the substation and realize the on-site operation risk perception of the substation.

[0098] ③ Substation on-site operation risk perception: 1) Substation on-site operation risk assessment index system: Comprehensively measure the uncertain factors in the substation on-site operation process, identify risk factors for each operation step one by one, and comprehensively measure in combination with the operation accidents and accident management regulations of power enterprises over the years. The on-site operation risk elements of the substation are summarized into 5 categories: operation method, weather environment, operation time, operation personnel and related equipment; Figure 7 The detailed substation on-site operation risk assessment index system is shown as follows.

[0099] 2) Substation on-site operation risk assessment: Given that the substation on-site operation risk assessment indicators restrict each other, when using a single indicator to evaluate the substation on-site operation risk, it will have a certain impact on the accuracy of the evaluation result. Considering comprehensively, the expert ranking method is adopted, that is, voting is carried out according to expert suggestions. After multiple rounds of consultation and feedback of opinions, finally a certain weight is assigned to each indicator, and then weighted average is used to obtain the average value of the importance assignment of each indicator; The weights of each indicator are determined by the analytic hierarchy process.

[0100] 1. Determine the scale value and construct the judgment matrix: Obtain the average of the importance assignments of each indicator through expert advice. The importance scale is determined by the scores of experts, and then construct the indicator judgment matrix .

[0101] 2. Solve the weight coefficient: Determine the weight vector according to the root method , the maximum eigenvalue , respectively: , , where is the order of the evaluation matrix; is the average of the importance assignments of each indicator; are the importance scale values of the th indicator respectively.

[0102] 3. Hierarchical sorting of risk indicators and consistency verification: Let be the average random consistency, and solve the consistency ratio : , , if is lower than 0.1, then the consistency of the judgment matrix is reasonable.

[0103] After determining the weights of each indicator, define the risk levels of on-site operations in the substation, respectively represent the risk values of the operation type, operation method, operation environment, operation time, operation personnel (i.e., the patrol robot inspects and evaluates the corresponding risks of the illegal operations of the operation personnel, the experience of the person in charge, safety awareness, emergency response ability, and the qualifications of the construction unit, etc.) and related equipment; then the on-site operation risk assessment value is:[[]] , the work ticket is a written instruction allowed by the management department for on-site operations in the substation. The staff carry out on-site operations with the ticket. The patrol robot obtains the work ticket information and sets as the benchmark risk value for each single operation. During the single-day operation time, when a single work ticket contains more than one operation, select the benchmark risk value of the single operation with the greatest risk as the benchmark risk value of the operation type on this work ticket , which is:[[]] , when several operations are carried out in parallel, that is, when several teams carry out maintenance at the same time and the number of work tickets exceeds two, the risk value is:[[]] , where is the risk value when the number of work tickets exceeds two; is the total risk value of the number of operators; are the risk coefficients of the 1st, 2nd, 3rd, , th operations after removing the work ticket with the greatest risk respectively; is the risk value of the work ticket with the highest risk among several operations; The solution formula for is: where

[0104] Step 5: Calculation of the substation inspection cycle based on reliability analysis: Through the functional relationship between task reliability and inspection time, the inspection cycle of possible fault points is obtained.

[0105] In the present invention, ① FTA analysis of the reliability of intelligent substations: The fault tree model of the intelligent substation is as shown in Figure 8 . In the fault tree model, all the logic gates are "OR" gates, and the minimal cut sets can be obtained as process bus fault, line protection fault, transformer protection fault, bus protection fault, circuit breaker fault, current transformer fault, voltage transformer fault, communication link fault, line fault, bus fault, transformer fault, and merging unit fault.

[0106] ② Reliability analysis of possible fault points: 1) Calculation of reliability data of possible fault points: Since the failure of any device connected to a possible fault point will cause a fault, the calculation formula for the availability of the possible fault point can be obtained: where represents the availability of the possible fault point; represents the availability of the th device connected to the possible fault point.

[0107] The calculation formula for the unavailability of the possible fault point: where represents the unavailability of the possible fault point; represents the unavailability of the th device connected to the possible fault point.

[0108] The calculation formula for the failure rate of the possible fault point: where represents the failure rate of the possible fault point; represents the failure rate of the th device connected to the possible fault point.

[0109] 2) Determination of the weight of reliability data by the analytic hierarchy process: The analytic hierarchy process is used to determine the weights of availability, unavailability, and failure rate. After weighted calculation, reliability data is obtained; the expert group judges and scores according to the nine-scale method to obtain the element in the judgment matrix. The quantization and meaning of the items defined by the nine-scale method are shown in Table 1; in the present invention, the judgment matrix of the three reliability indexes of availability, unavailability, and failure rate is shown as follows: I = [ 1 9 8 1 / 9 1 1 / 2 1 / 8 2 1 ] ( 43 ) , due to the limited subjective understanding of experts, the constructed judgment matrix may deviate significantly from consistency, leading to problems in weight confirmation. Therefore, it is necessary to conduct a consistency test on the judgment matrix, and the steps are as follows: (1) Calculate the maximum eigenvalue of the judgment matrix ; (2) Calculate the consistency index , where represents the order of the judgment matrix; (3) If , it means that the judgment matrix meets the consistency test; (4) If , then calculate the random consistency , where is related to the order of the matrix. When , , indicates that the judgment matrix meets the consistency test.

[0110] Table 1 Nine-point scale method

[0111] scale meaning 1 Indicates that when two factors are compared, they have equal importance 3 Indicates that when two factors are compared, one is slightly more important than the other 5 Indicates that when two factors are compared, one is significantly more important than the other 7 Indicates that when two factors are compared, one is particularly more important than the other 9 Indicates that when two factors are compared, one is extremely more important than the other 2，4，6，8 The median of the above two adjacent judgments represents the transitional nature between importance judgments

[0112] ③ Inspection cycle calculation: Calculate the task reliability of possible fault points after assigning weights : , where in the formula, represents the availability; represents the unavailability; represents the failure rate.

[0113] Considering the influence of main equipment such as transformers and circuit breakers on the inspection cycle and based on the operation data of unmanned intelligent substations in a certain area, a substation task reliability function is proposed: , and the calculated result of the inspection cycle of possible fault points is . After converting it to days and rounding off the decimal points, the actual inspection cycle of possible fault points can be obtained. From the task reliability and inspection cycle of possible fault points, it can be seen that the higher the reliability requirement, the shorter the inspection cycle; and the inspection cycles of possible fault points with the same nature are the same.

[0114] The present invention relates to an intelligent inspection method for a substation based on intelligent recognition and edge computing. During use, in the inspection process of the inspection robot of the present invention, the sensing layer uses an electronic theodolite and a laser rangefinder to collect various data of on-site operations in the substation, and uses the edge computing platform to retrieve the collected on-site operation data, obtain the risk data in the on-site operation data of the substation, transmit it to the cloud platform, construct an on-site operation risk evaluation index system based on the risk data, determine the weight of each index by using the analytic hierarchy process, conduct risk assessment, and real-time sense the on-site operation risk of the substation, effectively avoiding the occurrence of on-site operation accidents; for the intelligent substation inspection cycle calculation method based on reliability analysis of the present invention, first, the minimum cut sets of equipment that cause substation system failures are obtained by fault tree analysis (FTA); then, through the logical connection between possible fault points and related equipment in a typical 3 / 2 main wiring mode substation, and various original reliability data of equipment (availability, unavailability, failure rate), the reliability data of possible fault points are calculated; then, the weight relationship of the three types of reliability data of availability, unavailability, and failure rate is obtained by using the analytic hierarchy process, and the task reliability of possible fault points is calculated through weighted calculation; finally, through the functional relationship between the task reliability and the inspection cycle, the inspection cycle of possible fault points is obtained; the present invention has the advantages of accurate recognition based on scenarios, repositioning based on the database, rapid recovery and positioning, and real-time risk perception.

Claims

1. The intelligent inspection method of substation based on intelligent identification and edge computing is characterized by: The method comprises the following steps: Step 1: Substation scene image preprocessing: grayscale, brightness balance, image contrast improvement, and redundant noise removal are used to solve the problems of unclear and distorted images and improve the quality of the image to be identified; Step 2: Substation intelligent robot inspection under scene recognition: input the scene features into the multi-layer reverse convolution layer in the convolutional neural network, obtain the inspection target activation function, and complete the inspection; Step 3: Database-based repositioning method: Use the database to store positioning values. When positioning mismatch occurs, obtain the positioning values ​​stored in the database to initialize the particles, thereby achieving rapid positioning recovery. Step 4: Risk perception of substation on-site operations based on edge computing: Collect various data of substation on-site operations, obtain risk data, conduct risk assessment, and perceive substation on-site operation risks in real time; Step 5: Calculation of substation inspection cycle based on reliability analysis: The inspection cycle of possible fault points is obtained through the functional relationship between task reliability and inspection time.

2. The substation intelligent inspection method based on intelligent identification and edge computing according to claim 1 is characterized in that: The substation scene image preprocessing in step 1 is specifically as follows: Step 1.1: Use RGB to describe the color information in the robot patrol image, and use weighted average to quantify the color information into pixel grayscale values: , where It is the gray value of the image after weighted average processing; is the pixel in the image; Step 1.2: Use the histogram equalization method to solve the problems of uneven image color, overexposure or too dark color. First determine the grayscale level of the original image, and then calculate the corresponding pixel points on this basis to obtain the grayscale histogram: , where is the number of pixels with different gray levels in the image; is the gray level corresponding to the pixel; Step 1.3: Use the transformation relationship to correct the image grayscale value and count the number of pixels to make the image brightness uniform; then enhance the image contrast through differentiation and use the gradient algorithm to achieve differential transformation to make the target feature elements clearer. Set the image function to ,So In image coordinates The gradient vector in is: , from the above calculation results, we can know that The magnitude of the gradient change directly affects the size of the image function amplitude unit. If it is a digital image, the difference calculation can be used to obtain: ; Step 1.4: Use linear filtering algorithm to remove redundant noise, assuming Existence pixels, the image after filtering is ,Right now: , where The value is a positive integer; for The collection of pixels in the collection.

3. The substation intelligent inspection method based on intelligent identification and edge computing according to claim 1, characterized in that: The intelligent robot patrol of the substation under the scene recognition in step 2 includes the following steps: Step 2.1: Scene feature extraction: The processed image is decomposed into multiple scales by wavelet transform method. The decomposed image is classified into high-frequency and low-frequency parts. The mean difference and variance of contrast are used to describe the average intensity contrast and grayscale fluctuation difference between the target area and the background area, clarify the details of the image in the vertical and horizontal directions, and suppress background noise. Step 2.2: In-scene fault recognition based on convolutional neural network: Use convolutional neural network to classify and recognize features. When the convolutional neural network is running, it is necessary to connect the convolution layer before the input layer to remove images that do not contain target feature information and differentiate the structure to obtain the final effective image to be identified.

4. The intelligent inspection method for substations based on intelligent identification and edge computing according to claim 3 is characterized in that: When extracting the target features from the scene features in step 2.1, the directional characteristics in the horizontal direction of the scene need to be Directional characteristics in the direction perpendicular to the scene Constructed as a rectangular sliding window with spatial significance; For the current pixel The target area defined on: According to the above formula, the target area size is , in this area and Set to a value larger than the size of the detected target. and is the background contrast of the current image pixel in two directions, we can get: , , where the size of the background area is twice that of the target background; according to the value of the above formula, after defining the rectangular sliding window in the horizontal and vertical directions, the mean difference and variance features of the target contrast in the two directions can be extracted respectively to realize the feature extraction of the target image.

5. The intelligent inspection method for substations based on intelligent identification and edge computing according to claim 3, characterized in that: The in-scene fault recognition based on convolutional neural network in step 2.2 is specifically as follows: Step 2.21: The convolution operation is divided into two types, namely continuous and discrete. The continuous convolution calculation formula is: , the discrete convolution calculation formula is: , where is the height of the input patrol image; is the number of neurons for continuous convolution and discrete; is the weight of the next neuron; is the number of channels occupied by the image; Step 2.22: Discrete convolutional networks use linear operations, and the corresponding convolution kernels can also be called filters. There are multiple feature variables in the front layer of the network, and the similarity between different feature variables is comprehensively calculated. , we can get the input value of the activation function and obtain the output graph. There are many convolution features in any graph, and the expression is: , where For the Layer feature maps; and are the activation function and input image set in the formula respectively; Step 2.23: If the convolutional layer The next layer is named , then according to the error back propagation algorithm, we can know the value of the neuron signal. During the training process, the error signal is used for upsampling operation to obtain a more accurate error value: , where is the weight gain coefficient; For the Tier feature map mapping deviation; is the weight result in the upsampling layer; Step 2.24: Any layer in the network can be connected, so using the Kronecker product we can get: , where For upsampling; Step 2.25: Using the error signal image, sum and calculate all target information in the scene image, and obtain the deviation gradient: , where For output and training samples, the training iterations of the two are combined as ; Step 2.26: Calculate the kernel function weight difference in the network model through the back propagation algorithm, and the gradient sum of all weight values ​​is: .

6. The intelligent inspection method for substations based on intelligent identification and edge computing according to claim 1, characterized in that: The database-based relocation method in step 3 is specifically as follows: the robot is in state Positioning loss occurs when the coordinate value is Due to the frequency limit of the positioning status information release, the robot cannot obtain the positioning loss status in time, and will continue to move along The robot moves forward until the navigation module detects that the positioning is lost. At this time, the robot has reached the coordinate value , and satisfy: ,in, It is the linear velocity value when the robot loses positioning; The time from when the robot loses its positioning to when the navigation module receives the positioning loss status; the coordinate value after speed compensation is , which satisfies the following formula: ,make Then you can restore the positioning.

7. The substation intelligent inspection method based on intelligent identification and edge computing according to claim 1, characterized in that: The substation on-site operation risk perception based on edge computing in step 4 is specifically as follows: Step 4.1: Build a substation on-site operation risk perception architecture based on edge computing: The perception layer of the inspection robot uses electronic theodolites, digital rebound instruments, and laser rangefinders to collect substation operation data in real time. The sensor equipment on the substation site is connected to the edge computing platform through wireless conversion and communication aggregation to achieve the transmission of on-site operation data. The substation on-site operation data is identified through edge devices, and an edge computing model is built to solve the risk data in the substation on-site operation data and transmit it to the cloud platform; Step 4.2: Obtaining substation field operation risk data based on edge computing model; Step 4.3: Substation on-site operation risk perception: Establish a substation on-site operation risk assessment index system and conduct substation on-site operation risk assessment.

8. The intelligent inspection method for substations based on intelligent identification and edge computing according to claim 7, characterized in that: The acquisition of substation field operation risk data based on the edge computing model in step 4.2 is specifically as follows: the edge computing network contains several network nodes, Represents the substation on-site operation risk data information vector: , where Indicates the dispatching procedure for starting the calculation of substation site operation risk data information; are the first, second and third regression parameters, respectively; Solving the risk data model of server substation field operation is: , where The occurrence time of various risk data during substation on-site operations; In the substation The edge computing devices and cloud computing devices on the network branches and the cloud computing center Data communication Second delay; The time it takes for data fusion between edge devices and cloud servers to be completed; The time required for data fusion; the edge computing network includes edge devices; Duration of risk data generated during substation on-site operations for: , where The time for cloud computers and network nodes to perform device setup; For the The time it takes for an edge device to receive data; For the substation The first The time required to set up an edge computing device.

9. The substation intelligent inspection method based on intelligent identification and edge computing according to claim 1, characterized in that: The calculation of the substation inspection cycle based on reliability analysis in step 5 is specifically as follows: Step 5.1: FTA analysis of smart substation reliability: The minimum cut set obtained through the fault tree model is process bus fault, line protection fault, transformer protection fault, bus protection fault, circuit breaker fault, current transformer fault, voltage transformer fault, communication link fault, line fault, bus fault, transformer fault, and merging unit fault; Step 5.2: Reliability analysis of possible failure points; Step 5.3: Inspection cycle calculation: Calculate the reliability of the possible fault point task after assigning weights : , where represents the availability rate; represents the unavailability rate; represents the failure rate; considering the impact of major equipment such as transformers and circuit breakers on the inspection cycle, and proposing a substation task reliability function based on the operation data of unmanned smart substations in a certain area : .

10. The intelligent inspection method for substations based on intelligent identification and edge computing according to claim 9, characterized in that: The reliability analysis of possible fault points in step 5.2 includes the following steps: Step 5.21: Calculation of reliability data of possible fault points: Since failure of any device connected to a possible fault point will cause a fault to occur, the formula for calculating the availability of a possible fault point can be obtained: , where Indicates the availability of possible fault points; Indicates the first The availability rate of each device; the calculation formula for the unavailability rate of possible fault points is: , where Indicates the unavailability rate of possible fault points; Indicates the first The unavailability rate of each device; the calculation formula for the failure rate of possible fault points is: , where Indicates the failure rate of possible fault points; Indicates the first Failure rate of each device; Step 5.22: Determine the weight of reliability data using the analytic hierarchy process: Use the analytic hierarchy process to determine the weights of availability, unavailability, and failure rates, and obtain the reliability data through weighted calculation; the expert group uses the nine-point scale method to judge and score the elements in the judgment matrix. , the judgment matrix of the three reliability indicators of availability, unavailability and failure rate is shown as follows: .

Images