High-voltage live working safety early warning method based on neural network

Abstract

The invention relates to a neural network-based high-voltage live-line work safety early warning method. The method comprises the following steps of 1, establishing a high-voltage transmission line electric field calculation model by using a surface charge method; 2, establishing an overall three-dimensional model of the transformer substation; 3, constructing a signal processing circuit of the power frequency electric field sensor, capturing electric field information in real time through the power frequency electric field sensor, and converting the electric field information into a processable electric signal; 4, real-time data from the power frequency electric field sensor are received, feature extraction and classification are carried out by adopting an early warning method based on a convolutional neural network, and potential safety hazards are identified; 5, establishing a neural network-based high-voltage live working safety early warning comprehensive platform integrating data acquisition, processing, analysis and early warning; the method has the advantages that the neural network algorithm is adopted, the distance and the azimuth angle are accurately judged, live-line work is early warned, and the accident rate is greatly reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of live-line working on power grids, and in particular relates to a safety early warning method for high-voltage live-line working based on a neural network. Background Art

[0002] At present, the voltage level of my country's power grid is constantly increasing, the coverage of the power grid is constantly expanding, and the grid structure is becoming more and more complex, which has caused prominent safety issues for power system inspections, live operations personnel and equipment, and has put forward higher requirements for safety management; with the increase in the voltage level of my country's power grid and the expansion of the power grid coverage, the safety issues of power system inspections and live operations have become increasingly prominent, especially under high voltage conditions, the increase in safety distance and the enhancement of the induced electric field around the human body, which has brought great safety hazards to operation and maintenance personnel. The greatest safety hazard to power operation and maintenance personnel is electric shock, including direct electric shock and indirect electric shock. Direct electric shock accidents can be avoided by strengthening operating standards and safety management, while for the prevention of indirect electric shock, it is necessary to maintain a certain safe distance from the live body. However, power operation and maintenance personnel will inevitably approach high-voltage equipment during work, putting them in a more dangerous environment. This is the main cause of power accidents and safety hazards. Traditional visual observation and manual operation can no longer meet the needs of modern power grid safety management. Due to the lack of effective technical means and monitoring and early warning methods, it cannot provide necessary and objective reference information for operation and maintenance personnel, resulting in defects and hidden dangers in safety risk management. Using modern science and technology to strengthen the research and development of active defense power safety devices is an urgent problem to be solved in the new stage of my country's power grid development. Therefore, it is very necessary to provide a neural network-based high-voltage live operation safety early warning method that adopts a neural network algorithm, accurately judges distance and azimuth, warns of live operations, and greatly reduces the accident rate. Summary of the Invention

[0003] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a neural network-based high-voltage live working safety warning method that adopts a neural network algorithm, accurately judges distance and azimuth, warns of live working, and greatly reduces the accident rate.

[0004] The object of the present invention is achieved by: a neural network-based high-voltage live working safety early warning method, the method comprising the following steps:

[0005] Step 1: Use the surface charge method to establish a high-voltage transmission line electric field calculation model;

[0006] Step 2: Establish the overall 3D model of the substation;

[0007] Step 3: Build a signal processing circuit for the power frequency electric field sensor, capture the electric field information in real time through the power frequency electric field sensor, and convert it into a processable electrical signal;

[0008] Step 4: Receive real-time data from the power frequency electric field sensor and use a convolutional neural network-based early warning method to extract and classify features to identify potential safety hazards;

[0009] Step 5: Establish a comprehensive safety warning platform for high-voltage live working based on neural networks that integrates data collection, processing, analysis and warning.

[0010] The step 1 of establishing a high-voltage transmission line electric field calculation model using the surface charge method specifically includes the following steps:

[0011] Step 1.1: Surface charge method;

[0012] Step 1.2: Physical model of transmission line.

[0013] The surface charge method in step 1.1 is as follows: there is a conductor in a vacuum with a surface area of ​​s1. The charge density distribution on the conductor surface around the point r' is ds. The potential at any point r in the electric field is given by the following formula: Divide the conductor surface into n small pieces ΔS i (i=1,2,...,n), let the charge density distribution σ(r′) on each small block be a constant, that is, σ(r′)=σ i (2), the surface integral on the conductor surface is the sum of the integrals on each small block. From equations (1) and (2), we can get: Then select n arbitrary matching points r on the conductor surface j (j=1,2,...,n), solve the potential boundary conditions σ i , that is, to solve the i (i=1,2,...,n) is the matrix equation system of unknown variables: In the formula, the coefficient matrix p ji for: Solving for p ji The global coordinates should be converted to local coordinates for calculation. The vector expression and normal vector of any point on the element ds in the global coordinates can be expressed in the local coordinates as follows: From formula (4), we can find σ i , according to σ i The electric field intensity component at any point p in space can be calculated, that is:

[0014] Where,

[0015] According to the superposition theorem, we can get:

[0016]

[0017] The transmission line physical model in step 1.2 includes the following steps:

[0018] Step 1.21: Transmission line physical model: Simplify the transmission line model as follows: ① Treat the power frequency electric field of the transmission line as a quasi-static field; ② Assume that the earth is a good conductor with zero potential; ③ The equivalent radius of the split conductor commonly used in cylindrical high-voltage transmission lines Where R is the split conductor radius, n is the number of secondary conductors, and r is the secondary conductor radius;

[0019] Step 1.22: Catenary equation: When the span of overhead transmission lines is short, a small sag is generally used. Therefore, a parabola equation can be used to describe this type of catenary. The catenary equation is: Where, γ is the conductor load, that is, the load per unit conductor length cross section; h0 is the shortest distance between the overhead line and the ground; σ0 is the stress at the lowest point of the conductor;

[0020] Step 1.23: Calculation of the transmission line model: Since the conductor surface is considered as a cylindrical surface, cylindrical elements can be used for surface segmentation. Taking the local coordinate uv, the θ and x coordinates of the segmented surface element are: The y and z coordinates of the subdivision surface can be expressed as: Where y0 and z0 are the y and z coordinates of the central axis; r0 is the equivalent radius of the wire; the expression of the subdivision surface element vector is: R =x(v)i+y(u)j+z(u)k,0≤u,v≤1(12), the tangent vectors t1 and t2, the normal vector n and the infinitesimal area ds at a point T on the split surface are: For any point j in the field, the potential coefficient is: According to equations (4)-(9), the field strength at any point in space can be calculated.

[0021] The step 2 of establishing the overall three-dimensional model of the substation includes the following steps:

[0022] Step 2.1: Acquire 3D data: Use 3D laser scanning technology to acquire relevant 3D data information;

[0023] Step 2.2: Data acquisition: When the 3D laser scanner measures a certain point P of an object with its own coordinate system XYZ and origin O(0,0,0), the laser transmitter in the scanner emits a laser pulse signal. After diffuse reflection from the surface of the object to be measured, the signal returns to the receiver along almost the same path. By calculating the phase difference of the reflected light, the distance S between the scanned point P and the scanner can be obtained; the controller synchronously records the observed values ​​of the horizontal scanning angle α and the vertical scanning angle θ of each laser pulse signal, and internally determines the relative coordinates P (x P ,y P ,z P )for: The 3D laser scanner can obtain the reflection intensity of the object while obtaining the spatial coordinates. The measured object has different reflectivity to the laser beam due to its surface characteristics. The reflected light intensity L P Satisfy the following formula: L P =δL0(16), where L0 is the initial light intensity emitted by the 3D laser scanner; δ is the reflectivity of different objects;

[0024] Step 2.3: 3D modeling of substation.

[0025] The three-dimensional modeling of the substation in step 2.3 includes the following steps:

[0026] Step 2.31: Noise removal: Preprocess the acquired point cloud data to remove erroneous and error points in the original point cloud, identify and classify the point cloud data, and remove the data set other than the actual valid scanned objects;

[0027] Step 2.32: Point cloud stitching: The point cloud data scanned at different observation points are transformed into the same coordinate system and unified using the iterative nearest point algorithm. Based on the point cloud data registration, it is necessary to find the nearest point of a given point set in another point set and calculate the distance to the nearest point, that is, to calculate the rotation and translation transformation matrices R and T that satisfy the following objective function, that is: f(R,T)=min∑(Rp i +Tq j ) 2 (17), where P(p i )、Q(q j ) are two selected point cloud sets, satisfying 1≤i,j≤n. When P and Q are two independently scanned point cloud data sets with related features, the transformation matrix can be obtained. After calculation, the two point cloud data sets are merged to obtain a complete new data set;

[0028] Step 2.33: Surface reconstruction: After processing a large number of discrete points in the point cloud dataset through a grid reconstruction algorithm, a polygonal model under a given accuracy constraint is obtained. The three-dimensional contour model of the substation can be obtained by fitting the point cloud. For the terrain model, an evenly spaced grid is used for fitting. After the evenly spaced grid fitting modeling, the terrain model is formed. The terrain model is combined with the substation three-dimensional model, and a standard three-dimensional model is generated from the original data to form a complete three-dimensional model data.

[0029] The signal processing circuit of the industrial frequency electric field sensor constructed in step 3 includes a signal acquisition circuit, a signal conditioning circuit, a signal amplifier circuit, a gain control circuit, a low-pass filter circuit, an analog-to-digital conversion circuit, a digital-to-analog conversion circuit, a communication module circuit and a digital signal processing circuit.

[0030] The convolutional neural network-based early warning method in step 4 includes the following steps:

[0031] Step 4.1: Establish a scoring system;

[0032] Step 4.2: Generate dataset;

[0033] Step 4.3: Build a one-dimensional convolutional neural network model and input the dataset into the model for training;

[0034] Step 4.4: Complete training and determine model structure and parameters: Use loss function and mean absolute error (MAE) to reflect the model training process. The mean absolute error is an indicator of performance measurement, and the loss function is the mean square error (MSE), as shown in the following formula: In the formula, y is the true value, that is, the expert judgment value; y′ is the predicted value; is the mean of y; m is the number of samples; the determination coefficient R2 is used to evaluate the generalization ability of the model: Use L2 regularization to limit the weight size to avoid overfitting of the model: J = J(θ) + λR(ω), Where J(θ) is the original loss function; λ is the regularization coefficient; R(ω) is the model complexity; ω is the weight.

[0035] The neural network-based high-voltage live working safety warning platform established in step 5 includes a data acquisition module, a data processing module, a user interface module and an alarm system. The alarm system includes multiple sets of measuring plate modules, a single-chip microcomputer, a power supply module, a display module and an alarm module.

[0036] The neural network-based high-voltage live working safety warning platform established in step 5 adopts a four-layer architecture including a data acquisition layer, a data processing layer, a warning analysis layer and a user interface layer.

[0037] The beneficial effects of the present invention are as follows: the present invention is a safety early warning method for high-voltage live working based on a neural network. During use, the present invention uses a surface charge method / simulated charge method to establish a high-voltage transmission line electric field calculation model; then establishes an overall three-dimensional model of the substation; and uses a power frequency electric field sensor to capture electric field information in real time and convert it into a processable electrical signal; then receives real-time data from the power frequency electric field sensor, and uses a convolutional neural network-based early warning method to extract and classify features to identify potential safety hazards; after the model training is completed, it can be used to predict the equipment orientation in live working. By inputting relevant data (such as electric field strength, voltage level, etc.), the model will output predicted orientation information. Based on the predicted orientation information and combined with a preset safety threshold, an early warning message can be generated. If the predicted orientation approaches a dangerous area or exceeds the safety threshold, the system will issue an early warning to remind the operator to pay attention to safety. According to the actual working environment and changes in data, the system's early warning algorithm and model structure are continuously optimized and improved to improve the system's early warning accuracy and reliability; the present invention has the advantages of using a neural network algorithm, accurately judging distance and azimuth, warning live working, and greatly reducing the accident rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a cross-sectional view of the conductor of the present invention.

[0039] Figure 2 It is a cross-sectional view of the cylindrical conductor of the present invention.

[0040] Figure 3 Schematic diagram of the terrestrial laser scanner system and coordinate system of the present invention.

[0041] Figure 4 This is a typical one-dimensional convolutional neural network structure diagram of the present invention.

[0042] Figure 5 A flow chart is constructed for the early warning model of the present invention.

[0043] Figure 6 This is a schematic diagram of the SMOTE algorithm of the present invention.

[0044] Figure 7 This is a structural diagram of the one-dimensional convolutional neural network model of the present invention.

[0045] Figure 8 This is a schematic diagram of the early warning model training process of the present invention.

[0046] Figure 9 Schematic diagram of the electrode structure of the present invention.

[0047] Figure 10 It is a schematic diagram of the structure of the alarm system of the present invention.

[0048] Figure 11 This is a flow chart of the system signal processing of the present invention.

[0049] Figure 12 This is a flow chart of the adaptive neural network training program of the present invention. DETAILED DESCRIPTION

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

[0051] Example 1

[0052] like Figure 1-8 As shown, a neural network-based high-voltage live working safety warning method comprises the following steps:

[0053] Step 1: Use the surface charge method to establish a high-voltage transmission line electric field calculation model;

[0054] Step 2: Establish the overall 3D model of the substation;

[0055] Step 3: Build a signal processing circuit for the power frequency electric field sensor, capture the electric field information in real time through the power frequency electric field sensor, and convert it into a processable electrical signal;

[0056] Step 4: Receive real-time data from the power frequency electric field sensor and use a convolutional neural network-based early warning method to extract and classify features to identify potential safety hazards;

[0057] Step 5: Establish a comprehensive safety warning platform for high-voltage live working based on neural networks that integrates data collection, processing, analysis and warning.

[0058] The step 1 of establishing a high-voltage transmission line electric field calculation model using the surface charge method specifically includes the following steps:

[0059] Step 1.1: Surface charge method: The basic idea of ​​the surface charge method is based on the uniqueness theorem of electric field. A set of fictitious surface charge units are used to replace the surface charge units with equal values, and the charge density of each unit is set to a constant. The superposition theorem is applied to superimpose the electric fields generated by these surface charge units in space to obtain the spatial electric field distribution generated by the original continuous distributed charge. The cross-section diagram of the conductor is shown in the figure below. Figure 1 As shown, there is a conductor in a vacuum with a surface area of ​​s1. The charge density distribution on the infinitesimal area ds around point r' on the surface of the conductor is σ(r'). Then the potential at any point r in the electric field is given by the following formula: Figure 1 In the example, the conductor surface is divided into n small pieces ΔS i (i=1,2,...,n), let the charge density distribution σ(r′) on each small block be a constant, that is, σ(r′)=σ i(2), the surface integral on the conductor surface is the sum of the integrals on each small block. From equations (1) and (2), we can get: Then select n arbitrary matching points r on the conductor surface j (j=1,2,...,n), solve the potential boundary conditions σ i , that is, to solve the i (i=1,2,...,n) is the matrix equation system of unknown variables: In the formula, the coefficient matrix p ji for: Solving for p ji The global coordinates should be converted to local coordinates for calculation. The vector expression and normal vector of any point on the element ds in the global coordinates can be expressed in the local coordinates as follows: From formula (4), we can find σ i , according to σ i The electric field intensity component at any point p in space can be calculated, that is: Where, According to the superposition theorem, we can get:

[0060]

[0061] Step 1.2: Physical model of transmission line: This includes the following steps:

[0062] Step 1.21: Transmission line physical model: Simplify the transmission line model as follows: ① Treat the power frequency electric field of the transmission line as a quasi-static field; ② Assume that the earth is a good conductor with zero potential; ③ The equivalent radius of the split conductor commonly used in cylindrical high-voltage transmission lines Where R is the split conductor radius, n is the number of secondary conductors, and r is the secondary conductor radius;

[0063] Step 1.22: Catenary equation: When calculating the 3D field, sag must be considered, so a suitable catenary equation is often needed to describe the shape of the actual transmission line. When the span of overhead transmission lines is short, a low sag method is generally used. Therefore, a parabola equation can be used to describe this type of catenary. The catenary equation is: Where, γ is the conductor load, that is, the load per unit conductor length cross section; h0 is the shortest distance between the overhead line and the ground; σ0 is the stress at the lowest point of the conductor;

[0064] Step 1.23: Calculation of the transmission line model: Since the conductor surface is considered as a cylindrical surface, Use cylindrical elements to divide the surface, such as Figure 2 As shown, take the local coordinate uv, then the θ and x coordinates at the subdivision surface are: Where θ0 is the θ coordinate of points 1 and 2; x0 is the x coordinate of points 1 and 4; θ1 is the θ coordinate of points 3 and 4; x1 is the x coordinate of points 2 and 3; the y and z coordinates at the subdivision surface can be expressed as: Where y0 and z0 are the y and z coordinates of the central axis; r0 is the equivalent radius of the wire; the expression of the subdivision surface element vector is: R =x(v)i+y(u)j+z(u)k,0≤u,v≤1(12), the tangent vectors t1 and t2, the normal vector n and the infinitesimal area ds at a point T on the split surface are: For any point j in the field, the potential coefficient is: According to equations (4)-(9), the field strength at any point in space can be calculated.

[0065] In this embodiment, the present invention can also use the simulated charge method to establish an electric field calculation model for a high-voltage transmission line. Specifically, the simulated charge method is used to establish an electric field calculation model for a high-voltage transmission line, and the electric field distribution at different heights of the high-voltage transmission line is calculated and analyzed. The COMSOL finite element simulation software is used to simulate and analyze the spatial electric field intensity distribution near the high-voltage transmission lines and power towers with single-circuit horizontal wiring forms of different voltage levels and double-circuit wiring forms on the same tower with different wiring forms. An overall three-dimensional model of the substation is established, and the electric field intensity distribution within the substation and the electric field intensity distribution curve along its patrol path are analyzed through COMSOL simulation.

[0066] The simulated charge method simplifies electric field calculations by assuming a series of discrete charges to simulate a continuous charge distribution. The electric field distribution of high-voltage transmission lines varies at different altitudes, primarily because the charge distribution and the geometry of the surrounding environment change with altitude. The simulated charge method can calculate the electric field strength and distribution at different altitudes, allowing us to understand the distribution of the electric field within the substation and possible areas of excessively high or low electric field strength. It can also analyze the electric field strength distribution curve along the line inspection path, providing safety guidance for line inspectors.

[0067] To construct an electric field calculation model for a high-voltage transmission line, we first collect relevant information such as the transmission line's geometric parameters, voltage level, and conductor material. Then, based on the basic principles of the simulated charge method, we set a large number of simulated charges on the conductors, and determine the position and charge of each simulated charge through iterative calculations. Finally, based on the interaction between the charges, we calculate the electric field strength at each point in space.

[0068] The installation location and monitoring range of the early warning device of the early warning platform can be determined according to the characteristics of the electric field distribution to ensure that the early warning device can accurately detect abnormal electric field changes. In addition, the trigger threshold and alarm mechanism of the early warning device can be designed according to the change law of the electric field strength to improve the accuracy and reliability of the early warning device. Through the application of the simulated charge method and COMSOL finite element simulation software, the electric field distribution of high-voltage transmission lines and substations can be comprehensively and accurately analyzed to ensure that the early warning device can accurately monitor and respond to abnormal electric field changes, thereby improving the safety and reliability of the power system.

[0069] The step 2 of establishing the overall three-dimensional model of the substation includes the following steps:

[0070] Step 2.1: Acquire 3D data: Use 3D laser scanning technology to acquire relevant 3D data information;

[0071] Step 2.2: Data acquisition: When the 3D laser scanner measures a certain point P of an object with its own coordinate system XYZ and origin O(0,0,0), the laser transmitter in the scanner emits a laser pulse signal. After diffuse reflection from the surface of the object to be measured, the signal returns to the receiver along almost the same path. By calculating the phase difference of the reflected light, the distance S between the scanned point P and the scanner can be obtained; the controller synchronously records the observed values ​​of the horizontal scanning angle α and the vertical scanning angle θ of each laser pulse signal, and internally determines the relative coordinates P (x P ,y P ,z P )for: The 3D laser scanner can obtain the reflection intensity of the object while obtaining the spatial coordinates. The measured object has different reflectivity to the laser beam due to its surface characteristics. The reflected light intensity L P Satisfy the following formula: L P =δL0(16), where L0 is the initial light intensity emitted by the 3D laser scanner; δ is the reflectivity of different objects; when objects of different materials, such as steel, walls, trees, rubber, etc., are measured by the rangefinder, different reflected light intensity values ​​will be formed due to their own different reflectivities δ. In later applications, the reflected light intensity values ​​can be used as a reference to preliminarily classify the point cloud, providing data support for the accurate identification of models of objects in the same scene layer, such as metal wire pipes on the wall, windows of the house, etc., and also facilitating more intelligent segmentation of the model in subsequent software processing; for the scene to be measured, its scene point cloud data is a data set composed of a series of points obtained by transforming the coordinates using the formula. Through the point data set, the spatial appearance can be redrawn in the 3D scene;

[0072] Step 2.3: 3D modeling of substation: including the following steps:

[0073] Step 2.31: Noise removal: Preprocess the acquired point cloud data to remove erroneous and error points in the original point cloud, identify and classify the point cloud data, and remove the data set other than the actual valid scanned objects;

[0074] Step 2.32: Point cloud stitching: The point cloud data scanned at different observation points are transformed into the same coordinate system and unified using the iterative nearest point algorithm. Based on the point cloud data registration, it is necessary to find the nearest point of a given point set in another point set and calculate the distance to the nearest point, that is, to calculate the rotation and translation transformation matrices R and T that satisfy the following objective function, that is: f(R,T)=min∑(Rp i +Tq j ) 2 (17), where P(p i )、Q(q j ) are two selected point cloud sets, satisfying 1≤i,j≤n. When P and Q are two independently scanned point cloud data sets with related features, the transformation matrix can be obtained. After calculation, the two point cloud data sets are merged to obtain a complete new data set;

[0075] Step 2.33: Surface reconstruction: After processing a large number of discrete points in the point cloud dataset through a grid reconstruction algorithm, a polygonal model under a given accuracy constraint is obtained. The three-dimensional contour model of the substation can be obtained by fitting the point cloud. For the terrain model, an evenly spaced grid is used for fitting. After the evenly spaced grid fitting modeling, the terrain model is formed. The terrain model is combined with the substation three-dimensional model, and a standard three-dimensional model is generated from the original data to form a complete three-dimensional model data.

[0076] In this embodiment, the establishment of an overall three-dimensional model of the substation can provide accurate spatial information for the early warning platform, enabling accurate judgment of the location and operating range of operators, thereby avoiding false alarms and missed alarms; secondly, the three-dimensional model can simulate the electric field distribution and equipment status under different operating scenarios, providing rich data support for the early warning platform; finally, the three-dimensional model can also be combined with virtual reality technology to provide operators with an immersive safety training and drill environment, thereby improving their safety awareness and operational skills.

[0077] The convolutional neural network-based early warning method in step 4 includes the following steps:

[0078] Step 4.1: Establish a scoring system and create labels for power frequency electric field sensor monitoring data. The scoring system includes a power frequency electric field sensor monitoring database. The system can randomly select a historical time t and display the historical power frequency electric field sensor monitoring data and its characteristic parameters at that time. The interval of t is 3 hours, including 0-25 for early warning danger, 25-50 for weak warning danger, 50-75 for medium warning danger, and 75-100 for strong warning danger.

[0079] Step 4.2: Generate dataset;

[0080] Step 4.3: Build a one-dimensional convolutional neural network model and input the dataset into the model for training;

[0081] In this embodiment, a typical one-dimensional convolutional neural network is as follows Figure 4 As shown in the figure, its structure mainly includes input layer, convolution layer, pooling layer, fully connected layer and output layer; the convolution layer is similar to the neurons in the feedforward neural network, which can extract features from the input data. It contains multiple convolution kernels, which perform convolution operations on the input feature vector and construct the output feature vector using a nonlinear activation function: Where l is the lth layer network; is the output of the l-th layer network; is the input of the l-th layer network; ω is the convolution kernel; b is the bias; f is the activation function; commonly used activation functions are ReLU function, sigmoid function, ReLU:f1(x)=max(0,x), The pooling layer can extract features and filter information to avoid model overfitting. Maximum pooling is commonly used and is generally placed after the convolutional layer. The fully connected layer is used to "flatten" the features, that is, to connect all feature vectors end to end to form a one-dimensional vector, and finally the output layer outputs the target result.

[0082] Step 4.4: Complete training and determine model structure and parameters: Use loss function and mean absolute error (MAE) to reflect the model training process. The mean absolute error is an indicator of performance measurement, and the loss function is the mean square error (MSE), as shown in the following formula: In the formula, y is the true value, that is, the expert judgment value; y′ is the predicted value; is the mean of y; m is the number of samples; the determination coefficient R2 is used to evaluate the generalization ability of the model: Use L2 regularization to limit the weight size to avoid overfitting of the model: J = J(θ) + λR(ω), Where J(θ) is the original loss function; λ is the regularization coefficient; R(ω) is the model complexity; ω is the weight.

[0083] The present invention is a safety warning method for high-voltage live operations based on neural networks. During use, the present invention uses the surface charge method / simulated charge method to establish a high-voltage transmission line electric field calculation model; then establishes an overall three-dimensional model of the substation; and uses the power frequency electric field sensor to capture electric field information in real time and convert it into a processable electrical signal; then receives real-time data from the power frequency electric field sensor, and uses a convolutional neural network-based warning method to perform feature extraction and classification to identify potential safety hazards; after the model training is completed, it can be used to predict the equipment orientation in live operations. By inputting relevant data (such as electric field strength, voltage level, etc.), the model will output predicted orientation information. Based on the predicted orientation information, combined with a preset safety threshold, a warning message can be generated. If the predicted orientation approaches a dangerous area or exceeds the safety threshold, the system will issue a warning to remind the operator Personnel should pay attention to safety, and continuously optimize and improve the system's early warning algorithm and model structure according to changes in the actual working environment and data, so as to improve the system's early warning accuracy and reliability; the present invention optimizes the neural network model, combines more advanced sensor technology and data processing methods, to achieve more accurate electric field measurement and faster early warning response, and reduce false alarm and missed alarm rates; the present invention is not only limited to high-voltage live working scenarios, but the early warning system can also be applied to other power facilities, equipment and working environments, such as substations, transmission lines, etc., to provide support for comprehensive safety monitoring of the power system; in addition, the present invention can also be combined with modern information technologies such as the Internet of Things and cloud computing to achieve intelligent and automated management, improve reliability and ease of use, and reduce operation and maintenance costs; the present invention has the advantages of using neural network algorithms, accurately judging distance and azimuth, warning live working, and greatly reducing accident rates.

[0084] Example 2

[0085] like Figure 9-12 As shown, a neural network-based high-voltage live working safety warning method comprises the following steps:

[0086] Step 1: Use the surface charge method to establish a high-voltage transmission line electric field calculation model;

[0087] Step 2: Establish the overall 3D model of the substation;

[0088] Step 3: Build a signal processing circuit for the power frequency electric field sensor, capture the electric field information in real time through the power frequency electric field sensor, and convert it into a processable electrical signal;

[0089] Step 4: Receive real-time data from the power frequency electric field sensor and use a convolutional neural network-based early warning method to extract and classify features to identify potential safety hazards;

[0090] Step 5: Establish a comprehensive safety warning platform for high-voltage live working based on neural networks that integrates data collection, processing, analysis and warning.

[0091] The signal processing circuit of the industrial frequency electric field sensor constructed in step 3 includes a signal acquisition circuit, a signal conditioning circuit, a signal amplifier circuit, a gain control circuit, a low-pass filter circuit, an analog-to-digital conversion circuit, a digital-to-analog conversion circuit, a communication module circuit and a digital signal processing circuit.

[0092] In this embodiment, since the signal collected directly from the sensor is often relatively weak, it needs to be amplified by a signal amplifier. During the signal processing process, for the power frequency electric field sensor, a low-pass filter needs to be designed to filter out high-frequency noise. The signal after amplification and filtering is usually an analog signal, while the neural network processes a digital signal. Therefore, the analog signal needs to be converted into a digital signal through an analog-to-digital converter (ADC). In the digital signal processing stage, the signal can be further processed and analyzed, including signal spectrum analysis, time domain analysis, feature extraction, etc. The purpose of these processing steps is to extract key features that can reflect the electric field changes from the original signal to provide input for subsequent neural network prediction. The signal features after digital signal processing are fed into a trained neural network model for prediction. The neural network model can predict future electric field change trends or potential safety risks based on historical data and current signal features. The prediction results can be presented to the user through a display, sound and light alarm, etc., so that safety measures can be taken in a timely manner. The signal processing circuit of the power frequency electric field sensor can accurately capture electric field changes and provide high-quality input data for neural network prediction through reasonable signal acquisition, amplification, filtering, analog-to-digital conversion and digital signal processing.

[0093] First, the weak signal output by the sensor is amplified by the signal amplification module to increase the signal amplitude and signal-to-noise ratio; then, the signal is filtered by the filtering module to remove noise and interference components, thereby improving the purity and reliability of the signal; finally, the analog signal is converted into a digital signal by the digitization module to facilitate subsequent digital processing and analysis;

[0094] Among them, the AD circuit is responsible for converting the analog signal output by the sensor into a digital signal for use by the subsequent digital processing circuit; the DA circuit converts the processed digital signal into an analog signal for driving the alarm device or transmitting it to the remote monitoring center; the communication module is responsible for transmitting the warning information to the monitoring center in real time, using an Ethernet-based communication solution and data transmission through the TCP / IP protocol. The communication module also has data encryption and verification functions to ensure the security and reliability of data transmission; low-pass filter: Since the frequency of the industrial frequency electric field signal is low (usually 50Hz or 60Hz), a low-pass filter needs to be designed to remove high-frequency noise and interference; notch filter: If there are interference signals in the environment that are close to the frequency of the industrial frequency signal (such as harmonics of other power supply equipment), a notch filter is required to eliminate these interferences; anti-aliasing filter: Before the ADC, an anti-aliasing filter is required to limit the bandwidth of the input signal to prevent aliasing effects.

[0095] The neural network-based high-voltage live working safety warning platform established in step 5 includes a data acquisition module, a data processing module, a user interface module and an alarm system. The alarm system includes multiple sets of measuring plate modules, a single-chip microcomputer, a power supply module, a display module and an alarm module.

[0096] In this embodiment, ① the power supply module adopts two modes of power supply: solar energy and CT power supply, with built-in energy storage battery and charge and discharge protection circuit to ensure stable and reliable power supply of the device. At the same time, the power supply voltage of different circuit parts is different, and the power supply module is needed to match the voltage; the alarm system is powered by a 3.7V lithium battery, and the charging module adopts a constant current and constant voltage linear power management chip TLC4054, which can be directly charged from the USB port. The preset charging voltage is 4.2V, the accuracy reaches ±1%, and it has an over-discharge protection function; in order to provide a stable voltage for the microcontroller and external components, a voltage stabilizing circuit is designed. In order to prevent the voltage of the signal acquisition module from being affected by other components, a separate voltage stabilizing circuit is used. Voltage stabilization: The voltage stabilization module uses the AMS1117-3.3 voltage stabilization chip with an output voltage of 3.3V. In order to improve the efficiency of voltage stabilization, the input voltage of the voltage stabilization module is first increased to 4.5V by the boost module, and then the 4.5V voltage is stabilized to 3.3V through the AMS1117-3.3; ② The measuring plate module monitors the electric field strength in real time, and obtains electric field strength data at multiple angles and positions through the plate measurement modules distributed in 5 directions. The spacing between the plates is fixed, which helps to ensure the accuracy and repeatability of the data. When the plate module is close to the wire, the voltage is small, and when the distance is far, the voltage is large. Its main function is to obtain data by real-time monitoring of the electric field strength; ③ Alarm module Built-in high-decibel speakers, stroboscopic lights, laser projection devices and other devices can promptly warn of the dangers that may be caused by being too close to high-voltage power lines. At the same time, the set values ​​can be adjusted through the display circuit and buttons. The laser projection equipment may be used to project a warning sign or image when approaching high-voltage power lines. It can project obvious warning information in specific areas (such as around high-voltage power lines) to further remind people to pay attention to safety. The main function of the alarm module is to issue a warning in time when the equipment or personnel are too close to the high-voltage power lines. The built-in sensors (such as electric field sensors) are used to detect the distance to the high-voltage power lines. When the distance is less than the safety threshold, the alarm module will activate the above-mentioned warning device. Users can set the distance to the high-voltage power lines. Safety distance threshold, or adjust the warning level of sound and light source, so that the alarm module can adapt to different working environments and user needs; ④ The single chip microcomputer can obtain the vertical distance between the device and the wire by looking up the table by measuring and storing the characteristic curve of the plate output voltage corresponding to different voltages in advance. The in-plane distance between the device and the wire can be calculated through the voltage relationship between the two plates. The spatial distance between the device and the wire can be obtained through the data of the three basic modules (plates); whether the device is in the safe working area can be calculated. When in the working area, the device works normally, but when in the danger zone, the system will issue an alarm message in time. At the same time, a communication module can be added to send data to terminals such as mobile phones;

[0097] The core of the measuring plate module is two plates. When the plates are in an electric field, a voltage will be generated between the two plates. The voltage is related to the electric field strength between the plates and the distance between the two plates. For example, Figure 9 As shown, for example, when the two plates A / B are in an electric field environment, positive charges are generated in the opposite direction of the electric field, and negative charges are generated in the direction of the electric field. The electric field generated by the high-voltage wire is divergent, with the strongest field strength at the center point. The farther away from the center, the smaller the field strength. The measuring plate module uses two plates with a certain distance, short-circuiting the negative poles of all plates. When the distance between the plates is constant, the closer the distance between the plates and the transmission line, the greater the voltage generated between the plates, and the farther the distance between the plates and the transmission line, the smaller the voltage generated between the plates.

[0098] The alarm system of the present invention adopts five groups of measuring electrode modules, which are distributed on the other surfaces of the cube device except the bottom. During measurement, the device is leveled through a mechanical mechanism, and the voltage between each group of plates is introduced into the single-chip computer system through wires. A signal limiting device protection system is added to the single-chip computer system. After the signal is input into the single-chip computer, the received signal is processed by the developed neural network algorithm to obtain the distance between the current device and the high-voltage power line, thereby issuing an alarm signal.

[0099] The neural network-based high-voltage live operation safety warning platform established in step 5 adopts a four-layer architecture including a data acquisition layer, a data processing layer, a warning analysis layer and a user interface layer; the data acquisition layer is responsible for real-time collection of data from sensors and monitoring camera equipment, including information such as electric field strength, equipment status, and personnel location; the data processing layer performs preprocessing, feature extraction, and classification and recognition operations on the collected data, and extracts key information for subsequent warning analysis; the warning analysis layer performs in-depth analysis and prediction on the processed data based on the neural network model, determines whether there are safety hazards, and generates warning signals; the user interface layer provides users with a friendly operation interface and visual display, which facilitates users to view warning information, monitor the operation site, and perform system settings.

[0100] In this embodiment, the signal processing flow of the neural network-based high-voltage live working safety warning platform is as follows: Figure 11As shown, based on the classification and characteristics of current power grid equipment (including primary and secondary equipment of substations), combined with prior knowledge, a decision tree set is established; for the real-time transmission of remote signals, the syntax and semantic recognition of signal content are introduced, the substation equipment type, equipment characteristics and alarm content are separated, and simultaneous instantaneous activation and semi-activation are determined in the decision tree set; in the fault decision process, a large amount of real-time data from different channels is classified according to prior rules: suggestions include fault, abnormality, over-limit, displacement and notification; the fault discrimination decision module determines that the source end includes: circuit breaker, knife switch, grounding knife switch, busbar, transformer, transformer winding, capacitive reactance, load, AC line segment endpoint, protection node and terminal equipment; based on the decision results of the prior fault discrimination decision set and combined with the power grid topology, the effectiveness of the predetermined fault is identified; the telesignaling signal is immediately read in, various devices are set in the power grid topology, and the load data of each conductor, busbar, transformer and so on in the telemetry signal are combined to establish a linkage decision tree set for secondary identification;

[0101] Through secondary linkage fault identification, the effectiveness of the fault can be checked at the grid topology level. For example, if a switch is disconnected due to a fault, but real-time topology analysis reveals that there is still load telemetry in the switch-associated line, secondary identification will temporarily suppress the fault and run the suspension mechanism identification again after a specific delay to provide a final fault judgment. Secondary linkage fault identification based on topology correlation can effectively improve the fault recognition rate based on prior fault identification decisions and reduce the possibility of false alarms. Based on the scheduling data of the maintenance plan, the contents of the maintenance plan and operation and maintenance plan are established as a decision set based on the Gantt chart. Based on the secondary judgment results of the topology-related linkage fault identification, the maintenance plan scheduling process is combined with the tertiary linkage identification to mark the telemetry signals of equipment actions caused by maintenance work and operation and maintenance work, identify the changing characteristics of the fault, and eliminate the probability of false alarms.

[0102] Among them, the human posture acquisition circuit: The human posture acquisition circuit can capture and record the posture information of the human body. The human posture usually includes parameters such as position, direction, speed, acceleration, etc. In order to realize the acquisition of human posture, a sensor that can measure height and three-axis acceleration is required. The height sensor: It can measure the height of the human body relative to a reference point (such as the ground), such as ultrasonic, laser ranging, air pressure sensor, etc.; the three-axis acceleration sensor can measure the acceleration of an object on three axes (usually X, Y, and Z axes) to calculate the dynamic changes of the object, such as moving speed, direction, inclination, etc. By selecting appropriate sensors to achieve height and three-axis acceleration measurement, the main purpose is to collaboratively determine the height of the person's ascent and the gradient of the height change. For example, when the person rises, the height sensor will detect the increase in height, while the three-axis acceleration sensor may detect upward acceleration. The gradient of height change refers to the rate of change of height over time, that is, the speed at which the person rises or falls. This can be obtained by differentiating or differentially calculating the height data, which can assist in the coordinated processing of alarm and electric field sensor data.

[0103] The human posture acquisition circuit can also assist in the alarm function. When the human posture reaches a preset threshold (such as exceeding a certain height or descending too quickly), the system will generate an alarm signal. At the same time, it can process the data of the electric field sensor, and the electric field sensor data can be processed in conjunction with the human posture data to provide more comprehensive and accurate monitoring and judgment.

[0104] The high-voltage live working safety warning platform based on neural network uses neural network to process information to obtain accurate distance and direction information. Specifically, it is as follows: ① Initialize all weights and thresholds of the neural network model and set the initial weights and thresholds, including: 1) Determine the network structure: First, it is necessary to determine the structure of the neural network, including the number of neurons in the input layer, hidden layer and output layer; 2) Random initialization: Use a random number generator to generate an initial value for each weight and threshold. These initial values ​​are usually randomly selected within a small range such as [-1, 1] or [0, 1]; 3) Save the initial value: Save the generated initial weight and threshold in the parameters of the neural network for use in subsequent training processes; ② Train the network, that is, the learning process of the network: obtain several sets of relationship data between the electric field azimuth distance and the voltage level through actual measurement. In order to make accurate subsequent judgments, it is necessary to measure more sets of data samples, input the data samples into the neural network, and obtain the actual output after the neural network is processed (the output is the prediction or classification result of the neural network for these data); ③ According to the error back propagation algorithm, continuously adjust the weights and calculate the node error; ④ Return to ② and continue training the network until the error meets the requirements, such as Figure 12 The neural network algorithm flowchart shown.

[0105] The present invention is a safety warning method for high-voltage live operations based on a neural network. During use, the high-voltage live operation safety warning platform based on a neural network of the present invention can monitor the distance between operation and maintenance personnel and high-voltage live equipment in real time. When the distance is less than the safety threshold / when the operating equipment or personnel enter the dangerous area, the alarm system promptly issues an audible and visual alarm to prompt the operator, thereby effectively avoiding the occurrence of electric shock accidents and ensuring the safety of workers and equipment. The present invention uses a neural network algorithm to process data, obtains accurate orientation, and gives warning information to provide safety warnings for live operations and prevent accidents. The present invention adopts a dual power supply mode of solar power supply and CT power supply to make the operation of the device more stable and reliable. The present invention uses data from 5 azimuth measurement terminals for processing to accurately judge the distance and azimuth angle of the wire from the equipment, which serves as a warning and knowledge for live operations and greatly reduces the accident rate. The present invention has the advantages of using a neural network algorithm, accurately judging distance and azimuth angle, warning live operations, and greatly reducing the accident rate.

Claims

1. A neural network-based safety early warning method for high-voltage live working, characterized by: The method comprises the following steps: Step 1: Use the surface charge method to establish a high-voltage transmission line electric field calculation model; Step 2: Establish the overall 3D model of the substation; Step 3: Build a signal processing circuit for the power frequency electric field sensor, capture the electric field information in real time through the power frequency electric field sensor, and convert it into a processable electrical signal; Step 4: Receive real-time data from the power frequency electric field sensor and use a convolutional neural network-based early warning method to extract and classify features to identify potential safety hazards; Step 5: Establish a comprehensive safety warning platform for high-voltage live working based on neural networks that integrates data collection, processing, analysis and warning.

2. The neural network-based high-voltage live working safety early warning method according to claim 1, characterized in that: The step 1 of establishing a high-voltage transmission line electric field calculation model using the surface charge method specifically includes the following steps: Step 1.1: Surface charge method; Step 1.2: Physical model of transmission line.

3. The neural network-based high-voltage live working safety early warning method according to claim 2, characterized in that: The surface charge method in step 1.1 is as follows: there is a conductor in a vacuum with a surface area of ​​s1. The charge density distribution on the conductor surface around the point r' is ds. The potential at any point r in the electric field is given by the following formula: Divide the conductor surface into n small pieces ΔS i (i=1,2,...,n), let the charge density distribution σ(r′) on each small block be a constant, that is, σ(r′)=σ i (2), the surface integral on the conductor surface is the sum of the integrals on each small block. From equations (1) and (2), we can get: Then select n arbitrary matching points r on the conductor surface j (j=1,2,...,n), solve the potential boundary conditions σ i , that is, to solve the i (i=1,2,...,n) is the matrix equation system of unknown variables: In the formula, the coefficient matrix p ji for: Solving for p ji The global coordinates should be converted to local coordinates for calculation. The vector expression and normal vector of any point on the element ds in the global coordinates can be expressed in the local coordinates as follows: From formula (4), we can find σ i , according to σ i The electric field intensity component at any point p in space can be calculated, that is: Where, According to the superposition theorem, we can get:

4. The neural network-based high-voltage live working safety early warning method according to claim 2, characterized in that: The transmission line physical model in step 1.2 includes the following steps: Step 1.21: Transmission line physical model: Simplify the transmission line model as follows: ① Treat the power frequency electric field of the transmission line as a quasi-static field; ② Assume that the earth is a good conductor with zero potential; ③ The equivalent radius of the split conductor commonly used in cylindrical high-voltage transmission lines Where R is the split conductor radius, n is the number of secondary conductors, and r is the secondary conductor radius; Step 1.22: Catenary equation: When the span of overhead transmission lines is short, a small sag is generally used. Therefore, a parabola equation can be used to describe this type of catenary. The catenary equation is: Where, γ is the conductor load, that is, the load per unit conductor length cross section; h0 is the shortest distance between the overhead line and the ground; σ0 is the stress at the lowest point of the conductor; Step 1.23: Calculation of the transmission line model: Since the conductor surface is considered as a cylindrical surface, cylindrical elements can be used for surface segmentation. Taking the local coordinate uv, the θ and x coordinates of the segmented surface element are: The y and z coordinates of the subdivision surface can be expressed as: Where y0 and z0 are the y and z coordinates of the central axis; r0 is the equivalent radius of the wire; the expression of the subdivision surface element vector is: R =x(v)i+y(u)j+z(u)k,0≤u,v≤1(12), the tangent vectors t1 and t2, the normal vector n and the infinitesimal area ds at a point T on the split surface are: For any point j in the field, the potential coefficient is: According to equations (4)-(9), the field strength at any point in space can be calculated.

5. The neural network-based high-voltage live working safety early warning method according to claim 1, characterized in that: The step 2 of establishing the overall three-dimensional model of the substation includes the following steps: Step 2.1: Acquire 3D data: Use 3D laser scanning technology to acquire relevant 3D data information; Step 2.2: Data acquisition: When the 3D laser scanner measures a certain point P of an object with its own coordinate system XYZ and origin O(0,0,0), the laser transmitter in the scanner emits a laser pulse signal. After diffuse reflection from the surface of the object to be measured, the signal returns to the receiver along almost the same path. By calculating the phase difference of the reflected light, the distance S between the scanned point P and the scanner can be obtained; the controller synchronously records the observed values ​​of the horizontal scanning angle α and the vertical scanning angle θ of each laser pulse signal, and internally determines the relative coordinates P (x P ,y P ,z P )for: The 3D laser scanner can obtain the reflection intensity of the object while obtaining the spatial coordinates. The measured object has different reflectivity to the laser beam due to its surface characteristics. The reflected light intensity L P Satisfy the following formula: L P =δL0(16), where L0 is the initial light intensity emitted by the 3D laser scanner; δ is the reflectivity of different objects; Step 2.3: 3D modeling of substation.

6. The neural network-based high-voltage live working safety early warning method according to claim 5, characterized in that: The three-dimensional modeling of the substation in step 2.3 includes the following steps: Step 2.31: Noise removal: Preprocess the acquired point cloud data to remove erroneous and error points in the original point cloud, identify and classify the point cloud data, and remove the data set other than the actual valid scanned objects; Step 2.32: Point cloud stitching: The point cloud data scanned at different observation points are transformed into the same coordinate system using an iterative nearest point algorithm. Based on the point cloud data registration, it is necessary to find the nearest point in another point set of a given point set and calculate the distance to the nearest point. In other words, the rotation and translation transformation matrices R and T that satisfy the following objective function are calculated, namely: Where, For two selected point cloud sets, satisfying 1≤i,j≤n, when P and Q are two independently scanned point cloud data sets with related features, the transformation matrix can be obtained. After calculation, the two point cloud data sets are merged to obtain a complete new data set; Step 2.33: Surface reconstruction: After processing a large number of discrete points in the point cloud dataset through a grid reconstruction algorithm, a polygonal model under a given accuracy constraint is obtained. The three-dimensional contour model of the substation can be obtained by fitting the point cloud. For the terrain model, an evenly spaced grid is used for fitting. After the evenly spaced grid fitting modeling, the terrain model is formed. The terrain model is combined with the substation three-dimensional model, and a standard three-dimensional model is generated from the original data to form a complete three-dimensional model data.

7. The neural network-based high-voltage live working safety early warning method according to claim 1, characterized in that: The signal processing circuit of the industrial frequency electric field sensor constructed in step 3 includes a signal acquisition circuit, a signal conditioning circuit, a signal amplifier circuit, a gain control circuit, a low-pass filter circuit, an analog-to-digital conversion circuit, a digital-to-analog conversion circuit, a communication module circuit and a digital signal processing circuit.

8. The neural network-based high-voltage live working safety early warning method according to claim 1, characterized in that: The convolutional neural network-based early warning method in step 4 includes the following steps: Step 4.1: Establish a scoring system; Step 4.2: Generate dataset; Step 4.3: Build a one-dimensional convolutional neural network model and input the dataset into the model for training; Step 4.4: Complete training and determine model structure and parameters: Use loss function and mean absolute error (MAE) to reflect the model training process. The mean absolute error is an indicator of performance measurement, and the loss function is the mean square error (MSE), as shown in the following formula: In the formula, y is the true value, that is, the expert judgment value; y′ is the predicted value; is the mean of y; m is the number of samples; the determination coefficient R2 is used to evaluate the generalization ability of the model: Use L2 regularization to limit the weight size to avoid overfitting of the model: J = J(θ) + λR(ω), Where J(θ) is the original loss function; λ is the regularization coefficient; R(ω) is the model complexity; ω is the weight.

9. The neural network-based high-voltage live working safety early warning method according to claim 1, characterized in that: The neural network-based high-voltage live working safety warning platform established in step 5 includes a data acquisition module, a data processing module, a user interface module and an alarm system. The alarm system includes multiple sets of measuring plate modules, a single-chip microcomputer, a power supply module, a display module and an alarm module.

10. The neural network-based high-voltage live working safety early warning method according to claim 9, characterized in that: The neural network-based high-voltage live working safety warning platform established in step 5 adopts a four-layer architecture including a data acquisition layer, a data processing layer, a warning analysis layer and a user interface layer.

Images