An overhead transmission line space state recognition method and system based on magnetic field sensing and physical information neural network

Abstract

The present application relates to the technical field of non-contact monitoring of the spatial state of high-voltage overhead transmission lines, and in particular to an overhead transmission line spatial state recognition method and system based on magnetic field sensing and physical information neural networks. The method provided by the present application mainly includes generating a magnetic field feature dataset based on a 3D electromagnetic model, dividing a training set and a test set according to a preset ratio; using the magnetic field feature dataset as input to train a double-branch physical information neural network, which is configured to collect current magnetic field data, input the standardized data into the trained model, and output the probability distribution of several states of the current line. Through the above method, the background field features are extracted by a remote reference sensor and are subjected to reverse mapping and elimination, thereby filtering out electromagnetic interference from adjacent loops and ground mirror currents from a physical mechanism, and greatly improving the recognition accuracy of the model under complex multi-circuit lines on the same tower.

Description

Technical Field

[0001] This invention relates to the field of non-contact monitoring technology for the spatial status of high-voltage overhead transmission lines, and more specifically, to a method and system for identifying the spatial status of overhead transmission lines based on magnetic field sensing and physical information neural networks. Background Technology

[0002] High-voltage overhead transmission lines, as the core carriers of cross-regional energy allocation, require real-time and accurate monitoring of their spatial conditions (such as sag and wind deflection) to ensure the safe and stable operation of the power grid. When lines experience excessive sag due to overload or temperature rise, or excessive wind deflection under strong winds, serious faults such as phase-to-phase flashover and short circuits can easily occur, directly threatening power supply reliability. However, existing monitoring technologies have significant limitations. Traditional manual inspections rely mainly on on-site surveys by maintenance personnel, which is extremely inefficient and easily affected by the experience of maintenance personnel and weather conditions. It is difficult to comprehensively cover large-scale lines and capture instantaneous anomalies. Contact-type online monitoring equipment needs to be directly installed on live conductors, which not only places extremely stringent requirements on the equipment's insulation performance but also makes it prone to leakage and damage under high-voltage environments. Furthermore, the installation and maintenance of such equipment require power outages, which undoubtedly increases maintenance costs and impacts power supply. While machine vision-based image recognition technology achieves monitoring functions using cameras and algorithms, it is significantly affected by factors such as rain, fog, low light at night, and strong winds. Furthermore, cameras are susceptible to contamination or obstruction, creating blind spots. Non-contact magnetic sensing technology, due to its non-invasiveness, resistance to extreme weather, and good stability, is applied to line monitoring. High-sensitivity, low-cost magnetoresistive sensors such as TMR / AMR can simultaneously monitor current, sag, and wind deflection angle. However, the coupling relationship between current and magnetic field signals is complex, and precise decoupling is extremely difficult, resulting in a significant reduction in the accuracy of the retrieved sag and wind deflection parameters.

[0003] When this type of technology moves towards practical engineering applications, it still faces three major challenges that urgently need to be addressed: First, the problem of multi-circuit coupling and interference from complex electromagnetic environments. In complex engineering scenarios where multiple circuits are erected on the same tower or operate adjacent to each other, the magnetic field of the target line is deeply coupled with the background magnetic field. Existing decoupling algorithms struggle to isolate the background field, resulting in a significant reduction in the accuracy of the sag and wind deflection parameters obtained through inversion. Second, the "physical distortion" defect of pure data-driven models. When encountering extreme weather and causing spatial anomalies, abnormal sample data is extremely scarce. Traditional pure data-driven deep learning models (such as conventional CNNs or RNNs) often rely solely on numerical fitting for blind extrapolation when dealing with these unseen extreme conditions. The output results frequently violate the basic spatial distribution laws of electromagnetic fields (such as the principle of magnetic flux continuity), leading to frequent false alarms and missed alarms. Third, the lack of a reliable physical self-calibration mechanism. Existing monitoring systems mostly output a single "state probability," which is prone to outputting high-confidence erroneous conclusions when sensors are subjected to transient electromagnetic pulse interference, lacking a fault-tolerant early warning system that incorporates physical priors. It is difficult to meet the monitoring requirements of smart grids for line status that are "highly accurate, highly adaptable, and highly physically reliable".

[0004] In summary, existing methods are insufficient to meet the need for accurate monitoring and identification of line spatial status under complex environmental conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for spatial state identification of overhead transmission lines based on magnetic field sensing and physical information neural networks. This method overcomes the drawbacks of the above monitoring methods and has the advantages of non-contact installation, strong environmental adaptability and high precision. It provides reliable technical support for real-time and accurate monitoring of sag and wind deviation anomalies in high-voltage overhead transmission lines.

[0006] This invention is achieved through the following technical solution: In a first aspect, the present invention provides a method for spatial state identification of overhead transmission lines based on magnetic field sensing and physical information neural networks, comprising: Acquire and preprocess multi-state transmission line magnetic field data, construct a 3D electromagnetic model of high-voltage overhead transmission line based on the multi-state transmission line magnetic field data, generate magnetic field feature dataset based on the 3D electromagnetic model, and divide the training set and test set according to a preset ratio. A dual-branch physical information neural network is trained using a magnetic field feature dataset as input, and the model parameters are optimized through a total loss function. The training set is then used to complete the model training. Collect current magnetic field data, standardize it, input it into the trained model, and output the probability distribution of several states of the current line. Preferably, the acquisition and preprocessing of multi-state transmission line magnetic field data includes: Spatial coordinates of conductor sag in windless conditions:

[0007] Spatial coordinates of the traverse under wind deflection conditions

[0008] In the formula, For the first Phase conductors in Coordinates along the axial direction, This represents the sag value of the conductor. For the conductor along Position coordinates along the axis For tower span, The distance from the wire hanging point to the ground. For the first Phase conductors in time Coordinates along the axial direction, for The wind deflection angle at any given moment, For the first Phase conductors in time Coordinates along the axis, For the first Phase conductors in The distance from the origin along the axis.

[0009] Preferably, the acquisition of current magnetic field data further includes deploying triaxial magnetoresistive sensors to maximize the capture of line state characteristics, including: Define the state parameter vector of the core space of the line to be identified, establish the sensitivity matrix of the magnetic field characteristics to the state parameter vector, and construct the global sensitivity matrix based on the output of several sensors; Construct a Fisher information matrix about the state parameter vector, and based on the Fisher information matrix and the global sensitivity matrix, construct an objective function for coordinate optimization that maximizes the Fisher information matrix.

[0010] Preferably, the state parameter vector includes:

[0011] The sensitivity matrix includes:

[0012] The Fisher information matrix includes:

[0013] In the formula, For the state parameter vector, For wind deflection angle, For the sensitivity matrix, , , These are the characteristic components of the magnetic field in the x, y, and z directions, respectively. For Fisher's information matrix, This is the global sensitivity matrix. The standard deviation is denoted as .

[0014] Preferably, the objective function includes:

[0015]

[0016] In the formula, This is the coordinate array of the 1st to the nth sensor. Let be the spatial coordinate array of the s-th sensor. For wires, Let s be the shortest spatial distance from the s-th sensor to the adjacent wire. For safe distance.

[0017] Preferably, the dual-branch physical information neural network includes: The input layer receives the magnetic field feature vector, which is then processed by the encoder layer to extract features. The feature transformation relationship of the encoder layer includes:

[0018] In the formula, This represents the 64-dimensional intermediate feature output by the first-layer encoder. This represents the 32-dimensional abstract feature output by the second-layer encoder; and This represents the weights and biases of the first-level linear transformation. and This represents the weights and biases of the second-level linear transformation. For the input magnetic field feature vector, For activation function, This represents the discard rate.

[0019] Preferably, the total loss function includes:

[0020] In the formula, For the total loss function, It is classification loss. It is a physical loss. It is the regularization loss. , , All are weighting coefficients. This represents the spatial gradient constraint loss for the magnetic field.

[0021] Preferably, the classification loss includes:

[0022] In the formula, For the sample size, Let be the probability that the i-th sample corresponds to the true label. Let be the true state label of the i-th sample.

[0023] Physical losses include:

[0024] In the formula, Let be the theoretical magnetic field value of the p-th phase conductor in the i-th sample at the s-th sensor. This represents the corresponding magnetic field value predicted by the model.

[0025] Regularization loss includes:

[0026] In the formula, In a two-branch physical information neural network, the set of model weights A single trainable weight parameter within, This is the set of model weights.

[0027] The magnetic field spatial gradient constraint loss includes:

[0028] In the formula, , , These are the spatial position variables in the three directions of the X-axis, Y-axis, and Z-axis in the spatial coordinate system.

[0029] Preferably, the probability distribution of several states of the current line is output, which includes calculating the physical confidence level and outputting the current line state based on the physical confidence level.

[0030] In the formula, For physical confidence level, The predicted magnetic field value of the p-th phase conductor at the s-th sensor is output by the physical prediction branch of the model. This is the theoretical magnetic field value of the p-th phase conductor at the s-th sensor, calculated based on the Biot-Savart law.

[0031] Secondly, the present invention also provides a spatial state identification system for overhead transmission lines based on magnetic field sensing and physical information neural networks, including the aforementioned method for spatial state identification of overhead transmission lines based on magnetic field sensing and physical information neural networks, and further including: The data processing module is configured to acquire and preprocess multi-state transmission line magnetic field data, construct a 3D electromagnetic model of the high-voltage overhead transmission line based on the multi-state transmission line magnetic field data, generate a magnetic field feature dataset based on the 3D electromagnetic model, and divide the training set and test set according to a preset ratio. The neural network module is configured to train a dual-branch physical information neural network using a magnetic field feature dataset as input, optimize model parameters through a total loss function, and complete model training using a training set. The output module is configured to collect the current magnetic field data, input the standardized data into the trained model, and output the probability distribution of several states of the current line.

[0032] The technical solution of the present invention has at least the following advantages and beneficial effects: The method provided by this invention mainly includes generating a magnetic field feature dataset based on a 3D electromagnetic model, dividing it into a training set and a test set according to a preset ratio; using the magnetic field feature dataset as input to train a dual-branch physical information neural network, configured to collect current magnetic field data, which is then standardized and input into the trained model, outputting the probability distribution of several states of the current line. Through this method, background field features are extracted using a remote reference sensor and eliminated through reverse mapping, filtering out electromagnetic interference from adjacent loops and ground mirror currents from a physical mechanism, greatly improving the model's recognition accuracy under complex multi-circuit lines on the same tower.

[0033] To address the weak feature decoupling caused by blindly deploying sensor arrays: the conventional empirical deployment method was abandoned, and a Jacobian matrix of the magnetic field on the sag and wind deflection parameters was established. D-optimal space search was performed with the goal of maximizing the determinant of the Fisher Information Matrix (FIM). This ensured that the sensor array was deployed at the coordinate system nodes with the highest feature sensitivity and the strongest orthogonality, thus achieving perfect decoupling of sag and wind deflection features from the hardware source.

[0034] Overcoming the generalization bottleneck of purely data-driven models under extreme conditions: A magnetic field spatial gradient constraint based on automatic differentiation is introduced into the PINN model. This forces the model to strictly adhere to field theory laws such as the magnetic flux continuity principle of Maxwell's equations while learning the data distribution, enabling it to make accurate inferences consistent with physical laws even when faced with extreme and abnormal weather conditions lacking training samples.

[0035] A fault-tolerant self-calibration system with high physical reliability is constructed: breaking through the decision-making blind spot of traditional AI that relies solely on "probability thresholds," it utilizes the dual-branch characteristics of PINN to calculate "physical confidence." When encountering non-modeled abnormal interference such as bird flight or transient electromagnetic pulses, a three-level early warning mechanism combining probability and confidence can effectively identify "false anomalies" and initiate secondary investigations, keeping the misclassification rate below 1.5%, fully meeting the stringent requirements of wide-area high-reliability monitoring for smart grids. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a diagram showing the tower model and sensor placement locations according to an embodiment of the present invention; Figure 2 This is a diagram of the dual-branch PINN neural network model architecture according to an embodiment of the present invention; Figure 3 This is a flowchart of the data acquisition and preprocessing process according to an embodiment of the present invention; Figure 4 This is a flowchart of line status identification and output according to an embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0039] Please refer to Figures 1-4 The present invention provides a method for spatial state identification of overhead transmission lines based on magnetic field sensing and physical information neural networks, comprising: S1: Acquire and preprocess multi-state transmission line magnetic field data, construct a 3D electromagnetic model of high-voltage overhead transmission line based on the multi-state transmission line magnetic field data, generate magnetic field feature dataset based on the 3D electromagnetic model, and divide the training set and test set according to a preset ratio. A three-dimensional electromagnetic field model was established using COMSOL to simulate three states: normal, abnormal sag, and abnormal wind deflection. Samples were generated in conjunction with MATLAB, and an environmental magnetic field physical stripping algorithm was introduced to eliminate background interference. Magnetic field components were collected by tower-mounted sensors to form a feature matrix, and training and test sets were divided and standardized.

[0040] S2: Train a dual-branch physical information neural network using the magnetic field feature dataset as input, optimize the model parameters through the total loss function, and complete the model training using the training set; Build and train a deep physical constraint bi-branch PINN model: Using standardized magnetic field features as input, build a bi-branch PINN model with an encoder, including a classification branch and a physical prediction branch; extract the magnetic field spatial gradient through automatic differentiation technology, design a total loss function that includes physical field evolution constraints, and train it with the Adam optimizer until the verification accuracy meets the target.

[0041] S3: Collect current magnetic field data, standardize it, input it into the trained model, and output the probability distribution of several states of the current line. Based on the Fisher Information Matrix (FIM) maximization criterion, six triaxial magnetic sensors were optimally arranged and placed at the conductor suspension point and the middle of the tower to collect the magnetic field components of the six sensors.

[0042] After the collected data is standardized, it is input into the trained PINN. The physical confidence score is calculated by combining the classification probability and the physical prediction bias. A three-level state warning with self-calibration capability is executed, and the model accuracy and misclassification rate are verified using the test set.

[0043] The method provided by this invention mainly includes generating a magnetic field feature dataset based on a 3D electromagnetic model, dividing it into a training set and a test set according to a preset ratio; using the magnetic field feature dataset as input to train a dual-branch physical information neural network, configured to collect current magnetic field data, which is then standardized and input into the trained model, outputting the probability distribution of several states of the current line. Through this method, background field features are extracted using a remote reference sensor and eliminated through reverse mapping, filtering out electromagnetic interference from adjacent loops and ground mirror currents from a physical mechanism, greatly improving the model's recognition accuracy under complex multi-circuit lines on the same tower.

[0044] An exemplary embodiment of the present invention includes acquiring and preprocessing multi-state transmission line magnetic field data, comprising: Establish a spatial coordinate system with the midpoint of the tower crossarm as the origin. The x-axis is perpendicular to the conductor direction, the y-axis is along the tower height direction, and the z-axis is parallel to the conductor direction. The spatial coordinates corresponding to the conductor sag in a windless state are as follows:

[0045] Spatial coordinates of the traverse under wind deflection conditions

[0046] In the formula, For the first Phase conductors in Coordinates along the axis, This is the sag value of the conductor (normally taken as 3 meters). For the conductor along Position coordinates along the axis For tower span, The distance from the wire hanging point to the ground. For the first Phase conductors in time Coordinates along the axis, for The wind deflection angle at any given moment, For the first Phase conductors in time Coordinates along the axis, For the first Phase conductors in The distance from the origin along the axis.

[0047] like Figure 1 As shown, Figure 1 This is a tower model and sensor placement diagram of an embodiment of the spatial state identification method for overhead transmission lines based on magnetic field sensing and physical information neural network of the present invention. The diagram marks the deployment positions of 6 sensors (S1~S6): S1 corresponds to the A phase conductor, with coordinates (-2.5m, 2m, 0m); S2 corresponds to the B phase conductor, with coordinates (0m, 2m, 0m); S3 corresponds to the C phase conductor, with coordinates (2.5m, 2m, 0m); S4, S5, and S6 are fixed in the middle section of the tower body below the crossarm, with coordinates (-2m, 5m, 0m), (0m, 5m, 0m), and (2m, 5m, 0m) respectively. All sensors maintain a uniform distance of 2m from adjacent conductors.

[0048] Secondly, based on the Biot-Safar law, the relationship between the magnetic field component of the p-th phase conductor at the s-th sensor and the conductor parameters is established as follows:

[0049] in, , The first Phase conductor in the first Generated at each sensor , , Components of axial magnetic induction intensity , , They are respectively , , Axial unit vector, The magnetic permeability of vacuum; Represents the first in the spatial coordinate system The position coordinates of each sensor, and the meanings of the remaining parameters are consistent with the formula for spatial coordinates of the conductor. Let be the power frequency load current flowing through the p-th phase conductor. Let be the spatial coordinates of the p-th phase conductor in the x-axis direction of the spatial coordinate system. Let be the spatial coordinates of the p-th phase conductor in the y-axis direction of the spatial coordinate system.

[0050] A dataset was generated using COMSOL-MATLAB co-simulation, simulating three fault types: 0 (normal), 1 (abnormal sag), and 2 (abnormal wind deflection). 1000 samples were generated for each fault type, totaling 3000 samples. Data from six sensors were extracted. , , The components form an 18-dimensional feature vector, which is then divided into training and test sets in a 75:25 ratio and standardized.

[0051] in, These are the standardized magnetic field eigenvalues. These are the original magnetic field characteristic values. This represents the sample mean for that feature dimension. This is the sample standard deviation for this feature dimension. After standardization, the mean of all features is 0 and the variance is 1.

[0052] To overcome background magnetic field interference caused by multiple lines erected on the same tower, this embodiment introduces an environmental magnetic field physical stripping algorithm in the data preprocessing stage. This is achieved using sensors deployed in the middle section of the tower. (Because it is far from the conductor, its sensitivity to high displacement fluctuations of the target conductor is low, but its perception of the macroscopic background magnetic field is strong.) This serves as an environmental reference point. An environmental field estimation operator is constructed. For the measured total magnetic field Perform physical stripping:

[0053] This stripping step effectively eliminates background interference caused by adjacent loops or ground mirror currents, providing high signal-to-noise ratio target line feature input for subsequent neural networks. This refers to the pure magnetic field signal generated solely by the target monitoring line after the environmental magnetic field has been stripped away. The environmental background magnetic field signal was synchronously collected by three reference sensors, S4, S5, and S6, which were deployed in the middle section of the tower.

[0054] In one exemplary embodiment of the present invention, acquiring current magnetic field data further includes deploying a triaxial magnetoresistive sensor to maximize the capture of line state characteristics, including: Define the state parameter vector of the core space of the line to be identified, establish the sensitivity matrix of the magnetic field characteristics to the state parameter vector, and construct the global sensitivity matrix based on the output of several sensors; Construct a Fisher information matrix about the state parameter vector, and based on the Fisher information matrix and the global sensitivity matrix, construct an objective function for coordinate optimization that maximizes the Fisher information matrix.

[0055] Specifically, the TMR2103 triaxial magnetoresistive sensor was selected, according to... Figure 1 The coordinates of the sensors are marked. To ensure that the sensor array can capture the line condition characteristics to the maximum extent and to solve the problem of "difficult decoupling between sag and wind deflection" caused by the coupling of current and magnetic field, this embodiment uses a spatial coordinate array of 6 tower-mounted sensors. It is a mathematical optimization configuration based on the Fisher Information Matrix (FIM), and its specific theoretical support and derivation process are as follows: First, define the vector of spatial state parameters of the line core to be identified as follows: ,in This is the conductor sag value. The angle of wind deflection.

[0056] Let the first One sensor ( The measured magnetic field characteristic components are: Establish the magnetic field characteristic pair state parameter vector. The sensitivity matrix, i.e., the Jacobian matrix. :

[0057] By combining the outputs of the six sensors, a global Jacobian matrix is ​​constructed. Its dimensions are .

[0058] The observed noise follows independent and identically distributed Gaussian white noise (covariance matrix is...). Under the following assumptions, For the identity matrix that matches the dimension of the global sensitivity matrix, the system has respect to the state parameters. Fisher Information Matrix Defined as:

[0059] In the formula, For the state parameter vector, For wind deflection angle, For the sensitivity matrix, , , These are the characteristic components of the magnetic field in the x, y, and z directions, respectively. For Fisher's information matrix, This is the global sensitivity matrix (global Jacobian matrix). The standard deviation is denoted as .

[0060] The inverse of the Fisher information matrix represents the Cramer-Rao lower bound (CRLB) of the parameter estimation, i.e., the minimum extremum of the covariance. To achieve optimal decoupling of the "sag" and "wind deflection" characteristics and minimize the uncertainty ellipsoidal volume in parameter identification, this invention employs the D-optimality criterion. The objective function for sensor coordinate optimization is established as maximizing the determinant of the Fisher information matrix:

[0061]

[0062] In the formula, This is the coordinate array of the 1st to the nth sensor. Let be the spatial coordinate array of the s-th sensor. For wires, Let s be the shortest spatial distance from the s-th sensor to the adjacent wire. For safe distance.

[0063] By solving the above nonlinear constrained optimization problem, the final determination is... Figure 1 The optimal sensor spatial coordinate array is shown. This optimization process ensures the sensor... Precisely deployed where the spatial gradient of the magnetic field of the conductor changes most drastically (i.e., the partial derivatives) and The absolute value is the largest) and the two parameters are most orthogonal to each other (i.e. The key region (where off-diagonal elements tend to zero) completely solves the problem of "weak feature coupling generalization" caused by the blind layout of traditional arrays from a physical level.

[0064] The sensor synchronously acquires 18-dimensional magnetic field feature vectors at a sampling frequency of no less than 100Hz. The acquired signals are processed by a 5th-order Butterworth low-pass filter.

[0065] in, This expression represents the transfer function of a Butterworth low-pass filter. Represents a complex frequency variable. This is the cutoff frequency of the filter. This is the filter order, and the filter is used to remove high-frequency noise from magnetic field signals. like Figure 3 As shown, Figure 3 This is a flowchart illustrating the data acquisition and preprocessing process of an embodiment of the spatial state identification method for overhead transmission lines based on magnetic field sensing and physical information neural networks according to the present invention. The flowchart fully presents the entire chain logic of "sensor signal acquisition - filtering and preprocessing - data encapsulation - wireless transmission - local caching": the filtered signal is transmitted to the STM32L476 microcontroller through the SPI interface, encapsulated into JSON format, and then uploaded to the edge gateway via LoRa wireless communication (transmission distance exceeds 1km). At the same time, 24 hours of data are locally cached to prevent data loss due to network interruption. In one exemplary embodiment of the present invention, the dual-branch physical information neural network includes: like Figure 2 As shown, Figure 2 This is a diagram of a dual-branch PINN neural network model architecture, representing an embodiment of the spatial state identification method for overhead transmission lines based on magnetic field sensing and physical information neural networks according to the present invention. The diagram shows the hierarchical structure of the model from left to right: the input layer receives an 18-dimensional magnetic field feature vector, which is then processed by the encoder layer to extract features. The feature transformation relationship of the encoder layer is as follows:

[0066] In the formula, This represents the 64-dimensional intermediate feature output by the first-layer encoder. This represents the 32-dimensional abstract feature output by the second-layer encoder; and This represents the weights and biases of the first-level linear transformation. and This represents the weights and biases of the second-level linear transformation. For the input magnetic field feature vector, For activation function, The discard rate is 0.2 or 0.1.

[0067] Based on features, a dual-output branch is constructed: the classification branch outputs three state probability distributions; the physics prediction branch outputs 18-dimensional theoretical magnetic field components. The total loss function is:

[0068] In the formula, For the total loss function, It is classification loss. It is a physical loss. It is the regularization loss. , , All are weighting coefficients. For magnetic field spatial gradient constraint loss, , These are weighting coefficients used to balance the contributions of the three types of losses. This invention features a unique magnetic field spatial gradient constraint loss. The predicted magnetic field is calculated using an automatic differentiation technique. The first-order partial derivative with respect to spatial coordinates ensures that it strictly satisfies the principle of flux continuity of the Biot-Savart law in the source-free region. ).

[0069] Specifically, the classification loss (sparse cross-entropy) includes:

[0070] In the formula, For the sample size, Let be the probability that the i-th sample corresponds to the true label. The true state label of the i-th sample is 0 for normal, 1 for abnormal sag, and 2 for abnormal wind deviation.

[0071] Physical loss (mean squared error) includes:

[0072] In the formula, Let be the theoretical magnetic field value of the p-th phase conductor in the i-th sample at the s-th sensor. This represents the corresponding magnetic field value predicted by the model.

[0073] Regularization loss (L2 regularization) includes:

[0074] In the formula, In a two-branch physical information neural network, the set of all weights A single trainable weight parameter within, This is the set of model weights.

[0075] The Adam optimizer (learning rate 0.001), batch size 32, early stopping strategy (termination if no improvement in accuracy is verified after 5 consecutive epochs), 50 training epochs, training on GPU based on PyTorch framework, and training time no more than 9.9s were used.

[0076] The magnetic field spatial gradient constraint loss includes:

[0077] In the formula, , , These are the spatial position variables in the three directions of the X-axis, Y-axis, and Z-axis in the spatial coordinate system.

[0078] Introduction The forced model not only fits the data numerically, but also obeys the electromagnetic field evolution law in the physical space manifold, which greatly enhances the model's generalization ability under extreme and abnormal samples.

[0079] An exemplary embodiment of the present invention, such as Figure 4 As shown, Figure 4 This is a flowchart illustrating the line state identification and output process of an embodiment of the spatial state identification method for overhead transmission lines based on magnetic field sensing and physical information neural networks according to the present invention. The flowchart clearly shows the identification process: the model not only outputs the probability distribution of the three states through classification branches, but also... Simultaneously, the theoretical and measured magnetic fields output by the physical prediction branch are extracted, and the physical confidence level is calculated:

[0080] In the formula, For physical confidence level, The predicted magnetic field value of the p-th phase conductor at the s-th sensor is output by the physical prediction branch of the model. This is the theoretical magnetic field value of the p-th phase conductor at the s-th sensor, calculated based on the Biot-Savart law.

[0081] Based on probability and physical priors, this method implements a three-level state warning self-checking mechanism: Level 1 Warning (Continuous Monitoring): When the probability of an abnormal state... At that time, the physical laws and classification results were highly consistent, indicating that the line was in the early stage of deformation; Level 2 Early Warning (Interference Investigation): When the probability of an abnormal state... but If the pure data classification results contradict the laws of physical space, it is determined that the system is subjected to strong external transient electromagnetic interference, marked as "to be confirmed", and compared with the first 3 sampling points to filter out false anomalies; Level 3 Warning (Emergency Alert): When the probability of an abnormal state... and When a high-confidence confirmation of abnormal line sag or wind deviation occurs, an alarm is immediately output.

[0082] Recognition performance is evaluated using the following metrics:

[0083] in, For classification accuracy, To correctly identify the number of samples under normal conditions, To correctly identify the number of samples for sag anomalies, To correctly identify the number of samples for wind drift anomalies. This represents the total number of test samples.

[0084] The relationship between misclassification rate and classification accuracy is established as follows:

[0085] in, Misclassification rate For classification accuracy, it is used to quantify the probability of the model making a mistake.

[0086] Therefore, this invention employs the aforementioned method for spatial state identification of overhead transmission lines based on magnetic field sensing and physical information neural networks. It utilizes a tower-mounted magnetic sensor array optimized based on Fisher information matrix deployment to collect three-dimensional magnetic field data. Combined with an environmental field physical stripping algorithm and standardized preprocessing, and through a PINN model embedding deep physical constraints such as automatic differential spatial gradients, it not only achieves non-contact and accurate identification of normal line conditions, abnormal sag, and abnormal wind deflection, but also innovatively constructs a three-level early warning self-checking mechanism based on physical confidence. This invention effectively overcomes the industry challenges of multi-circuit magnetic field coupling interference in complex environments and the weak generalization of purely data-driven models. It possesses significant advantages such as convenient installation, strong anti-interference capability, and high identification accuracy and physical reliability. A spatial state identification system for overhead transmission lines based on magnetic field sensing and physical information neural networks, comprising the aforementioned method for spatial state identification of overhead transmission lines based on magnetic field sensing and physical information neural networks, and further comprising: The data processing module is configured to acquire and preprocess multi-state transmission line magnetic field data, construct a 3D electromagnetic model of the high-voltage overhead transmission line based on the multi-state transmission line magnetic field data, generate a magnetic field feature dataset based on the 3D electromagnetic model, and divide the training set and test set according to a preset ratio. The neural network module is configured to train a dual-branch physical information neural network using a magnetic field feature dataset as input, optimize model parameters through a total loss function, and complete model training using a training set. The output module is configured to collect the current magnetic field data, input the standardized data into the trained model, and output the probability distribution of several states of the current line.

[0087] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0088] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. This computer software product, stored in a storage medium, includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0089] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for spatial state identification of overhead transmission lines based on magnetic field sensing and physical information neural networks, characterized in that, include: Acquire and preprocess multi-state transmission line magnetic field data, construct a 3D electromagnetic model of high-voltage overhead transmission line based on the multi-state transmission line magnetic field data, generate magnetic field feature dataset based on the 3D electromagnetic model, and divide the training set and test set according to a preset ratio. A dual-branch physical information neural network is trained using a magnetic field feature dataset as input, and the model parameters are optimized through a total loss function. The training set is then used to complete the model training. Collect current magnetic field data, standardize it, input it into the trained model, and output the probability distribution of several states of the current line.

2. The method for spatial state identification of overhead transmission lines based on magnetic field sensing and physical information neural networks according to claim 1, characterized in that, The acquisition and preprocessing of multi-state transmission line magnetic field data includes: Spatial coordinates of conductor sag in windless conditions: Spatial coordinates of the traverse under wind deflection conditions In the formula, For the first Phase conductors in Coordinates along the axis, This represents the sag value of the conductor. For the conductor along Position coordinates along the axis For tower span, The distance from the wire hanging point to the ground. For the first Phase conductors in time Coordinates along the axis, for The wind deflection angle at any given moment, For the first Phase conductors in time Coordinates along the axial direction, For the first Phase conductors in The distance from the origin along the axis.

3. The method for spatial state identification of overhead transmission lines based on magnetic field sensing and physical information neural networks according to claim 2, characterized in that, The acquisition of current magnetic field data also includes deploying triaxial magnetoresistive sensors to maximize the capture of line state characteristics, including: Define the state parameter vector of the core space of the line to be identified, establish the sensitivity matrix of the magnetic field characteristics to the state parameter vector, and construct the global sensitivity matrix based on the output of several sensors; Construct a Fisher information matrix about the state parameter vector, and based on the Fisher information matrix and the global sensitivity matrix, construct an objective function for coordinate optimization that maximizes the Fisher information matrix.

4. The method for spatial state identification of overhead transmission lines based on magnetic field sensing and physical information neural networks according to claim 3, characterized in that, The state parameter vector includes: The sensitivity matrix includes: The Fisher information matrix includes: In the formula, For the state parameter vector, For wind deflection angle, For the sensitivity matrix, , , These are the characteristic components of the magnetic field in the x, y, and z directions, respectively. For Fisher's information matrix, This is the global sensitivity matrix. The standard deviation is denoted as .

5. The method for spatial state identification of overhead transmission lines based on magnetic field sensing and physical information neural networks according to claim 4, characterized in that, The objective function includes: In the formula, This is the coordinate array of the 1st to the nth sensor. Let be the spatial coordinate array of the s-th sensor. For wires, Let s be the shortest spatial distance from the s-th sensor to the adjacent wire. For safe distance.

6. The method for spatial state identification of overhead transmission lines based on magnetic field sensing and physical information neural networks according to claim 5, characterized in that, The dual-branch physical information neural network includes: The input layer receives the magnetic field feature vector, which is then processed by the encoder layer to extract features. The feature transformation relationship of the encoder layer includes: In the formula, This represents the 64-dimensional intermediate feature output by the first-layer encoder. This represents the 32-dimensional abstract feature output by the second-layer encoder; and This represents the weights and biases of the first-level linear transformation. and This represents the weights and biases of the second-level linear transformation. For the input magnetic field feature vector, For activation function, This represents the discard rate.

7. The method for spatial state identification of overhead transmission lines based on magnetic field sensing and physical information neural networks according to claim 6, characterized in that, The total loss function includes: In the formula, For the total loss function, It is classification loss. It is a physical loss. It is the regularization loss. , , All are weighting coefficients. This represents the spatial gradient constraint loss for the magnetic field.

8. The method for spatial state identification of overhead transmission lines based on magnetic field sensing and physical information neural networks according to claim 7, characterized in that, The classification loss includes: In the formula, For the sample size, Let be the probability that the i-th sample corresponds to the true label. Let be the true state label of the i-th sample; Physical losses include: In the formula, Let be the theoretical magnetic field value of the p-th phase conductor in the i-th sample at the s-th sensor. The corresponding magnetic field value predicted by the model; Regularization loss includes: In the formula, In a two-branch physical information neural network, the set of model weights A single trainable weight parameter within, For the set of model weights; The magnetic field spatial gradient constraint loss includes: In the formula, In the spatial coordinate system, axis, shaft and Spatial position variables in the three directions of the axis.

9. A method for spatial state identification of overhead transmission lines based on magnetic field sensing and physical information neural networks according to claim 8, characterized in that, The probability distribution of several states of the current line is output, which includes calculating the physical confidence level and outputting the current line state based on the physical confidence level: In the formula, For physical confidence level, The predicted magnetic field value of the p-th phase conductor at the s-th sensor is output by the physical prediction branch of the model. This is the theoretical magnetic field value of the p-th phase conductor at the s-th sensor, calculated based on the Biot-Savart law.

10. A spatial state identification system for overhead transmission lines based on magnetic field sensing and physical information neural networks, characterized in that, The method for spatial state identification of overhead transmission lines based on magnetic field sensing and physical information neural networks, as described in any one of claims 1-9, further includes: The data processing module is configured to acquire and preprocess multi-state transmission line magnetic field data, construct a 3D electromagnetic model of the high-voltage overhead transmission line based on the multi-state transmission line magnetic field data, generate a magnetic field feature dataset based on the 3D electromagnetic model, and divide the training set and test set according to a preset ratio. The neural network module is configured to train a dual-branch physical information neural network using a magnetic field feature dataset as input, optimize model parameters through a total loss function, and complete model training using a training set. The output module is configured to collect the current magnetic field data, input the standardized data into the trained model, and output the probability distribution of several states of the current line.

Images