A method and system for monitoring the sag of a power transmission line based on a non-contact magnetic array

Abstract

The present application belongs to the field of power transmission line state monitoring, and relates to a power transmission line sag monitoring method and system based on a non-contact magnetic array, comprising: collecting multi-dimensional space magnetic field data; constructing an electromagnetic physical prior ratio sequence of a three-phase conductor; constructing a space-time feature matrix and obtaining a standardized space-time feature; inputting the standardized space-time feature into a pre-space unmixing network to output a pure space feature; obtaining a micro kinematics rate feature; constructing the micro kinematics rate feature into a micro kinematics sliding prediction window sequence and inputting it into a differential attention long short-term memory network to obtain a trained differential attention long short-term memory network; inputting the current micro kinematics sliding prediction window sequence into the trained differential attention long short-term memory network to output a micro downward movement rate prediction value; obtaining an absolute sag prediction value and performing a sag out-of-limit early warning; and stripping extreme current jump and three-phase phasor crosstalk to overcome saturation collapse of a data-driven model in a monotonic extrapolation domain.

Description

Technical Field

[0001] This invention relates to the field of power transmission line condition monitoring technology, specifically to a method and system for monitoring power transmission line sag based on a non-contact magnetic array. Background Technology

[0002] With the rapid development of smart grids and ultra-high voltage (UHV) transmission technologies, the safe and stable operation of overhead transmission lines is crucial to national energy security. Conductor sag is a core safety parameter for assessing the operational status of a line. During long-term operation, factors such as high ambient temperatures, high-load operation heat generation, and extreme icy weather can easily cause conductors to undergo thermal expansion and contraction, material creep, or ice accumulation, leading to a significant increase in mid-span sag. If the sag is excessive, the safe distance between the conductor and the ground or crossings will be severely insufficient, easily triggering line flashovers, grounding short circuits, or even large-scale power outages.

[0003] Currently, the power grid industry widely uses contact-based monitoring methods such as tilt sensors and tension sensors, which require power outages for installation on towers and are prone to mechanical freezing and damage in freezing, rain, and snow environments. Visual image monitoring technologies based on visible light or infrared thermal imaging are easily rendered useless in adverse weather conditions such as heavy fog, heavy rain, or nighttime when visibility is extremely low due to lens obstruction. To overcome the inherent shortcomings of these contact and optical solutions, non-contact magnetic field sensing technology based on the Biot-Savart Law has emerged. This technology uses a sensor array installed on a tower to measure the spatial magnetic field generated by the conductor to determine its spatial position. It boasts unique advantages such as convenient installation and complete immunity to adverse weather conditions, and is becoming a research hotspot in the field of intelligent sensing.

[0004] Despite the promising prospects of non-contact magnetic field sensing technology, existing physical inversion methods face insurmountable physical limitations in strong electromagnetic and complex power grid operating environments. First, the magnetic field strength at a point in space is not only affected by the spatial distance between conductors but also directly proportional to the magnitude of the transient transmission current flowing through them. In actual power grid operation, the switching of user-side loads or adjustments to dispatching strategies can cause highly random and violent high-frequency jumps in transmission current, leading to drastic changes in the spatial magnetic field. Traditional single-point or single-layer sensor systems, lacking a physical differential reference, cannot separate the "magnetic field abrupt change caused by load surges" from the "slow increase in magnetic field caused by actual conductor slump," making them highly susceptible to false alarms. Second, real high-voltage transmission lines are three-phase AC systems (phases A, B, and C), with a strict 120-degree complex phase difference between the three phases, arranged asymmetrically in space. This means that any single-phase sensor not only receives the magnetic field of the target phase but is also subject to complex phasor spatial crosstalk caused by the diffusion of the magnetic fields of the other two phases. Under the premise of random current fluctuations, this three-phase AC nonlinear superposition exhibits extremely complex time-varying coupling characteristics. Conventional linear filtering or simple ratio methods cannot achieve complete decoupling of different phases. Ultimately, the remaining small nonlinear offset will be continuously amplified as the electromagnetic environment deteriorates, making it difficult to meet the high-precision early warning requirements of engineering-level monitoring with absolute error.

[0005] In recent years, an increasing number of researchers have attempted to introduce multivariate temporal deep learning models such as Long Short-Term Memory (LSTM) networks to directly establish an end-to-end nonlinear mapping from mixed magnetic field characteristics to absolute sag values, hoping to circumvent complex physical decoupling modeling through the fitting capabilities of neural networks. However, research on the long-timescale evolution characteristics of sag shows that traditional deep temporal prediction networks face a severe "out-of-distribution (OOD) saturation collapse" bottleneck in this specific physical scenario. The sag evolution of transmission lines typically exhibits a long-period, monotonic cumulative trend (e.g., slow sag caused by high-temperature creep or continuous icing), which means that when the model is deployed in practice, the input feature values ​​and target sag values ​​often exceed the maximum range seen in the historical training set. When faced with such an unknown, extremely monotonic extrapolation domain, the activation functions inside the traditional LSTM network (such as Tanh and Sigmoid gates) are instantly filled by unprecedentedly large values, reaching the saturation limit and causing the backpropagation gradient to approach zero. In order to minimize prediction loss, the network will experience mode collapse, directly outputting a flat constant line of historical averages, and completely losing the ability to extrapolate the true downward trend of the sag.

[0006] Therefore, this application provides a method and system for monitoring the sag of transmission lines based on a non-contact magnetic array. Starting from the underlying physical mechanism, it can completely isolate extreme current jumps and three-phase phasor crosstalk, and overcome the saturation collapse of the data-driven model in the monotonic extrapolation domain. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for monitoring the sag of transmission lines based on a non-contact magnetic array, addressing the following problems: existing non-contact magnetic field sensing technologies are highly susceptible to extreme load fluctuations under strong electromagnetic interference, suffer from severe phasor crosstalk in three-phase AC power, and traditional deep time-series prediction networks are prone to out-of-distribution (OOD) saturation collapse when dealing with long-period monotonic sag data. The specific solution is as follows: A method for monitoring the sag of transmission lines based on a non-contact magnetic array includes: S1: acquiring multi-dimensional spatial magnetic field data through a double-layer magnetoresistive sensor array deployed directly below the crossarm of the tower; the double-layer magnetoresistive sensor array is divided into upper and lower layers along the tower height direction, with multiple magnetic sensors deployed in each layer; S2: constructing an electromagnetic physical prior ratio sequence of the three-phase conductors based on the spatial coordinates of each magnetic sensor in the double-layer magnetoresistive sensor array; S3: concatenating the electromagnetic physical prior ratio sequence with the original magnetic field components of the multi-dimensional spatial magnetic field data to construct a spatiotemporal feature matrix, and standardizing the spatiotemporal feature matrix to obtain standardized spatiotemporal features; S4: inputting the standardized spatiotemporal features into a pre-processing spatial demixing network to filter out three-phase phasor crosstalk and output clean spatial features. S5: Apply a first-order difference operator to the clean space features for stabilization to obtain micro-kinematic velocity features; S6: Construct the micro-kinematic velocity features into a micro-kinematic sliding prediction window sequence, input it into the differential attention long short-term memory network, and optimize the network parameters through a composite physical constraint total loss function to obtain a trained differential attention long short-term memory network; S7: Collect the current real-time magnetic field data, calculate the current micro-kinematic sliding prediction window sequence, input the current micro-kinematic sliding prediction window sequence into the trained differential attention long short-term memory network, and output the small downward displacement rate prediction value; S8: Perform numerical integration reconstruction based on the small downward displacement rate prediction value to obtain the absolute sag prediction value, and execute the sag over-limit warning based on the absolute sag prediction value.

[0008] Further, S1 includes: S11: Establishing a spatial rectangular coordinate system with the ground projection of the midpoint of the tower crossarm as the origin; the x-axis of the spatial rectangular coordinate system is parallel to the direction of the conductor, the y-axis is along the direction of the tower crossarm and perpendicular to the direction of the conductor, and the z-axis is vertically upward along the tower height; S12: Deploying an upper sensor array and a lower sensor array in the spatial rectangular coordinate system; the height of the upper sensor array above the ground is a first preset height, and the height of the lower sensor array above the ground is a second preset height, wherein the first preset height is greater than the second preset height. Height; S13: Deploy magnetic sensors according to the vertical mapping logic of independent pairing of each phase conductor, thereby forming three pairs of magnetic sensors corresponding to phase A, phase B and phase C respectively; The vertical mapping logic requires that the magnetic sensors in the upper sensor array correspond one-to-one with the magnetic sensors in the lower sensor array, and be located on the same y-axis coordinate and face the vertical projection plane of the same phase conductor; S14: The three pairs of magnetic sensors synchronously collect the horizontal magnetic induction intensity superimposed modulus component and the vertical magnetic induction intensity superimposed modulus component of their respective corresponding phase conductors to form the multidimensional spatial magnetic field data.

[0009] Furthermore, S2 includes: S21: For any phase conductor, calculate the amplitude of the magnetic induction intensity measured by the magnetic sensor on the upper layer of that phase conductor: ; in, For being in the first The upper layer directly below the phase conductor The amplitude of magnetic flux density measured by each sensor; is the vacuum permeability constant; For the first The effective value of transient load current flowing through high-voltage overhead transmission lines at any given moment; Pi; The vertical height of the upper magnetoresistive sensor array from the horizontal ground surface; This refers to the static suspension height of the transmission line at the insulator hanging point; For the first Phase conductor at time The absolute sag value at mid-span; To take the absolute value; S22: Calculate the amplitude of the magnetic flux density measured by the magnetic sensor in the lower layer of the conductor in this phase: ; in, For being in the first The lower layer directly below the phase conductor The amplitude of magnetic flux density measured by each sensor; The vertical height of the lower magnetoresistive sensor array from the horizontal ground surface; S23: Based on the magnetic flux density amplitudes measured by the upper and lower magnetic sensors of this phase conductor, construct the electromagnetic physics prior ratio: ; in, The a priori ratio of the electromagnetic physics of the space magnetic field; It is the smoothing minimum constant; S24: Calculate the electromagnetic physical prior ratios of phases A, B, and C respectively to obtain the electromagnetic physical prior ratio sequence composed of the electromagnetic physical prior ratios of the three phases.

[0010] Furthermore, S3 includes: S31: splicing the electromagnetic physical prior ratios of the three phases in the electromagnetic physical prior ratio sequence with the horizontal and vertical magnetic field components of the six magnetic sensors in a preset order to form a fifteen-dimensional spatiotemporal feature matrix; S32: dividing the spatiotemporal feature matrix into a training set and a test set according to a preset ratio; S33: standardizing each spatiotemporal feature dimension using the sample mean and sample standard deviation of that feature dimension on the training set to obtain the standardized spatiotemporal feature.

[0011] Furthermore, S5 includes: S51: subtracting the pure space features at the current moment from the pure space features at the previous moment to obtain a micro-evolution rate vector; S52: subtracting the absolute sag value at the current moment from the absolute sag value at the previous moment to obtain the true physical micro-downward displacement rate; S53: using the micro-evolution rate vector as the micro-kinematic rate feature to map the non-stationary absolute position domain features to the stationary velocity domain.

[0012] Furthermore, the input to the differential attention long short-term memory network is a sequence of microkinematic sliding prediction windows, and the output is a predicted value of a small downward shift rate, including: S61: Constructing a sequence of microkinematic sliding prediction windows with a preset step size: ; in, For micro-kinematic sliding time window; and The pure space feature sequences are respectively in and The gradient sequence of small changes between adjacent time steps; For time variables; The preset temporal receptive field length; S62: Input the microkinematic sliding prediction window sequence into the long short-term memory cells of the differential attention long short-term memory network time-by-time, and extract the hidden states of the sequence time-by-time: ; in, For each time step within the window The temporal hidden state vector; The output gate activation value of the long short-term memory cell at time step τ; For Hadamah accumulation; It is the hyperbolic tangent function; The cellular state of long short-term memory cells at time step τ; S63: Feed the sequential hidden states at each time step into the temporal attention layer and calculate the dynamic weights for attention allocation at each time step: ; in, Dynamic weights are assigned to attention; exp is an exponential function; Let T be the scoring weight vector for the attention mechanism; T is the transpose of the matrix. The scoring weight matrix for the attention mechanism; is the bias vector for the attention mechanism; m is the index of the time step traversed when normalizing and summing the hidden states within the sliding window; Let be the temporal hidden state vector at the m-th time step; S64: Weighted summation of the hidden states at each time step using attention-assigned dynamic weights to generate a high-dimensional context vector: ; in, It is a high-dimensional context vector; For each time step within the window; S65: Input the high-dimensional context vector into the terminal fully connected regression layer and decode to map out the predicted value of the tiny downward shift rate.

[0013] Furthermore, the total loss function of the composite physical constraints is: ; ; ; ; in, This represents the total loss during network backpropagation. , and These are the weight coefficients for the first, second, and third hyperparameters, respectively. The mean squared error loss is the micro-rate regression; M is the batch size during network training; i is the sample variable. and The first The predicted micro-downshift rate and the actual downshift rate for each sample; For the first The true downward shift rate of each sample; To smooth the constraint loss of the time-series evolution of the conductor; For L2 structure regularization loss; The set of all trainable parameters for the Diff-ALSTM network Individual weights within; This is the set of trainable parameters for the Diff-ALSTM network.

[0014] Furthermore, the step of reconstructing the absolute sag prediction value based on the predicted value of the minute downward displacement rate includes: S81: taking the start or retest time of online monitoring as the origin of integration, extracting the true static absolute sag at that time as the reference sag; S82: starting from the origin of integration, accumulating the predicted value of the minute downward displacement rate at each time point to obtain the cumulative downward displacement; S83: adding the reference sag to the cumulative downward displacement to reconstruct the predicted value of the absolute sag at the current time.

[0015] Furthermore, the step of executing sag limit overshoot warning based on the absolute sag prediction value includes: S84: when the absolute sag prediction value exceeds a preset percentage of the overshoot value and the predicted value of the small downward movement rate increases monotonically in a positive direction, a first-level warning is output; S85: when the single-point sag increment jumps, but the small downward movement rate of subsequent continuous sampling points returns to zero or oscillates, a second-level warning is output; S86: when the absolute sag prediction value exceeds the safety design limit and the confidence level reaches a preset threshold after artifact removal, a third-level warning is output.

[0016] This invention also provides a transmission line sag monitoring system based on a non-contact magnetic array, comprising a double-layer non-contact magnetic array, a power supply module, a signal conditioning module, a feature processing module, a sag prediction module, and an early warning module. The double-layer non-contact magnetic array is deployed directly below the crossarm of the tower and is divided into upper and lower layers along the tower height, with multiple magnetic sensors in each layer. The power supply module is connected to the double-layer non-contact magnetic array and is used to drive each magnetic sensor to synchronously acquire the original magnetic field signal. The signal conditioning module is connected to the double-layer non-contact magnetic array and is used to process the original magnetic field signal to output multi-dimensional spatial magnetic field data. The feature processing module is connected to the signal conditioning module and is used to construct an electromagnetic physical prior ratio sequence of the three-phase conductors based on the spatial coordinates of each magnetic sensor in the double-layer magnetoresistive sensor array. The electromagnetic physical prior ratio sequence is concatenated with the original magnetic field components of the multi-dimensional spatial magnetic field data to construct a spatiotemporal feature matrix. The spatiotemporal feature matrix is ​​then standardized to obtain... Standardized spatiotemporal features are input into a pre-spatial demixing network to filter out three-phase phasor crosstalk and output clean spatial features. The clean spatial features are then stabilized using a first-order difference operator to obtain micro-kinematic velocity features. The sag deduction module, connected to the feature processing module, constructs a micro-kinematic sliding prediction window sequence from the micro-kinematic velocity features, inputs it into a differential attention long short-term memory network, and optimizes the network parameters using a composite physical constraint total loss function to obtain a trained differential attention long short-term memory network. Real-time magnetic field data is collected, and the current micro-kinematic sliding prediction window sequence is calculated. This sequence is then input into the trained differential attention long short-term memory network to output a small downward displacement rate prediction value. Numerical integration is performed based on the small downward displacement rate prediction value to reconstruct the absolute sag prediction value. The early warning module, connected to the sag deduction module, executes a sag over-limit early warning based on the absolute sag prediction value.

[0017] The present invention has the following advantages and beneficial effects: This invention achieves high-fidelity, high-precision continuous reconstruction of the absolute sag of high-voltage transmission lines under strong electromagnetic interference by constructing a two-layer spatial magnetic field geometric ratio method to remove dynamic current disturbances, utilizing a spatial unmixing network to eliminate complex phase crosstalk, and constructing a first-order differential micro-kinematic attention network (Diff-ALSTM) for model extrapolation training. This invention constructs a dual-layer sensor topology and a physical space magnetic field ratio method. Through low-level mathematical division, the transient current variables common to the numerator and denominator are perfectly reduced before being input into the deep network. The extracted prior ratio is transformed into a pure spatial geometric parameter, giving the system absolute immunity to extreme electromagnetic load jumps and greatly reducing the false sag warning rate in complex electromagnetic environments.

[0018] This invention splices the prior geometric ratio with the multidimensional original magnetic field components using tensors, and uses a pre-space unmixing network containing multi-level activation functions to implicitly solve the complex high-dimensional nonlinear complex superposition inverse operation, thoroughly cleaning up the diffusion pollution of non-target phases, and significantly improving the monitoring accuracy of multi-phase lines on the same tower.

[0019] This invention proposes a first-order difference micro-kinematic network concept, abandoning the direct prediction of non-stationary absolute spatial positions and instead forcing the network to learn the minute descent rate of the conductor within a stable velocity domain. Regardless of how unknown or extreme the absolute sag descends, its first-order difference rate remains stable. By combining an attention mechanism with post-Eulerian discrete integral reconstruction, the limitations of neural network extrapolation are fundamentally overcome at the mathematical level, achieving high-fidelity absolute sag reconstruction under extremely monotonic conditions.

[0020] This invention introduces a physical constraint loss for temporal smoothness into the total loss function, which forces the output of the neural network to strictly obey the physical laws of continuity and inertia of the mechanical motion of large power transmission lines. This greatly suppresses non-physical jumps in the prediction rate caused by accidental high-frequency noise from sensors, making the final integrally reconstructed absolute sag curve smoother and more noise-resistant. Attached Figure Description

[0021] Figure 1 This is an exemplary schematic diagram of the tower model and sensor placement provided by the present invention; Figure 2 This is an exemplary flowchart of the transmission line sag monitoring method provided by the present invention. Detailed Implementation

[0022] 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.

[0023] This invention provides a method for monitoring the sag of transmission lines based on a non-contact magnetic array, such as... Figure 2 As shown, its core process includes S1 to S8, specifically including the following: S1: Multidimensional spatial magnetic field data are collected by a double-layer magnetoresistive sensor array deployed directly below the crossarm of the tower; the double-layer magnetoresistive sensor array is divided into upper and lower layers along the height of the tower, and multiple magnetic sensors are deployed in each layer.

[0024] Directly below the tower crossarm refers to the space area covered by the tower crossarm's projection onto the ground, i.e., the space area traversed by the vertical projection of the three-phase conductors. Deploying the sensor array at this location ensures that each magnetic sensor faces the vertical projection plane of the target conductor, thereby capturing the conductor's magnetic field with the shortest pole pitch and the highest signal-to-noise ratio, while avoiding contact with the conductor upon climbing the tower. A dual-layer magnetoresistive sensor array refers to a collection of magnetic sensors arranged in upper and lower layers along the tower height direction. The upper layer is the upper sensor array, and the lower layer is the lower sensor array, with both layers sampling independently and synchronously. By setting up a dual-layer structure, the pole pitches of the upper and lower sensors directly below the same phase conductor are different. Dividing their measurements by the two layers cancels out common terms related to current, thus eliminating transient current interference. The magnetic sensors can be extremely low-noise TMR (tunneling magnetoresistive) sensors, and each magnetic sensor can be a triaxial magnetoresistive sensor to simultaneously acquire the horizontal and vertical magnetic field components. Multidimensional spatial magnetic field data refers to a set of magnetic field components containing multiple spatial dimensions, synchronously acquired by multiple magnetic sensors in a dual-layer magnetoresistive sensor array. For example, when using 6 magnetic sensors, each outputting magnetic induction intensity components in both horizontal and vertical directions, the multidimensional spatial magnetic field data is a set of 12 original magnetic field components.

[0025] In some embodiments, S1 further includes S11 to S14. S11: Establish a spatial rectangular coordinate system with the ground projection of the midpoint of the tower crossarm as the origin; the x-axis of the spatial rectangular coordinate system is parallel to the direction of the conductor, the y-axis is along the direction of the tower crossarm and perpendicular to the direction of the conductor, and the z-axis is vertically upward along the tower height.

[0026] The midpoint of a tower crossarm refers to the geometric midpoint of the crossarm along its length. Its ground projection is used as the origin of a coordinate system, which describes the spatial positions of each phase conductor and each sensor. The conductor's direction of travel refers to the direction in which the transmission conductor is erected, i.e., the span direction. For example, under the assumption of a two-dimensional cross-section for a long straight conductor, only the magnetic field distribution within the yz cross-section perpendicular to the conductor's direction needs to be considered.

[0027] S12: Deploy an upper sensor array and a lower sensor array in a spatial rectangular coordinate system; the height of the upper sensor array above the ground is a first preset height, and the height of the lower sensor array above the ground is a second preset height, wherein the first preset height is greater than the second preset height.

[0028] The upper sensor array is a collection of magnetic sensors deployed at a higher position, and the lower sensor array is a collection of magnetic sensors deployed at a lower position. The first preset height is the vertical height of the upper magnetoresistive sensor array from the horizontal ground. The second preset height is the vertical height of the lower magnetoresistive sensor array from the horizontal ground. ,and For example, take the first preset height. Second preset height .

[0029] S13: Deploy magnetic sensors according to the vertical mapping logic of independent pairing of each phase conductor, thereby forming three pairs of magnetic sensors corresponding to phase A, phase B and phase C respectively; the vertical mapping logic requires that the magnetic sensors in the upper sensor array correspond one-to-one with the magnetic sensors in the lower sensor array, and be located on the same y-axis coordinate and face the vertical projection plane of the same phase conductor.

[0030] Vertical mapping logic, where each phase conductor is independently paired (referred to as vertical mapping logic), refers to configuring a separate pair of magnetic sensors, one on the upper layer and one on the lower layer, for each phase conductor. These magnetic sensors share the same y-axis coordinate in the horizontal direction and are aligned vertically with the vertical projection plane of the corresponding phase conductor. This ensures that the measurement value of the sensor pair is determined solely by the difference in the pole distance from the phase conductor to the two sensors. For example, six magnetic sensors could be used. The upper layer is , , The lower layer is , , The three-phase conductors A, B, and C are arranged parallel to each other along the crossarm direction, with y-axis coordinates of -5m, 0m, and 5m respectively. The specific pairings and coordinates are as follows: The coordinates are (-5m, 15m, 0m). The coordinates are (-5m, 10m, 0m), both are located at y=-5m, strictly corresponding to and monitoring phase A conductor. ; The coordinates are (0m, 15m, 0m). The coordinates are (0m, 10m, 0m), both are located at y=0m, strictly corresponding to and monitoring the B-phase conductor. ; The coordinates are (5m, 15m, 0m). The coordinates are (5m, 10m, 0m), both are located at y=5m, and the C-phase conductor is strictly correlated and monitored. ; That is, sensor and Mapping phase A, and Mapping phase B, and By mapping phase C, all sensors are aligned with the vertical projection plane of the target conductor, thus realizing the spatial vertical mapping decoupling logic.

[0031] S14: Three pairs of magnetic sensors synchronously collect the superimposed modulus components of the horizontal magnetic induction intensity and the superimposed modulus components of the vertical magnetic induction intensity of their respective phase conductors to form multidimensional spatial magnetic field data.

[0032] Because of the nonlinear overlap characteristic of the alternating magnetic field in a three-phase AC system, the magnetic field received by any sensor does not originate solely from the target phase conductor, but is the result of the superposition of three-phase complex phasors. The horizontal magnetic induction intensity superposition modulus component refers to the effective value of the modulus of the three-phase complex phasors at the s-th sensor after superposition in the horizontal (y-axis) direction. The vertical magnetic flux density superposition modulus component refers to the effective value of the modulus of the three-phase complex phasors at the s-th sensor after being superimposed in the vertical (z-axis) direction. The expressions for the two are as follows: ; ; in, , For each phase conductor in the th... The superimposed modulus components of the horizontal and vertical magnetic induction intensity generated at each sensor; represents the vacuum permeability; s∈{1,2,…,6} represent the six independent TMR sensors arranged on the tower. For the first The anchorage coordinates of the phase conductor in the y-axis direction; Let be the AC complex phasor transient load current flowing through the p-th phase conductor, which includes a strict 120° electrical phase difference; , These are the horizontal and vertical coordinates of the s-th sensor in the static coordinate system, respectively. Let be the spatial horizontal coordinate of the p-th phase conductor at time t. To take the absolute value.

[0033] Thus, the six sensors each output superimposed modulus components in the horizontal and vertical directions, which together constitute the 12-dimensional original magnetic field components, namely the multidimensional spatial magnetic field data.

[0034] S2: Based on the spatial coordinates of each magnetic sensor in the dual-layer magnetoresistive sensor array, construct the electromagnetic physical prior ratio sequence of the three-phase conductor.

[0035] The electromagnetic physics prior ratio sequence refers to the sequence composed of the electromagnetic physics prior ratios of phases A, B, and C. The electromagnetic physical prior ratio refers to the ratio of the magnetic induction intensity amplitudes measured by two layers of magnetic sensors directly below the same phase conductor. Its physical meaning is a dimensionless spatial geometric parameter that has been stripped of transient load current and is determined only by the geometric pole distance between the sensor and the conductor, serving as a physical prior feature of the deep network.

[0036] In some embodiments, S2 further includes S21 to S24. S21: For any phase conductor, based on the Biot-Savart law under the assumption of a two-dimensional cross-section of a long straight conductor, calculate the amplitude of the magnetic flux density measured by the upper magnetic sensor of that phase conductor. ; in, The magnitude of the magnetic flux density measured by the i-th sensor located directly below the p-th phase conductor. is the vacuum permeability constant; Let t be the effective value of the transient load current flowing through the high-voltage overhead transmission line at time t, which includes the high-frequency jump component caused by grid dispatching or sudden changes in end load; π is the mathematical constant pi. For a moment No. Phase conductor to the upper layer The absolute vertical pole distance of each sensor is calculated as follows, since the sensor and the wire are on the same vertical cross-section: ; H represents the vertical height of the upper magnetoresistive sensor array from the horizontal ground; H represents the static suspension height of the transmission line at the insulator hanging point. Let be the absolute sag value at mid-span of the p-th phase conductor at time t; To take the absolute value, the actual spatial vertical height coordinates of the p-th phase conductor at time t. satisfy: ; The absolute sag value at mid-span refers to the vertical sag of the p-phase conductor at the mid-span position (lowest point) relative to the insulator hanging point.

[0037] S22: Calculate the amplitude of the magnetic flux density measured by the magnetic sensor in the lower layer of the conductor in this phase: ; in, The magnitude of the magnetic flux density measured by the j-th sensor located directly below the p-th phase conductor; For a moment No. Phase conductor to the lower layer The absolute vertical pole distance of each sensor is calculated as follows: ; This represents the vertical height of the lower magnetoresistive sensor array above the horizontal ground.

[0038] S23: Based on the magnetic flux density amplitudes measured by the upper and lower magnetic sensors of this phase conductor, construct the electromagnetic physics prior ratio: ; in, The a priori ratio of the electromagnetic physics of the space magnetic field; This is a smoothing constant used to prevent overflow in the denominator calculation or the occurrence of singular zero values; for example, it is usually taken as a value of... As can be seen from the above equation, when the amplitudes of the upper and lower magnetic induction intensities are divided, the transient load current shared by the numerator and denominator is... With vacuum permeability Common terms such as π (pi) are completely canceled out; therefore... Transformed into being subject only to absolute sag With a known constant height , The dimensionless a priori characteristics of H-control make it immune to electrical interference caused by power grid dispatch or lightning surges, thus eliminating transient surge currents from a physical mechanism perspective.

[0039] S24: Calculate the electromagnetic physical prior ratios of phases A, B, and C respectively, to obtain the electromagnetic physical prior ratio sequence composed of the respective electromagnetic physical prior ratios of the three phases. .

[0040] S3: The electromagnetic physics prior ratio sequence is concatenated with the original magnetic field components of the multidimensional spatial magnetic field data to construct a spatiotemporal feature matrix. The spatiotemporal feature matrix is ​​then standardized to obtain standardized spatiotemporal features.

[0041] The raw magnetic field components refer to the horizontal and vertical magnetic field components output by each magnetic sensor in multidimensional spatial magnetic field data, which contain three-phase nonlinear phase crosstalk information. For example, the 12-dimensional raw magnetic field components of 6 sensors. A multidimensional feature vector is formed by concatenating the electromagnetic physics prior ratio sequence with the original magnetic field components using a high-dimensional tensor. The spatiotemporal feature matrix refers to the global multidimensional feature vector X(t) constructed by concatenating the electromagnetic physics prior ratio sequence with the original magnetic field components; because it contains both the prior ratios reflecting spatial geometric relationships and the original magnetic field components that change over time, it is called the spatiotemporal feature matrix. Standardization refers to performing a normalization operation (Z-Score standardization) on each dimension of the spatiotemporal feature matrix, removing the mean and dividing by the standard deviation, so that each dimension has zero mean and unit variance; the standardized spatiotemporal features are the features obtained after the above standardization process. .

[0042] In some embodiments, S3 further includes S31 to S33. S31: The electromagnetic physical prior ratios of the three phases in the electromagnetic physical prior ratio sequence are spliced ​​with the horizontal and vertical magnetic field components of the six magnetic sensors in a preset order to form a fifteen-dimensional spatiotemporal feature matrix.

[0043] The preset order refers to a pre-agreed, fixed order in which the features of each dimension are arranged into feature vectors; this order only needs to be consistent between the training and inference phases. For example, the order could be: first the three comparison values, then the horizontal and vertical components of each sensor. Based on this, the fifteen-dimensional spatiotemporal feature matrix X(t) is: ; That is, the N=15-dimensional global multidimensional spatiotemporal feature matrix is ​​obtained by splicing the 3-dimensional electromagnetic physics prior ratio sequence with the 12-dimensional original magnetic field components.

[0044] S32: Divide the spatiotemporal feature matrix into training and test sets according to a preset ratio. The preset ratio refers to the pre-defined proportion of samples to be divided into training and test sets. For example, a 70:30 ratio for the training and test sets. The training set is used to train the network parameters, and the test set is used to verify the model's tracking accuracy.

[0045] S33: For each spatiotemporal feature, standardize it using the sample mean and standard deviation of that feature on the training set to obtain the standardized spatiotemporal feature: ; in, The d-th eigenvalue is the standardized eigenvalue. For the original d-th eigenvalue, This represents the mean of the samples in the training set for this feature dimension. This represents the standard deviation of the feature dimension on the training set. After standardization, the mean of all features is 0 and the variance is 1, which can accelerate the convergence of subsequent deep neural networks.

[0046] It should be noted that, to maximize the capture of line status characteristics, a triaxial magnetoresistive sensor and supporting hardware data link system can be deployed. An extremely low-noise TMR tunnel magnetoresistive sensor should be selected, according to... Figure 1 The system employs a dual-layer mapping coordinate deployment; the sensor array is driven by a high-precision independent power supply module, synchronously acquiring the raw, weak voltage signals from each node, with a sampling frequency set to no less than 100Hz to capture the complete power frequency magnetic field envelope; the acquired analog signals are amplified by hardware and then processed using a built-in 5th-order Butterworth low-pass filter algorithm. ; in, Let the frequency response transfer function be that of the Butterworth low-pass filter. This represents the time-varying angular frequency variable of the input signal. This is the cutoff angular frequency of the filter. This filter is used to filter out stray radio frequency interference and high-frequency white noise from corona discharge in space.

[0047] S4: Input the standardized spatiotemporal features into the pre-spatial demixing network to filter out three-phase phasor crosstalk and output clean spatial features.

[0048] A pre-specified spatial demixing network is a feedforward neural network placed before the time-series prediction network to filter out three-phase AC phasor spatial crosstalk. For example, it can be constructed using a multilayer perceptron (MLP), specifically a two-layer fully connected network. This network implicitly fits the inverse matrix operation of the superposition of three-phase complex phasor spaces, demixing and separating the superimposed three-phase magnetic fields. Three-phase phasor crosstalk refers to the spatial overlap of alternating magnetic fields caused by the inherent 120° complex phase difference of the three-phase AC current in a real high-voltage transmission line. This causes any phase sensor to receive the target phase magnetic field while also experiencing complex phasor spatial crosstalk from the diffusion of the other two phase magnetic fields. Clean spatial features refer to the high signal-to-noise ratio spatial feature sequence output after the pre-specified spatial demixing network filters out the three-phase complex phasor spatial crosstalk. It has already removed the nonlinear electromagnetic overlap between phases B and C relative to phase A (and between each phase). The simple geometric ratio cannot eliminate the nonlinear error introduced by the overlap of three complex phasors, so this network is needed to further clean up phase crosstalk.

[0049] In some embodiments, the pre-spatial demixing network is a two-layer fully connected network, and S4 further includes S41 to S42. S41: The standardized spatiotemporal features are input into the first fully connected layer, activated by Gaussian error linear units, and output as intermediate transition abstract features. ; in, The intermediate transition abstract feature output by the first fully connected layer is the transition abstract feature that filters out some neighboring loop clutter; GELU is the Gaussian error linear unit activation function, which has probabilistic smoothing properties and is used to suppress the gradient vanishing phenomenon caused by zero crossover points in the input matrix. and The weights and biases of the first-level implicit space mapping; The input is standardized spatiotemporal features.

[0050] The first fully connected layer refers to the first linear transformation (dimensionality reduction) layer of the pre-spatial unmixing network. The first implicit spatial mapping refers to the nonlinear mapping process that linearly transforms and reduces the 15-dimensional normalized input before GELU activation. , That is, the weight matrix and bias vector of the mapping.

[0051] S42: Input the intermediate transition abstract features into the second fully connected layer, activate it through a Gaussian error linear unit, and output the pure spatial features: ; in, This represents the clean spatial characteristics output by the second fully connected layer. and These are the weights and biases for the second layer of feature dimensionality reduction.

[0052] The second fully connected layer refers to the second linear feature recombination transformation layer of the preceding spatial demixing network. For example, the second fully connected layer outputs a 16-dimensional pure abstract spatial feature sequence. It should be noted that the terms appearing in the claims of this invention... Both the spatiotemporal feature matrix and the standardized spatiotemporal feature input vector are fed into the pre-spatial unmixing network; that is, S4 uses the standardized spatiotemporal features... As input.

[0053] S5: Apply the first-order difference operator to the pure space features for stabilization to obtain the micro-kinematic velocity features.

[0054] First-order difference operator This refers to the operator that calculates the difference between adjacent moments in a continuous time series. The stabilization process refers to using a first-order difference operator to transform non-stationary absolute position domain data that monotonically accumulates over time into stationary velocity domain data that remains within a small numerical range, thereby avoiding the limit of absolute position extrapolation. The micro-kinematic velocity characteristic refers to the velocity vector obtained after first-order difference stabilization, which represents the gradient of small changes in the pure spatial characteristics between adjacent time steps. It describes the microscopic kinematics (velocity) evolution of the conductor rather than its absolute position.

[0055] The principle of introducing the first-order difference operator is that the sag of transmission conductors usually exhibits a long-period, monotonically decreasing evolution, and the absolute position... It will continuously exceed the maximum boundary of the historical training distribution (OOD), causing traditional networks to saturate and collapse; while regardless of the absolute sag The depth to which it is probed, and the rate of change of small intervals between adjacent moments. The value remained consistently small, thus eliminating the root cause of the saturation collapse from the underlying mechanism.

[0056] In some embodiments, S5 further includes S51 to S53. S51: Subtract the pure space features at the current moment from the pure space features at the previous moment to obtain the micro-evolution rate vector: ; Among them, the characteristics of pure space at the current moment are: The characteristics of the pure space in the previous moment were Micro-evolution rate vector It represents the gradient sequence of tiny changes in the pure spatial feature sequence between adjacent time steps, completely immune to the upper limit of absolute displacement values.

[0057] S52: Subtract the absolute sag value at the current moment from the absolute sag value at the previous moment to obtain the true physical minute downward displacement rate: ; Wherein, the absolute sag value at the current moment is The absolute sag value at the previous moment is Both are derived from the historical offline calibration absolute sag sequence of synchronous matching; physical minute downward shift rate This represents the actual minute downward displacement rate of the high-voltage overhead conductor sag between adjacent time steps, serving as the true rate benchmark (label) for network training.

[0058] S53: Using the micro-evolution rate vector as the micro-kinematic rate feature, the non-stationary absolute position domain feature is mapped to the stationary velocity domain.

[0059] Among them, the absolute position domain feature refers to the non-stationary feature characterized by the absolute spatial position of the conductor (absolute sag); the velocity domain refers to the stationary feature domain characterized by the rate of change between adjacent moments. Through S53, no matter how unknown the extreme region to which the absolute sag of the conductor descends, its first-order differential rate remains stable, thus breaking the limit of extrapolation ceiling of deep networks.

[0060] S6: Construct the micro-kinematic rate features into a micro-kinematic sliding prediction window sequence, input it into the differential attention long short-term memory network, and optimize the network parameters through a composite physical constraint total loss function to obtain the trained differential attention long short-term memory network.

[0061] Among them, the microkinematic sliding prediction window sequence (also known as the microkinematic sliding time window) A **temporal window sequence (L)** refers to a dynamic window sequence consisting of micro-kinematic rate features at L consecutive time points, sliding along the time axis with a preset temporal receptive field length L. It includes the gradients of all continuous micro-kinematic rate changes from time (t-L+1) to time t. A **Diff-ALSTM** network is a deep temporal network that takes first-order differencing micro-kinematic rate features as input and introduces a temporal attention mechanism on top of a Long Short-Term Memory (LSTM) network. Logically, it consists of a recurrent encoding layer (LSTM cells) for temporal feature extraction, an attention mechanism layer for dynamic weight adaptive allocation, and a fully connected regression decoder head (terminal fully connected regression layer) for continuous value rate mapping. A **composite physical constraint total loss function** is a weighted total loss function that simultaneously includes the prediction mean squared error loss, the temporal smoothness physical constraint loss, and the L2 structure regularization loss. A trained **Diff-ALSTM** network is a Diff-ALSTM model obtained by iteratively optimizing the weight parameters through backpropagation using the composite physical constraint total loss function and then converging.

[0062] In some embodiments, the input to the differential attention long short-term memory network is a sequence of microkinematic sliding prediction windows, and the output is a small downward shift rate prediction value, including S61 to S65.

[0063] S61: Construct a sequence of micro-kinematic sliding prediction windows with a preset step size: ; in, For micro-kinematic sliding time window; and These are the gradient sequences of small changes in the pure spatial feature sequence between adjacent time steps t-L+1 and t; t is the time variable; L is the preset step size, i.e., the temporal receptive field length, representing the number of consecutive time steps contained within a single sliding window, for example, L=10. The preset step size is the temporal receptive field length L mentioned above.

[0064] S62: Input the microkinematic sliding prediction window sequence into the long short-term memory cells of the differential attention long short-term memory network time-by-time, and extract the hidden states of the sequence time-by-time: ; in, Let τ be the temporal hidden state vector at each time step τ within the window, where τ∈[t-L+1,t]. is the output gate activation value of the long short-term memory cell at time step τ, used to control the proportion of the cell state output to the hidden state; ⊙ is the Hadamard product, which is the element-wise multiplication operation of corresponding vectors or matrices; tanh(·) is the hyperbolic tangent function. This refers to the cell state of long short-term memory cells at time step τ, which is used to carry and transmit long-term memory information.

[0065] Long Short-Term Memory (LSTM) cells refer to memory units in the recurrent coding layer that recursively extract temporal hidden states based on the forget gate, input gate, and output gate operation mechanism. The sequential hidden state is the aforementioned temporal hidden state vector. As an example illustrating the departmental evolution mechanism within long short-term memory cells, it strictly follows: ; ; ; ; ; ; in, , and These are the forget gate, input gate, and output gate for time step τ, respectively. The candidate cell state is represented by σ, which is the Sigmoid activation function. , , , and , , , These are the weight matrix and bias vector for each gate, respectively; Indicates the previous hidden state. Input with the current time step splicing.

[0066] S63: Feed the sequential hidden states at each time step into the temporal attention layer and calculate the dynamic weights for attention allocation at each time step: ; in, Dynamic weights are assigned to attention to characterize the importance of local feature segments at each time step to the prediction target; exp is an exponential function. T is the scoring weight vector (trainable parameters) for the attention mechanism; T is the transpose of the matrix (vector); The scoring weight matrix (trainable parameters) for the attention mechanism; is the bias vector (trainable parameter) for the attention mechanism; m is the index of the time step traversed when normalizing and summing the hidden states within the sliding window, m∈[t-L+1,t]; Let be the temporal hidden state vector at the m-th time step.

[0067] Among them, the dynamic weight of attention allocation refers to the dynamic weight coefficient obtained by normalized Softmax operation; , , All of these are trainable parameters that participate in error backpropagation updates within the attention mechanism; the denominator of the above equation represents all hidden states traversed within the sliding window. The normalized Softmax operation ensures that the sum of the weight coefficients for all time steps is 1.

[0068] S64: Weighted summation of the hidden states at each time step using attention-assigned dynamic weights to generate a high-dimensional context vector: ; in, τ is a high-dimensional context vector, which is the core vector of temporal context aggregation extracted by weighted fusion of the hidden states of each time step by attention weights; τ is each time step within the window, τ∈[t-L+1,t].

[0069] Here, the high-dimensional context vector refers to a high-dimensional feature vector that integrates information from all time steps within the window and aggregates it according to importance weights, serving as the input for regression decoding. To avoid interference with cell state... Confusion, here Specifically used to represent context vectors.

[0070] S65: Input the high-dimensional context vector into the terminal fully connected regression layer and decode to map out the predicted value of the tiny downward shift rate.

[0071] The terminal fully connected regression layer (i.e., the fully connected regression decoder head) is a fully connected layer located at the end of the network that decodes and maps high-dimensional context vectors into continuous rate values ​​through deep feedforward operations. The micro-downshift rate prediction value is the micro-downshift rate of the conductor predicted at the current time t. .

[0072] During the training of S6, the AdamW optimizer with momentum adaptive adjustment (e.g., an initial learning rate of 0.003) can be used, an early stopping strategy can be introduced, and offline parameter solidification can be completed based on the PyTorch framework until the verification accuracy meets the standard.

[0073] The total loss function for composite physical constraints is: ; ; ; ; in, This represents the total loss during network backpropagation. , and These are the weight coefficients for the first, second, and third hyperparameters, respectively, and the penalty weight coefficients for each loss sub-objective set based on engineering priors. The mean squared error loss is the micro-rate regression; M is the batch size during network training; i is the sample variable. and These are the predicted micro-downward shift rate and the actual downward shift rate for the i-th sample, respectively. This represents the actual downward shift rate of the (i-1)th sample. To smooth the constraint loss of the time-series evolution of the conductor; is the L2 structure regularization loss; w is a single weight within the set of all trainable parameters Ω of the Diff-ALSTM network; Ω is the set of trainable parameters of the Diff-ALSTM network.

[0074] Among them, the mean square error loss of micro-rate regression Used to characterize the accuracy of network rate fitting. Wire-time evolution smoothing constraint loss. This term is used to constrain the physical continuity of the macroscopic conductor sag evolution and suppress high-frequency non-physical jumps caused by accidental strong clutter. Essentially, it calculates the second-order differential derivative of the predicted rate in time (i.e., the local vertical acceleration of the conductor). Based on the prior mechanical inertia of the massive mass of the conductor spanning hundreds of meters (its falling acceleration approaches zero), it adds virtual mechanical damping to the network, forcing the model to learn the slowly varying inertial characteristics equivalent to conductor creep, thus suppressing sawtooth-like non-physical prediction jumps. L2 structure regularization loss. It is used to alleviate the overfitting of a model to a specific training distribution domain and prevent the model from getting overfitted.

[0075] S7: Collect the current real-time magnetic field data, calculate the current microkinematic sliding prediction window sequence, input the current microkinematic sliding prediction window sequence into the trained differential attention long short-term memory network, and output the predicted value of the small downward shift rate.

[0076] The current real-time magnetic field data refers to the multi-dimensional spatial magnetic field data acquired in real time by a dual-layer non-contact magnetic array during system operation. The current microkinematic sliding prediction window sequence refers to the microkinematic sliding prediction window sequence constructed by preprocessing the current real-time magnetic field data using the same methods as the training set, including standardization, spatial unmixing, and first-order difference stabilization, and then using a preset step size L. The predicted value of the minute downward displacement rate refers to the minute downward displacement rate of the conductor output by the trained differential attention long short-term memory network for the current moment. .

[0077] Specifically, the magnetic field components of six triaxial magnetic sensors, collected in real time by a double-layer array directly below the tower crossarm, are processed using the same physical ratio stripping, tensor concatenation, Z-score normalization, pre-space unmixing, and first-order difference processing as the training set. This process constructs the current micro-kinematic sliding prediction window sequence, which is then input into the trained Diff-ALSTM model at each time step to output the transient small downward displacement rate prediction value. .

[0078] S8: Based on the predicted value of the small downward rate, perform numerical integration to reconstruct the absolute sag prediction value, and execute the sag over-limit warning based on the absolute sag prediction value.

[0079] Numerical integral reconstruction refers to the process of restoring the absolute sag. The predicted absolute sag value refers to the high-fidelity mid-span absolute sag reconstruction value that is completely immune to extreme load jump interference and three-phase crosstalk. .

[0080] In some embodiments, numerical integration is performed based on the predicted value of the small downward displacement rate to reconstruct the absolute sag prediction value, including S81 to S83. S81: Taking the start or retest time of online monitoring as the origin of the integration reference, the true static absolute sag at that time is extracted as the reference sag.

[0081] Online monitoring startup refers to the initial moment when the system goes online and begins continuous monitoring; retesting time refers to the moment when offline calibration and verification are performed again during system operation; the origin of the integration reference refers to the deterministic moment that serves as the starting point for numerical integration accumulation. Static absolute sag refers to the sag at the origin of the integration reference. High-precision physical absolute sag obtained by offline optical or tensile force sensor calibration The reference sag is the static absolute sag. It serves as the deterministic origin and starting point for subsequent mathematical integration.

[0082] S82: Starting from the origin of the integral reference, the predicted values ​​of the small downward shift rate at each time point are accumulated to obtain the cumulative downward shift.

[0083] The cumulative downward displacement refers to the total absolute physical displacement increment obtained by accumulating the predicted small downward displacement rates at each moment. .

[0084] S83: Add the baseline sag to the cumulative downward shift to reconstruct the absolute sag prediction value at the current moment: ; in, The high-fidelity mid-span absolute sag prediction value for the current moment t, which is completely immune to complex three-phase crosstalk and extreme alternating load impacts; The origin of the integration reference The reference sag at the location; This represents the cumulative downward shift.

[0085] In some embodiments, a sag over-limit warning is executed based on the absolute sag prediction value, including S84 to S86, constituting a three-level warning mechanism with time-series self-checking capability. S84: When the absolute sag prediction value exceeds a preset percentage of the over-limit value and the small downward movement rate prediction value increases monotonically in a positive direction, a first-level warning is output.

[0086] The threshold value refers to the pre-set absolute sag threshold used to determine whether the sag exceeds the limit; the preset percentage refers to the percentage threshold set relative to the threshold value to trigger a Level 1 warning. For example, 85%. That is, when the absolute sag prediction value... Exceeding 85% of the limit and with a small downward shift rate When the sag increases monotonically and positively, it is determined to be the initial stage of sag increase, and a level one warning is issued (continue to monitor).

[0087] S85: When a single-point sag increment suddenly jumps, but the subsequent continuous sampling points have a small downward shift rate that returns to zero or oscillates, a level 2 warning is output.

[0088] The single-point sag increment refers to the increment of the absolute sag prediction value between two adjacent sampling points; subsequent consecutive sampling points refer to a series of consecutive sampling times after the jump point. That is, when the single-point sag increment experiences a sudden jump, but the subsequent consecutive sampling points show a small downward shift rate... When the signal returns to zero or oscillates, it is identified as strong transient non-physical electromagnetic pulse artifact interference, which is filtered out by internal time delay, and a level 2 warning (interference investigation) is output.

[0089] S86: When the absolute sag prediction value exceeds the safety design limit and the confidence level reaches the preset threshold after artifact removal, a level 3 warning is output.

[0090] The safety design limit refers to the upper limit of absolute sag determined based on the design requirements for safe clearance distance between the line and the ground and to crossings; confidence level refers to the measure of the reliability of the limit violation judgment result after smoothing and artifact removal; the preset threshold refers to the confidence threshold used to trigger the Level 3 warning. That is, when the predicted absolute sag value... If the safety design limit is exceeded and the confidence level reaches the preset threshold (extremely high confidence level) after smoothing and eliminating artifacts, a level 3 warning (emergency alarm, lockout level alarm) will be immediately output.

[0091] The present invention also provides a transmission line sag monitoring system based on a non-contact magnetic array, comprising a double-layer non-contact magnetic array, a power supply module, a signal conditioning module, a feature processing module, a sag prediction module, and an early warning module.

[0092] A dual-layer non-contact magnetic array is deployed directly below the crossarm of the tower, and is divided into upper and lower layers along the tower height, with multiple magnetic sensors installed in each layer. Its deployment method and coordinate mapping relationship are as described in S1 above. Figure 1 As stated above.

[0093] The power supply module is connected to a dual-layer non-contact magnetic array to drive each magnetic sensor to synchronously acquire raw magnetic field signals. The raw magnetic field signals refer to the unprocessed, weak voltage signals directly acquired by each magnetic sensor. For example, the power supply module is a high-precision independent power supply module that drives each node to synchronously acquire signals at a sampling frequency of at least 100Hz.

[0094] The signal conditioning module is connected to a dual-layer non-contact magnetic array to process the original magnetic field signal and output multi-dimensional spatial magnetic field data. For example, the signal conditioning module amplifies the original magnetic field signal in hardware and filters out radio frequency interference and high-frequency white noise from corona discharge using a 5th-order Butterworth low-pass filter algorithm to output multi-dimensional spatial magnetic field data.

[0095] The feature processing module is connected to the signal conditioning module. It constructs an electromagnetic physics prior ratio sequence for the three-phase conductors based on the spatial coordinates of each magnetic sensor in the dual-layer magnetoresistive sensor array. This sequence is then concatenated with the original magnetic field components of the multi-dimensional spatial magnetic field data to construct a spatiotemporal feature matrix. The spatiotemporal feature matrix is ​​standardized to obtain standardized spatiotemporal features. These standardized features are then input into a pre-processing spatial demixing network to filter out three-phase phasor crosstalk and output clean spatial features. Finally, a first-order difference operator is applied to the clean spatial features for stabilization, yielding micro-kinematic velocity features. The processing performed by this module corresponds to the aforementioned steps S2 to S5.

[0096] The sag estimation module is connected to the feature processing module. It constructs a microkinematic velocity feature sequence into a microkinematic sliding prediction window sequence, inputs it into a differential attention long short-term memory network (LSTM), and optimizes the network parameters using a composite physical constraint total loss function to obtain a trained LSTM network. It then collects real-time magnetic field data, calculates the current microkinematic sliding prediction window sequence, inputs it into the trained LSTM network, and outputs a small downward displacement rate prediction value. Based on this small downward displacement rate prediction value, it performs numerical integration reconstruction to obtain the absolute sag prediction value. The processing performed by this module corresponds to the reconstruction part in S6 to S8 mentioned above.

[0097] The early warning module is connected to the sag prediction module and is used to perform sag limit warning based on the absolute sag prediction value. It performs the three-level early warning mechanism as described in S84 to S86 above.

[0098] The aforementioned system can be further divided into the following hierarchical architecture: a bottom-level hardware multi-dimensional perception and preprocessing module, a physical prior adaptive demixing and cleaning module, a micro-kinematic feature extraction and deep network inference module, and an integral reconstruction and digital scheduling output module. These modules respectively implement the aforementioned data acquisition and preprocessing, physical prior ratio demixing, micro-kinematic difference and deep inference, and integral reconstruction and scheduling output functions, providing extremely low-latency safe clearance distance data support services for the power grid transmission monitoring and management platform and upper-level situational awareness applications.

[0099] This invention generates a dataset through co-simulation using COMSOL and MATLAB, and establishes a three-dimensional electromagnetic field model using COMSOL to simulate absolute sag. Monotonically and slowly increasing, accompanied by load current Under complex operating conditions with drastic and random fluctuations, a total of 20,000 time-series samples were generated using MATLAB, and data from 6 sensors were extracted. , The components form 12-dimensional original magnetic field features, which are divided into training and test sets in a 70:30 ratio.

[0100] Extreme electromagnetic storm conditions were simulated by applying a transient load current with disordered and violent jumps between 350A and 750A. The obtained dynamic tracking graph shows that the real reference absolute sag exhibits a rigorous extrapolation test distribution with a long-period monotonic drop from 10.6 meters to an extreme 12.0 meters. The absolute sag predicted by the reconstructed Diff-ALSTM perfectly overlaps with the reference curve, without saturation collapse degenerating into a straight line, and without any abnormal jumps or glitches due to current interference. The global determination coefficient is [not specified]. Breaking the 0.9995 barrier. Through the underlying geometric ratio method of this invention, the absolute error of the algorithm reconstruction is extremely stable and completely independent of the magnitude of the destructive transient load surge current.

[0101] In summary, this invention utilizes a dual-layer high-precision magnetic sensor array with strictly aligned mapping of phases A, B, and C to acquire three-dimensional alternating magnetic field data. Combined with a dimensionless physical ratio stripping algorithm and Z-Score preprocessing at the bottom layer, and through a Diff-ALSTM deep network embedding a first-order difference operator of micro-kinematics and attention-based weighting, and combined with a self-calibrating three-level early warning closed-loop mechanism for Euler discrete integral reconstruction, it effectively overcomes three major engineering challenges: strong alternating load jump impact, multi-phase nonlinear coupling interference, and weak extrapolation generalization of traditional time-series architectures. It has significant technical advantages such as convenient pure passive installation on site, strong resistance to current surge interference, and high absolute reconstruction accuracy of limit extrapolation.

[0102] 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 monitoring the sag of transmission lines based on a non-contact magnetic array, characterized in that, include: S1: Multidimensional spatial magnetic field data are collected by a double-layer magnetoresistive sensor array deployed directly below the crossarm of the tower; the double-layer magnetoresistive sensor array is divided into upper and lower layers along the height of the tower, and multiple magnetic sensors are arranged in each layer. S2: Based on the spatial coordinates of each magnetic sensor in the dual-layer magnetoresistive sensor array, construct the electromagnetic physical prior ratio sequence of the three-phase conductor; S3: The electromagnetic physics prior ratio sequence is concatenated with the original magnetic field components of the multidimensional spatial magnetic field data to construct a spatiotemporal feature matrix. The spatiotemporal feature matrix is ​​then standardized to obtain standardized spatiotemporal features. S4: Input the standardized spatiotemporal features into the pre-spatial demixing network to filter out three-phase phasor crosstalk and output clean spatial features; S5: Apply the first-order difference operator to the pure space features for stabilization to obtain the micro-kinematic velocity features; S6: Construct the micro-kinematic rate features into a micro-kinematic sliding prediction window sequence, input it into the differential attention long short-term memory network, and optimize the network parameters through the composite physical constraint total loss function to obtain the trained differential attention long short-term memory network; S7: Collect the current real-time magnetic field data, calculate the current microkinematic sliding prediction window sequence, input the current microkinematic sliding prediction window sequence into the trained differential attention long short-term memory network, and output the predicted value of the small downward shift rate. S8: Based on the predicted value of the small downward rate, perform numerical integration to reconstruct the absolute sag prediction value, and execute the sag over-limit warning based on the absolute sag prediction value.

2. The method for monitoring the sag of transmission lines based on a non-contact magnetic array according to claim 1, characterized in that, S1 includes: S11: Establish a spatial rectangular coordinate system with the ground projection of the midpoint of the tower crossarm as the origin; the x-axis of the spatial rectangular coordinate system is parallel to the direction of the conductor, the y-axis is along the direction of the tower crossarm and perpendicular to the direction of the conductor, and the z-axis is vertically upward along the tower height. S12: Deploy an upper sensor array and a lower sensor array in a spatial rectangular coordinate system; the upper sensor array is at a first preset height above the ground, and the lower sensor array is at a second preset height above the ground, wherein the first preset height is greater than the second preset height; S13: Deploy magnetic sensors according to the vertical mapping logic of independent pairing of each phase conductor, thereby forming three pairs of magnetic sensors corresponding to phase A, phase B and phase C respectively; the vertical mapping logic requires that the magnetic sensors in the upper sensor array correspond one-to-one with the magnetic sensors in the lower sensor array, and be located on the same y-axis coordinate and face the vertical projection plane of the same phase conductor. S14: The multidimensional spatial magnetic field data is formed by the simultaneous acquisition of the horizontal magnetic induction intensity superimposed modulus component and the vertical magnetic induction intensity superimposed modulus component of the corresponding phase conductor by three pairs of magnetic sensors.

3. The method for monitoring the sag of transmission lines based on a non-contact magnetic array according to claim 2, characterized in that, The S2 includes: S21: For any phase conductor, calculate the amplitude of the magnetic flux density measured by the magnetic sensor on the upper layer of that phase conductor: ； in, For being in the first The upper layer directly below the phase conductor The amplitude of magnetic flux density measured by each sensor; is the vacuum permeability constant; For the first The effective value of transient load current flowing through high-voltage overhead transmission lines at any given moment; Pi; The vertical height of the upper magnetoresistive sensor array from the horizontal ground surface; This refers to the static suspension height of the transmission line at the insulator hanging point; For the first Phase conductor at time The absolute sag value at mid-span; To take the absolute value; S22: Calculate the amplitude of the magnetic flux density measured by the magnetic sensor in the lower layer of the conductor in this phase: ； in, For being in the first The lower layer directly below the phase conductor The amplitude of magnetic flux density measured by each sensor; The vertical height of the lower magnetoresistive sensor array from the horizontal ground surface; S23: Based on the magnetic flux density amplitudes measured by the upper and lower magnetic sensors of this phase conductor, construct the electromagnetic physics prior ratio: ； in, The a priori ratio of the electromagnetic physics of the space magnetic field; It is the smoothing minimum constant; S24: Calculate the electromagnetic physical prior ratios of phases A, B, and C respectively to obtain the electromagnetic physical prior ratio sequence composed of the electromagnetic physical prior ratios of the three phases.

4. The method for monitoring the sag of transmission lines based on a non-contact magnetic array according to claim 1, characterized in that, The S3 includes: S31: The electromagnetic physical prior ratios of the three phases in the electromagnetic physical prior ratio sequence are spliced ​​together with the horizontal and vertical magnetic field components of the six magnetic sensors in a preset order to form a 15-dimensional spatiotemporal feature matrix. S32: Divide the spatiotemporal feature matrix into a training set and a test set according to a preset ratio; S33: For each spatiotemporal feature, standardize it using the sample mean and sample standard deviation of that feature on the training set to obtain the standardized spatiotemporal feature.

5. The method for monitoring the sag of transmission lines based on a non-contact magnetic array according to claim 1, characterized in that, The S5 includes: S51: Subtract the pure space features at the current moment from the pure space features at the previous moment to obtain the micro-evolution rate vector. S52: Subtract the absolute sag value at the current moment from the absolute sag value at the previous moment to obtain the true physical minute downward displacement rate. S53: Using the micro-evolution rate vector as the micro-kinematic rate feature, the non-stationary absolute position domain feature is mapped to the stationary velocity domain.

6. The method for monitoring the sag of transmission lines based on a non-contact magnetic array according to claim 1, characterized in that, The differential attention long short-term memory network takes a sequence of microkinematic sliding prediction windows as input and outputs a small downward shift rate prediction value, including: S61: Construct a sequence of micro-kinematic sliding prediction windows with a preset step size: ； in, For micro-kinematic sliding time window; and The pure space feature sequences are respectively in and The gradient sequence of small changes between adjacent time steps; For time variables; The preset temporal receptive field length; S62: Input the microkinematic sliding prediction window sequence into the long short-term memory cells of the differential attention long short-term memory network time-by-time, and extract the hidden states of the sequence time-by-time: ； in, For each time step within the window The temporal hidden state vector; The output gate activation value of the long short-term memory cell at time step τ; For Hadamah accumulation; It is the hyperbolic tangent function; The cellular state of long short-term memory cells at time step τ; S63: Feed the sequential hidden states at each time step into the temporal attention layer and calculate the dynamic weights for attention allocation at each time step: ； in, Dynamic weights are assigned to attention; exp is an exponential function; Let T be the scoring weight vector for the attention mechanism; T is the transpose of the matrix. The scoring weight matrix for the attention mechanism; is the bias vector for the attention mechanism; m is the index of the time step traversed when normalizing and summing the hidden states within the sliding window; Let be the temporal hidden state vector at the m-th time step; S64: Weighted summation of the hidden states at each time step using attention-assigned dynamic weights to generate a high-dimensional context vector: ； in, It is a high-dimensional context vector; For each time step within the window; S65: Input the high-dimensional context vector into the terminal fully connected regression layer and decode to map out the predicted value of the tiny downward shift rate.

7. The method for monitoring the sag of transmission lines based on a non-contact magnetic array according to claim 1, characterized in that, The total loss function of the composite physical constraints is: ； ； ； ； in, This represents the total loss during network backpropagation. , and These are the weight coefficients for the first, second, and third hyperparameters, respectively. The mean squared error loss is the micro-rate regression; M is the batch size during network training; i is the sample variable. and The first The predicted micro-downshift rate and the actual downshift rate for each sample; For the first The true downward shift rate of each sample; To smooth the constraint loss of the time-series evolution of the conductor; For L2 structure regularization loss; The set of all trainable parameters for the Diff-ALSTM network Individual weights within; This is the set of trainable parameters for the Diff-ALSTM network.

8. The method for monitoring the sag of transmission lines based on a non-contact magnetic array according to claim 1, characterized in that, The process of numerically integrating and reconstructing based on the predicted value of a small downward drift rate to obtain the predicted value of absolute sag includes: S81: Using the start or retest time of online monitoring as the origin of integration, extract the true static absolute sag at that time as the reference sag. S82: Starting from the origin of the integral reference, the predicted values ​​of the small downward shift rate at each time point are accumulated to obtain the cumulative downward shift; S83: Add the baseline sag to the cumulative downward displacement to reconstruct the absolute sag prediction value at the current moment.

9. The method for monitoring the sag of transmission lines based on a non-contact magnetic array according to claim 1, characterized in that, The method of performing sag limit exceedance warning based on absolute sag prediction value includes: S84: When the absolute sag prediction value exceeds the preset percentage of the over-limit value and the small downward rate prediction value increases monotonically in a positive direction, output a first-level warning; S85: When a single-point sag increment suddenly jumps, but the subsequent continuous sampling points have a small downward movement rate that returns to zero or oscillates, a level 2 warning is output; S86: When the absolute sag prediction value exceeds the safety design limit and the confidence level reaches the preset threshold after artifact removal, a level 3 warning is output.

10. A transmission line sag monitoring system based on a non-contact magnetic array, characterized in that, It includes a dual-layer non-contact magnetic array, a power supply module, a signal conditioning module, a feature processing module, a sag prediction module, and an early warning module; The dual-layer non-contact magnetic array is deployed directly below the crossarm of the tower and is divided into upper and lower layers along the height of the tower, with multiple magnetic sensors installed in each layer. The power supply module is connected to the double-layer non-contact magnetic array and is used to drive each magnetic sensor to synchronously acquire the original magnetic field signal; The signal conditioning module is connected to the dual-layer non-contact magnetic array and is used to process the original magnetic field signal and output multi-dimensional spatial magnetic field data. The feature processing module is connected to the signal conditioning module and is used to construct an electromagnetic physical prior ratio sequence of three-phase conductors based on the spatial coordinates of each magnetic sensor in the dual-layer magnetoresistive sensor array; the electromagnetic physical prior ratio sequence is spliced ​​with the original magnetic field components of the multi-dimensional spatial magnetic field data to construct a spatiotemporal feature matrix; the spatiotemporal feature matrix is ​​standardized to obtain standardized spatiotemporal features. The standardized spatiotemporal features are input into the pre-spatial demixing network to filter out three-phase phasor crosstalk and output clean spatial features. The pure space features are stabilized by applying a first-order difference operator to obtain the micro-kinematic velocity features. The sag estimation module is connected to the feature processing module. It is used to construct a microkinematic sliding prediction window sequence from the microkinematic rate features, input the sequence into the differential attention long short-term memory network, and optimize the network parameters through a composite physical constraint total loss function to obtain a trained differential attention long short-term memory network. It collects the current real-time magnetic field data, calculates the current microkinematic sliding prediction window sequence, inputs the current microkinematic sliding prediction window sequence into the trained differential attention long short-term memory network, and outputs the predicted value of the small downward displacement rate. Based on the predicted value of the small downward displacement rate, it performs numerical integration and reconstruction to obtain the predicted value of the absolute sag. The early warning module is connected to the sag calculation module and is used to perform sag limit exceeding early warning based on the absolute sag prediction value.

Images