ADSS optical cable potential measurement method and device of improved ViT network based on physical symbol fusion and medium

Abstract

The invention discloses an ADSS optical cable potential measurement method and device of an improved ViT network based on physical symbol fusion and a medium, and the method comprises the steps: collecting three-dimensional electric field distribution parameters of an ADSS optical cable, synchronously obtaining environment parameters, and constructing a multi-channel space-time tensor based on an environment-electric field coupling factor; inputting the multi-channel space-time tensor into the improved ViT network to generate a hidden space representation vector; a physical symbol fusion layer is introduced on the basis of the improved ViT network, and an improved ViT network based on physical symbol fusion is constructed; a training process is constrained through a neural differential equation gradient stabilization technology, a three-dimensional dynamic potential distribution field phi (x, y, z, t) is output, and an electric corrosion risk level Risk is calculated in real time based on potential gradient space-time characteristics. According to the invention, multi-modal environmental parameters are fused, physical priori knowledge is embedded, the dynamic risk deduction capability is provided, and high-accuracy and high-precision electric field measurement can be completed in a high-humidity environment.

Description

Technical Field

[0001] This invention relates to intelligent monitoring of power systems, and more particularly to an ADSS optical cable potential measurement method, device, and medium based on an improved ViT network using physical symbol fusion. Background Technology

[0002] ADSS (All-Dielectric Self-Supporting Optical Cable), as a critical infrastructure in power communication systems, is widely deployed in high-voltage transmission line corridors, serving the important function of sharing both optical communication and power transmission corridors. Operating long-term in high-voltage, strong electric field environments, its outer sheath is susceptible to electrolytic corrosion due to continuous exposure to power frequency electric fields and complex environmental factors (such as rain, fog, and corona discharge). This leads to a decrease in the cable's mechanical strength and premature failure, seriously threatening the reliability of power communication networks.

[0003] Currently, ADSS optical cable potential monitoring mainly relies on contact electric field sensors or indirect inference methods, which have the following limitations:

[0004] (1) High environmental sensitivity: Traditional sensors are easily affected by environmental parameters such as temperature, humidity and air pressure. In particular, the dielectric constant drifts under high temperature and high humidity conditions, resulting in electric field measurement distortion.

[0005] (2) Insufficient spatial resolution: Existing technologies are unable to accurately capture the electric field distortion characteristics of key areas such as the periphery of insulators and the ends of fittings;

[0006] (3) Lack of physical mechanism: Although pure data-driven methods (such as convolutional neural networks) can extract spatial features, they lack the embedding of the basic conservation law of electromagnetic field and have poor generalization ability under transient conditions (such as lightning strikes and switching operations).

[0007] (4) Lagging dynamic risk warning: Traditional methods rely solely on electric field strength threshold alarms, failing to integrate spatiotemporal gradients and historical fluctuation characteristics, thus failing to achieve early prediction of corrosion paths.

[0008] There is an urgent need for an ADSS optical cable potential measurement method that can integrate multimodal environmental parameters, embed physical prior knowledge, and have dynamic risk extrapolation capabilities, in order to overcome the bottleneck of accuracy and reliability in electric field sensing under complex environments. Summary of the Invention

[0009] Technical Objective: To address the shortcomings of existing technologies, this invention discloses an ADSS optical cable potential measurement method, device, and medium based on physical symbol fusion in an improved ViT network. The aim is to solve core problems such as electric field measurement distortion, insufficient spatial sensing accuracy, and delayed dynamic risk warning in high humidity environments.

[0010] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution.

[0011] An improved ViT network-based ADSS optical cable potential measurement method based on physical symbol fusion includes the following steps:

[0012] S1. Collect the three-dimensional electric field distribution parameters of the ADSS optical cable and simultaneously acquire environmental parameters to construct a multi-channel spatiotemporal tensor based on the environment-electric field coupling factor;

[0013] S2. Input the multi-channel spatiotemporal tensor into the improved ViT network to generate the latent space representation vector;

[0014] S3. Based on the improved ViT network, a physical symbol fusion layer is introduced to construct an improved ViT network based on physical symbol fusion. In the physical symbol fusion layer, a differentiable symbol layer is designed to rigidly embed the electromagnetic field conservation law and the charge continuity principle. The loss function integrates time-varying electromagnetic field constraints to eliminate transient operating condition distortion, nonlinear dielectric response model to quantify rain and fog material mutations, and a reinforcement learning-driven dynamic weight allocation mechanism.

[0015] S4. Train the improved ViT network based on physical symbol fusion, constrain the training process through the gradient stabilization technique of neural differential equations, output the three-dimensional dynamic potential distribution field Φ(x,y,z,t), and calculate the electro-erosion risk level Risk in real time based on the spatiotemporal characteristics of the potential gradient.

[0016] An electronic device includes: a memory for storing a computer program; and a processor for executing the computer program to cause the device to perform the ADSS optical cable potential measurement method for an improved ViT network based on physical symbol fusion as described above.

[0017] A computer storage medium storing a computer program, wherein when the computer program is run, a device running the computer program implements the ADSS optical cable potential measurement method for an improved ViT network based on physical symbol fusion as described above.

[0018] Beneficial effects: This invention can integrate multimodal environmental parameters, embed physical prior knowledge, and has dynamic risk inference capabilities, enabling it to complete high-accuracy and high-precision electric field measurements in high-humidity environments; it breaks through the bottleneck of accuracy and reliability of electric field sensing in complex environments. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention;

[0020] Figure 2 This is a schematic diagram of the improved Vision Transformer network structure of the present invention;

[0021] Figure 3This is a schematic diagram of the improved physical symbol layer network structure;

[0022] Figure 4 A schematic diagram of the improved Vision Transformer and physical symbol fusion network structure;

[0023] Figure 5 A schematic diagram illustrating the accuracy of the improved Vision Transformer and physical symbol fusion network;

[0024] Figure 6 A diagram showing the accuracy comparison of the improved Vision Transformer and the physical symbol fusion network under different environments. Detailed Implementation

[0025] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0026] Example

[0027] As attached Figure 1 As shown in this embodiment, an ADSS optical cable potential measurement method based on physical symbol fusion in an improved ViT network includes the following steps:

[0028] S1. Acquire the three-dimensional electric field distribution parameters of the ADSS optical cable and simultaneously obtain environmental parameters to construct a multi-channel spatiotemporal tensor based on the environment-electric field coupling factor; S1 includes the following steps:

[0029] S11. Collect the three-dimensional electric field distribution parameters of the ADSS optical cable through several sensor arrays, and simultaneously acquire environmental parameters, including temperature, humidity, and air pressure. The sensor arrays are deployed at preset intervals along the ADSS optical cable, such as 1.5 meters. Each sensor array includes a set of non-contact MEMS sensors, a digital temperature and humidity sensor, and a high-precision air pressure sensor. The set of non-contact MEMS sensors includes three orthogonal sensors along the X / Y / Z axes, with a measurement range of ±50kV / m and an accuracy of ±0.5kV / m, used to collect data from the ADSS optical cable. The three-dimensional electric field distribution parameters, i.e., the spatial electric field vector (Ex, Ey, Ez); the operating range of the digital temperature and humidity sensor is: temperature range -40℃ to +85℃, humidity range RH range 0% to 100%, used to collect the temperature and humidity parameters of the ADSS optical cable; the high-precision barometric pressure sensor has a range of 300 to 1100 hPa and an accuracy of ±0.1 hPa, used to collect the barometric pressure parameters of the ADSS optical cable; during the parameter acquisition process, a GPS timing module is used to achieve μs-level time synchronization between the spatial electric field vector (sampling rate 2kHz) and the environmental parameters (sampling rate 1Hz);

[0030] S12. Obtaining quantum state codes after constructing quantum states for environmental parameters: After normalizing the temperature, humidity, and air pressure parameters, quantum states are constructed, and quantum state codes are calculated; the normalization formulas include:

[0031]

[0032] Where ∝ represents the normalized humidity, RH, RH min RH max These represent the sampled value, minimum value, and maximum value during the humidity parameter sampling process, respectively; β represents the normalized temperature, T0, T1, and T2. min T max These represent the sampled value, minimum value, and maximum value during the temperature parameter sampling process; γ is the normalized air pressure, and P and P are the sampled values, minimum value, and maximum value, respectively. min P max These represent the sampled value, minimum value, and maximum value during the air pressure parameter sampling process; in this embodiment, environmental parameters are restricted to the [0,1] range through data normalization.

[0033] The computational formulas for constructing quantum states include:

[0034] ψ new =∝|RH>+β|T>+γ|P>

[0035] Where |RH>, |T>, and |P> are preset ground-state vectors for humidity, temperature, and air pressure, respectively, and ψ new Quantum state encoding for environmental parameters;

[0036] S13. Constructing the environment-electric field coupling factor based on quantum state encoding: Mapping environmental parameters to quantum probability models according to quantum state encoding, and obtaining the environment-electric field coupling factor for each environmental parameter by calculating the KL divergence.

[0037] The formula for calculating the KL divergence includes:

[0038]

[0039] Among them, D KL (p env ||p ref ) represents the KL divergence operator for environmental parameters, p env (i) represents the true distribution of environmental parameters, i.e., the sampled values, p ref (i) is the fitted distribution of environmental parameters; the KL divergence is calculated independently for each environmental parameter, namely the KL divergence of humidity, temperature and air pressure, and finally three values ​​are obtained.

[0040] The formulas for calculating the environment-electric field coupling factor include:

[0041]

[0042] in, The environmental-electric field coupling factor. <E peak > represents the expected peak value of the electric field, directly derived from the uncompensated three-dimensional electric field distribution parameter E. raw E0 is the reference electric field strength, which is a static preset value that does not depend on real-time sampling. It is set before system deployment and remains unchanged during operation to ensure the stability of the coupling factor. At this time, the calculated environment-electric field coupling factor also has three values.

[0043] S14. Compensate the three-dimensional electric field distribution parameters based on the environment-electric field coupling factor to obtain the compensated three-dimensional electric field distribution parameters, thus completely eliminating dielectric constant drift. The calculation formula for the compensated three-dimensional electric field distribution parameters includes:

[0044]

[0045] Among them, E raw E represents the three-dimensional electric field distribution parameters obtained from sampling. comp For the compensated three-dimensional electric field distribution parameters, E comp =(E x,comp E y,comp E z,comp );

[0046] S15. Constructing a multi-channel spatiotemporal tensor based on the compensated three-dimensional electric field distribution parameters: The compensated three-dimensional electric field distribution parameters are decomposed into three physically meaningful components: the original electric field intensity, the electric field direction angle, and the electric field vertical tilt angle. These are then normalized to obtain the multi-channel spatiotemporal tensor. Its calculation formula includes:

[0047]

[0048] After normalization, it is represented as:

[0049] θ' = θ / π + 0.5

[0050]

[0051] Where |E| is the original electric field strength, and θ and θ' are the electric field direction angles before and after normalization, respectively; These are the vertical tilt angles of the electric field before and after normalization;

[0052] S1 integrates quantum environmental sensing with data, capturing the three-dimensional electric field distribution and environmental parameters of the optical cable in real time through a non-contact MEMS sensor array. It innovatively adopts the quantum superposition principle to construct an environment-electric field coupling factor, that is, introduces a quantum-inspired environment-electric field coupling mechanism. This mechanism maps temperature, humidity, and air pressure into a quantum probability model, dynamically compensates for environmental interference through variational optimization, and introduces an environment-electric field coupling factor to completely solve the problem of dielectric constant drift in high humidity environments.

[0053] S2. Input the multi-channel spatiotemporal tensor into the improved ViT network to generate the latent space representation vector;

[0054] As attached Figure 2-4 As shown, the improved ViT network is an improved hierarchical Vision Transformer network, and the network architecture includes:

[0055] The input layer takes a multi-channel spacetime tensor as input, which is the electric field data after quantum environment compensation. The multi-channel spacetime tensor has a dimension of 16×16×3, where 16×16 is the spatial grid dimension and 3 represents the channel electric field component. Each independent spatial location in the multi-channel spacetime tensor, i.e., each point in the spatial grid, is defined as a spatial point. The spatial point is the basic unit of the multi-channel spacetime tensor and is processed point by point in the subsequent frequency domain processing layer to capture local frequency domain features.

[0056] The frequency domain processing layer, connected to the input layer, performs a 1024-point FFT transform on each spatial point and applies a bandpass filter to retain the 200-400kHz frequency band, i.e., the power corona characteristic frequency band. It outputs the filtered frequency domain feature tensor, specifically the frequency band data retained for each spatial point after FFT and bandpass filtering. The transfer function calculation formula used in the frequency domain processing layer includes:

[0057]

[0058] Where f is the frequency variable, W f (f) is the transfer function of the bandpass filter;

[0059] The spatial processing layer, connected to the input layer, is used to process the spatial features in the multi-channel spatiotemporal tensor and output a high-dimensional spatial feature tensor.

[0060] The spatial processing layer uses 3×3 convolutional kernels or multi-head self-attention mechanisms (such as 8-head attention) to traverse the spatial grid and compute the feature map for each point. For example, convolutional operations use the ReLU activation function to enhance nonlinear features and output intermediate feature maps; the self-attention mechanism captures global correlations between spatial points through query, key, and value vectors. The output is a high-dimensional spatial feature tensor with dimensions of 16×16×64, encoding rich spatial information for dual-domain fusion in the subsequent frequency-spatial joint attention layer. This output is combined with the frequency domain features of the frequency domain processing layer, and finally, a latent space representation vector is generated through a cross-scale feature fusion layer, supporting the computation of three-dimensional dynamic potential distribution fields.

[0061] The frequency-space joint attention layer, connected to the frequency domain processing layer and the spatial domain processing layer, is used to fuse the frequency domain feature tensors and spatial feature tensors output by the frequency domain processing layer and the spatial domain processing layer, outputting fused features. The frequency-space joint attention layer uses an 8-head self-attention layer query vector dimension. The calculation formula for the fused features includes:

[0062]

[0063] Among them, A fs (F, S) represents the fused features, where F is the frequency domain feature tensor output by the frequency domain processing layer, S is the spatial feature tensor output by the spatial domain processing layer, Q, K, and V are the query vector, key vector, and value vector, respectively, and the subscripts f and s represent the frequency domain and spatial domain, respectively. k For query vector dimensions;

[0064] The cross-scale feature fusion layer, connected to the frequency-space joint attention layer, decomposes the fused features into low-frequency and high-frequency feature maps, performs tensor product operations to obtain entanglement gating coefficients, and then calculates the final feature output, which is the latent space representation vector. Its calculation formula includes:

[0065]

[0066] F out =g×F low +(1-g)×F high

[0067] Where g is the entanglement gating coefficient, F low F high These are low-frequency feature maps and high-frequency feature maps, respectively. out The final feature is the latent space representation vector, and σ is the Sigmoid activation function.

[0068] The fused features contain rich multi-scale electric field information. The cross-scale feature fusion layer decomposes the fused features into low-frequency and high-frequency feature maps using frequency band separation techniques (such as wavelet transform or filtering). The low-frequency feature map captures the global gradual variation components (such as the background electric field distribution trend) in the fused features, while the high-frequency feature map extracts local high-frequency details (such as the distortion signal caused by corona discharge). Both inherit the global information of the fused features but focus on different frequency scales. Entangled gating coefficients are generated through tensor product operations, and the final output is a latent space representation vector, thereby enhancing the model's perception accuracy and dynamic risk warning capability for the multi-scale characteristics of ADSS optical cable potential.

[0069] S2 employs a hierarchical VisionTransformer architecture through joint frequency-space feature extraction and captures long-range spatial electric field correlations via a multi-head self-attention mechanism. It also innovatively introduces a frequency domain bandpass filter to focus on the corona interference frequency band unique to power systems. The improved ViT network significantly enhances the sensing sensitivity of the distortion region around the insulator.

[0070] S3, as attached Figure 2-4 As shown, a physical symbol fusion layer is introduced on the basis of the improved ViT network to construct an improved ViT network based on physical symbol fusion. The physical symbol fusion layer designs a differentiable symbol layer that rigidly embeds the electromagnetic field conservation law and the charge continuity principle. The loss function integrates time-varying electromagnetic field constraints to eliminate transient distortion, a nonlinear dielectric response model to quantify rain and fog material mutations, and a reinforcement learning-driven dynamic weight allocation mechanism. Among these, the nonlinear dielectric response model quantifies rain and fog material mutations by calculating the energy functional constraints, corresponding to the dynamic dielectric constant ω. dyn Dynamic dielectric constant dyn It is a key parameter that changes nonlinearly with environmental conditions (such as rain and fog).

[0071] The working process of the physical symbol fusion layer includes:

[0072] S31. Input the latent space representation vector into the physics-guided differentiable symbol layer, which includes electromagnetic field conservation constraints, charge continuity constraints, and energy functional constraints; couple the electromagnetic field conservation residual term, charge continuity residual term, and energy functional constraint term into the loss function.

[0073] The formulas for calculating electromagnetic field conservation constraints include:

[0074]

[0075] Among them, R em For the electromagnetic field conservation residual term, × represents the curl operator, E represents the electric field intensity vector (three-dimensional vector), i.e., the three-dimensional electric field distribution parameters acquired by the sensor array in S1; B represents the magnetic induction intensity vector (three-dimensional vector). The curl of the electric field is calculated by automatic differentiation, and the rate of change of magnetic flux is obtained by the difference between adjacent frames. That is, during the acquisition of the three-dimensional electric field distribution parameters of the ADSS optical cable, the three-dimensional electric field distribution parameters are acquired in real time at a sampling rate of 2kHz, forming time series data. The time series data consists of sampling frames at consecutive time points, and each frame contains the spatial electric field vector at one time point.

[0076] The formulas for calculating charge continuity constraints include:

[0077]

[0078] Among them, R charge Let J be the charge continuity residual term, J be the current density vector, and ρ be the charge density, derived from the electric field intensity divergence. Where ò0 is the vacuum permittivity. • For divergence operators, The electric field intensity divergence;

[0079] The calculation process for the current density vector is as follows:

[0080] The conductivity of the outer sheath material of ADSS optical cables is not constant, but strongly depends on environmental conditions, especially humidity and temperature. The role of environmental parameters has been emphasized in S1, and an environment-electric field coupling factor has been constructed. Similarly, a dynamic conductivity model can be established:

[0081] σ dyn =f(T,RH,P)

[0082] Where, σ dyn Let T be the dynamic conductivity, RH be the temperature, P be the relative humidity, and f() be the empirical formula fitted from experimental data.

[0083] The generalized Ohm's law provides a formula for calculating the current density vector:

[0084] J = σ dyn E comp

[0085] Where J is the current density vector, σ dyn E is the dynamic conductivity. comp The parameters are the three-dimensional electric field distribution after compensation.

[0086] The formulas for calculating energy functional constraints include:

[0087]

[0088] Among them, L symbol For the energy functional constraint term, ∫ Ω (...)dV is the volume integral operator, ò dyn The dynamic dielectric constant, Let μ be the gradient of the scalar potential function, and μ0 be the free permeability. The time-varying rate of magnetic field strength is obtained by discretization, i.e., by numerical integration of the spatial grid using the trapezoidal method.

[0089] S32. Adaptive reinforcement learning weight allocation strategy is adopted; including the following:

[0090] ① Define the state space Where s t MSE is a state-space vector. t Mean square error, Let t be the L2 norm of the gradient of the loss function, and t be the time step. For normalized time steps;

[0091] ② Output weight adjustment actions through a deep Q-network; the deep Q-network is a 3-layer fully connected network with 128, 64, and 3 neurons respectively; the deep Q-network (DQN) outputs the optimal weight adjustment actions based on the current state;

[0092] ③ Dynamic weight update: The weights of the three constraint terms—magnetic field conservation residual, charge continuity residual, and energy functional constraint—are dynamically updated. The importance of each constraint term in the total loss is not a preset constant, but is adjusted in real-time and adaptively based on the model's actual performance (such as convergence and gradient stability) as training progresses. The dynamic weight calculation formula includes:

[0093]

[0094] Where, λ k (t) represents the dynamic weight value, λ k (0) represents the initial weight values, η represents the decay rate, clip() represents the clipping function, Q() represents the deep Q-network, and s t Let a be the state vector. t For action vectors;

[0095] ④ The reward function guides the weight allocation strategy towards the optimal direction. The formula for the reward function includes:

[0096]

[0097] Where, r t The reward value is represented by α and β, which are weighting coefficients, and MSE is the mean squared error. The L2 norm of the gradient of the loss function is used to encourage the model to achieve a balance between reducing prediction error (MSE) and maintaining training stability (so that the gradient norm is not too large), thereby guiding the weight allocation strategy to evolve in the optimal direction.

[0098] S4. Train the improved ViT network based on physical symbol fusion, constrain the training process through the gradient stabilization technique of neural differential equations, output the three-dimensional dynamic potential distribution field Φ(x,y,z,t), and calculate the electro-erosion risk level Risk in real time based on the spatiotemporal characteristics of the potential gradient.

[0099] S4 outputs a three-dimensional dynamic potential distribution field. The improved ViT network based on physical symbol fusion is a risk model based on the fusion of spatiotemporal gradient and historical electric field fluctuation characteristics, which is the electro-erosion risk level calculation model. It generates electro-erosion risk level and evolution path map in real time to achieve early warning of corrosion.

[0100] Gradient stabilization techniques for neural differential equations include:

[0101] Gradient dynamic modeling:

[0102]

[0103] in, f is the gradient of the loss function, θ is the trainable parameters of the improved ViT network based on physical symbol fusion, such as the weight matrix and bias vector in the improved ViT network and the physical symbol fusion layer, and f neural It is a neutral function, t is the time variable, and E is the electric field intensity tensor;

[0104] f neural It is the core function of gradient stabilization, and its role is to dynamically adjust the training process through time, data, and gradient feedback. neural Receive time variable t, current gradient of loss function The electric field intensity tensor E is used as input, and the output is a gradient adjustment. Its purpose is to simulate the temporal evolution of the gradient, making its changes smooth and avoiding drastic fluctuations. For example, when the gradient norm is too large, the function may decay the gradient; when the gradient is too small, it amplifies the effective signal.

[0105] Triple control mechanism:

[0106] ① Anomaly detection threshold γ th =10 3

[0107] ② Implicit Euler update:

[0108]

[0109] ③ Attention gating retains effective signals; effective signals refer to gradient components or physical features that contribute positively to training optimization and physical consistency, including stable gradients, signals that conform to physical laws, and information related to key physical quantities.

[0110] The formula for calculating the risk level of electrical erosion is:

[0111]

[0112] Risk represents the risk value for electrical erosion. For spatial gradient terms, For the gradient operator of potential distribution, Entropy(E) is the L2 norm of the rate of change of the potential distribution with time. hist E represents the historical electric field entropy term. hist For historical electric field data sequences, E i For the i-th historical electric field data point, p(E) i ) is E i The probability of occurrence within a sliding window (e.g., 60 minutes), where N is the number of historical data points or partitions;

[0113] Spatial gradient integral is used to perform triangular meshing on the surface of the optical cable, with a mesh side length ≤ 2cm; historical electric field entropy is used for historical electric field analysis, with a sliding window taking 60 minutes of data, i.e., the sliding window for parameter acquisition in S1; early warning map generation:

[0114] Risk level Numerical range Warning colors Response measures Level I Risk<0.3 blue Normal monitoring Level II 0.3≤Risk<0.6 yellow Strengthen inspection Level III Risk ≥ 0.6 red Emergency treatment

[0115] The network training employs a three-stage training strategy:

[0116] Phase 1: Train only the VisionTransformer, i.e., an improved ViT network; the optimizer is AdamW (parameters β1 = 0.9, β2 = 0.999); the learning rate is 10. -3 Batch size is 32;

[0117] β1 and β2 are parameters belonging to the AdamW optimizer formula. AdamW is a variant of the Adam optimizer, and its standard update formula is usually expressed as:

[0118] m t =β1m t-1 +(1-β1)g t

[0119]

[0120] Where, mt For the first-order moment estimate, v t For second-order moment estimation, g t Let β1 be the first-order moment decay rate, β2 be the second-order moment decay rate, and m be the gradient vector. t-1 For the first moment estimate of the previous time step, v t-1 This is an estimate of the second moment of the previous time step;

[0121] β1 = 0.9 represents the exponential decay rate of the first-moment estimate (such as the mean of the gradient). A larger value indicates a higher weighting of historical gradient information, which helps stabilize training. β2 = 0.999 represents the exponential decay rate of the second-moment estimate (such as the mean of the squared gradient). A larger value results in a smoother estimate of the gradient variance, better adapting to learning rate adjustments.

[0122] When improving the ViT network, these parameters ensure stable gradient updates, avoid oscillations, and improve convergence efficiency. Choosing standard values ​​(0.9 and 0.999) is a common setting based on the Adam optimizer, balancing convergence speed and stability.

[0123] Phase 2: Jointly train the physical symbol layer, i.e., the improved ViT network based on physical symbol fusion; the optimizer is L-BFGS, the bottom weights of the improved ViT network are frozen, and the learning rate is 10. -5 ;

[0124] Phase 3: Fine-tuning the physics constraints, using SGD (momentum 0.9) as the optimizer and a learning rate of 10. -6 ;

[0125] The training set consists of 80% normal operating conditions + 15% abnormal operating conditions + 5% transient shocks;

[0126] Normal operating conditions (80%): Data content consists of ADSS optical cable data under stable operating conditions. This includes: three-dimensional electric field distribution parameters; environmental parameters; and multi-channel spatiotemporal tensors. Characteristics: The data is stable, without distortion or abrupt changes, and is used to train models to identify reference electric field patterns.

[0127] Abnormal Operating Conditions (15%): Data consists of optical cables under abnormal conditions, such as electric field distortion caused by corona discharge, local insulation degradation, or mechanical damage. This includes: distorted electric field parameters; environmental parameters. Features: The data contains local distortions to improve the model's ability to detect electric field distortions.

[0128] Transient Impacts (5%): Data consists of transient events such as lightning strikes, switching operations, or sudden changes in the electric field caused by external interference. This includes transient electric field parameters and environmental parameters. Characteristics: The data has a short timescale and changes rapidly, used to train models to handle dynamic events.

[0129] The primary label is the true three-dimensional dynamic potential distribution field Φ_true(x,y,z,t), and the secondary label is the electrolytic erosion risk level Risk_true. The three-stage training strategy (stages 1-3) relies on the accuracy of the labels. In stage 1, when training and improving the ViT network, Φ_true(x,y,z,t) is used as the primary label. In stages 2-3, during joint training, Risk-true is added to optimize the physical symbol layer.

[0130] The simulation verification process in this embodiment is as follows:

[0131] Figure 5 A schematic diagram illustrating the accuracy of the improved Vision Transformer and physical symbol fusion network; Figure 6 This diagram illustrates the accuracy comparison of the improved Vision Transformer and physical symbol fusion network under different environments. As can be seen from the diagram, the network structure of this invention outperforms other network structures in both accuracy and precision.

[0132] The present invention also discloses an electronic device, the device comprising: a memory for storing a computer program; and a processor for executing the computer program to enable the device to perform the aforementioned ADSS optical cable potential measurement method for an improved ViT network based on physical symbol fusion.

[0133] The present invention also provides a computer storage medium on which a computer program is stored. When the computer program is run, the device running the computer program implements the aforementioned ADSS optical cable potential measurement method for an improved ViT network based on physical symbol fusion.

[0134] In the embodiments of this application, the terms "first" and "second" (if they exist) are used only as name identifiers and do not represent the order of first and second.

[0135] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that all or part of the steps in the methods of the above embodiments can be implemented by means of software plus a general-purpose hardware platform. Based on this understanding, the technical solution of this application can be embodied in the form of a software product. This computer software product can be stored in a storage medium. The memory can be various types of memory, such as random access memory, read-only memory, flash memory, etc., such as read-only memory (ROM) / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which can be a personal computer, server, or network communication device such as a router) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0136] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for measuring the potential of ADSS optical cables in an improved ViT network based on physical symbol fusion, characterized in that, Includes the following steps: S1. Collect the three-dimensional electric field distribution parameters of the ADSS optical cable and simultaneously acquire environmental parameters to construct a multi-channel spatiotemporal tensor based on the environment-electric field coupling factor; S2. Input the multi-channel spatiotemporal tensor into the improved ViT network to generate the latent space representation vector; S3. Based on the improved ViT network, a physical symbol fusion layer is introduced to construct an improved ViT network based on physical symbol fusion. In the physical symbol fusion layer, a differentiable symbol layer is designed to rigidly embed the electromagnetic field conservation law and the charge continuity principle. The loss function integrates time-varying electromagnetic field constraints to eliminate transient operating condition distortion, nonlinear dielectric response model to quantify rain and fog material mutations, and a reinforcement learning-driven dynamic weight allocation mechanism. S4. Train the improved ViT network based on physical symbol fusion, constrain the training process through the gradient stabilization technique of neural differential equations, output the three-dimensional dynamic potential distribution field Φ(x,y,z,t), and calculate the electro-erosion risk level Risk in real time based on the spatiotemporal characteristics of the potential gradient.

2. The ADSS optical cable potential measurement method based on physical symbol fusion in an improved ViT network according to claim 1, characterized in that: S1, including the following steps: S11. Collect the three-dimensional electric field distribution parameters of the ADSS optical cable through several sensor arrays, and simultaneously acquire environmental parameters, including temperature, humidity and air pressure parameters. S12. Obtaining quantum state codes after constructing quantum states for environmental parameters: After normalizing the temperature, humidity, and air pressure parameters, construct quantum states and calculate the quantum state codes; S13. Constructing the environment-electric field coupling factor based on quantum state encoding: Mapping environmental parameters to a quantum probability model based on quantum state encoding, and obtaining the environment-electric field coupling factor by calculating KL divergence; S14. Compensate the three-dimensional electric field distribution parameters according to the environment-electric field coupling factor to obtain the compensated three-dimensional electric field distribution parameters. S15. Construct a multi-channel spatiotemporal tensor based on the compensated three-dimensional electric field distribution parameters: Decompose the compensated three-dimensional electric field distribution parameters into the original electric field intensity, electric field direction angle, and electric field vertical tilt angle, and perform normalization processing to obtain a multi-channel spatiotemporal tensor.

3. The ADSS optical cable potential measurement method based on physical symbol fusion in an improved ViT network according to claim 2, characterized in that: ADSS optical cables deploy a sensor array at preset intervals; each sensor array includes a set of non-contact MEMS sensors, a digital temperature and humidity sensor, and a high-precision barometric pressure sensor. The set of non-contact MEMS sensors includes three orthogonal sensors along the X / Y / Z axes.

4. The ADSS optical cable potential measurement method based on physical symbol fusion in an improved ViT network according to claim 2, characterized in that: The computational formulas for constructing quantum states include: ψ new =∝|RH>+β|T>+γ|P> Where |RH>, |T>, and |P> are preset ground-state vectors for humidity, temperature, and air pressure, respectively, and ψ new Quantum state encoding for environmental parameters; The formulas for calculating the environment-electric field coupling factor include: in, The environmental-electric field coupling factor. <E peak > represents the expected peak value of the electric field; E0 represents the reference electric field strength.

5. The ADSS optical cable potential measurement method based on physical symbol fusion in an improved ViT network according to claim 2, characterized in that: The formulas for calculating multichannel spatiotemporal tensors include: After normalization, it is represented as: θ' = θ / π + 0.5 Where |E| is the original electric field strength, and θ and θ' are the electric field direction angles before and after normalization, respectively; These are the vertical tilt angles of the electric field before and after normalization.

6. The ADSS optical cable potential measurement method based on physical symbol fusion in an improved ViT network according to claim 1, characterized in that: The improved ViT network is an improved hierarchical Vision Transformer network, and the network architecture includes: Input layer, the input is a multi-channel spatiotemporal tensor; The frequency domain processing layer, connected to the input layer, is used to perform a 1024-point FFT transformation on each spatial point and apply a bandpass filter to retain the 200-400kHz frequency band, i.e. the power corona characteristic frequency band, and output the filtered frequency domain feature tensor. The spatial processing layer, connected to the input layer, is used to process the spatial features in the multi-channel spatiotemporal tensor and output the spatial feature tensor. The frequency-space joint attention layer, connected to the frequency domain processing layer and the spatial domain processing layer, is used to fuse the frequency domain feature tensors and spatial feature tensors output by the frequency domain processing layer and the spatial domain processing layer, and output the fused features; the frequency-space joint attention layer adopts an 8-head self-attention layer query vector dimension; The cross-scale feature fusion layer, connected to the frequency-space joint attention layer, is used to perform tensor product operations on the low-frequency feature map and the high-frequency feature map, obtain the entanglement gating coefficients, and then calculate the final feature output, which is the latent space representation vector.

7. The ADSS optical cable potential measurement method based on physical symbol fusion in an improved ViT network according to claim 1, characterized in that: The working process of the physical symbol fusion layer includes: S31. Input the latent space representation vector into the physics-guided differentiable symbol layer, which includes electromagnetic field conservation constraints, charge continuity constraints, and energy functional constraints; couple the electromagnetic field conservation residual term and the charge continuity residual term in the loss function. The formulas for calculating electromagnetic field conservation constraints include: Among them, R em For electromagnetic field conservation residuals, ▽× is the curl operator, E is the electric field intensity vector, and B is the magnetic induction intensity vector; The formulas for calculating charge continuity constraints include: Among them, R charge Let J be the charge continuity residual term, J be the current density vector, and ρ be the charge density, which is derived from the electric field intensity divergence, ρ = ò0▽·E; where ò0 is the vacuum permittivity, ▽· is the divergence operator, and ▽·E is the electric field intensity divergence. The formulas for calculating energy functional constraints include: Among them, L symbol For the energy functional constraint term, ∫ Ω (...)dV is the volume integral operator, ò dyn Let φ be the dynamic permittivity, ▽φ be the gradient of the scalar potential function, and μ0 be the free permeability. The time rate of change of magnetic field strength; S32. Adaptive reinforcement learning weight allocation strategy is adopted; including the following: ① Define the state space Where s t MSE is a state-space vector. t Let be the mean squared error, ‖▽L‖2 be the L2 norm of the gradient of the loss function, and t be the time step. For normalized time steps; ② The deep Q-network outputs the weight adjustment action; the deep Q-network is a 3-layer fully connected network with 128, 64 and 3 neurons respectively; the deep Q-network outputs the optimal weight adjustment action based on the current state; ③ Dynamic weight update: The weights of the three constraint terms—magnetic field conservation residual, charge continuity residual, and energy functional constraint—are dynamically updated; the formulas for calculating the dynamic weight values ​​include: Where, λ k (t) represents the dynamic weight value, λ k (0) represents the initial weight values, η represents the decay rate, clip() represents the clipping function, Q() represents the deep Q-network, and s t Let a be the state vector. t For action vectors; ④ The reward function guides the weight allocation strategy towards the optimal direction. The formula for the reward function includes: r t =-(α·MSE+β·‖▽L‖2) Where, r t Let α be the reward value, β be the weighting coefficients, MSE be the mean squared error, and ‖▽L‖2 be the L2 norm of the gradient of the loss function.

8. The ADSS optical cable potential measurement method based on physical symbol fusion in an improved ViT network according to claim 1, characterized in that: The training process is constrained by gradient stabilization techniques based on neural differential equations. The network training adopts a three-stage training strategy: Phase 1: Train only VisionTransformer, i.e., improve the ViT network; The optimizer is AdamW; Phase 2: Jointly train the physical symbol layer, i.e., the improved ViT network based on physical symbol fusion; the optimizer is L-BFGS, and the bottom weights of the improved ViT network are frozen; Phase 3: Fine-tune the physical constraints, using SGD as the optimizer; The training set consists of 80% normal operating conditions + 15% abnormal operating conditions + 5% transient shocks.

9. An electronic device, characterized in that, The device includes: a memory for storing a computer program; and a processor for executing the computer program to cause the device to perform an ADSS optical cable potential measurement method for an improved ViT network based on physical symbol fusion as described in any one of claims 1-8.

10. A computer storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is run, the device running the computer program implements the ADSS optical cable potential measurement method for an improved ViT network based on physical symbol fusion as described in any one of claims 1-8.

Images