Ground wire damage identification method, system and equipment based on probability learning and medium

Abstract

The invention discloses a ground wire damage identification method, system and device based on probability learning and a medium, and belongs to the technical field of ground wire damage identification, and the method comprises the steps: obtaining ultrasonic guided wave reflection signal data of a ground wire, and obtaining map data according to the ultrasonic guided wave reflection signal data; constructing a probability deep learning model according to the map data; obtaining a damage type identification result based on the probability deep learning model and the atlas data, and determining an uncertainty quantitative index of the damage type identification result; and determining a ground wire damage diagnosis result according to the damage type identification result in combination with the uncertainty quantitative index. Through noise reduction processing and time-frequency domain joint representation conversion of ultrasonic guided wave reflection signal data, the problems of strong sound wave scattering effect and frequency dispersion caused by a multi-strand stranded structure of the steel strand are effectively solved, the signal-to-noise ratio of a time domain reflection wave packet is increased, and early-stage accurate capture of hidden corrosion such as inner-layer strand breakage and intergranular cracks is achieved.

Description

Technical Field

[0001] This invention relates to the field of grounding wire damage identification technology, and in particular to grounding wire damage identification methods, systems, devices and media based on probabilistic learning. Background Technology

[0002] With the rapid construction and intelligent upgrading of ultra-high voltage transmission networks, the damage to overhead ground wires (OPGW / steel strands), as key components for resisting lightning strikes and ensuring line insulation, caused by long-term exposure to meteorological environments is becoming increasingly prominent, directly threatening the reliability and economic efficiency of power grid operation and maintenance. Traditional detection methods mainly rely on periodic manual inspections, visual observation, and local sampling analysis, which have two fundamental drawbacks: First, the detection coverage is low and the timeliness is poor, making it difficult to capture the early evolution characteristics of hidden corrosion (such as internal strand breakage and intergranular cracks), resulting in defect detection lagging behind the actual damage process; Second, the detection results are highly subjective, lacking quantitative indicators to support them, and cannot accurately assess the degree of damage and remaining lifespan. Although ultrasonic guided wave technology has been introduced due to its advantages such as long-distance propagation and multi-modal sensitivity, In this field, practical applications are still limited by the strong acoustic wave scattering effect of the multi-strand stranded structure of steel strands—the guided wave signal repeatedly interacts with the strand interface during propagation, resulting in dispersion and mode conversion phenomena, which reduces the signal-to-noise ratio of the time-domain reflected wave packet; existing methods mostly rely on manual extraction of empirical features such as time delay and amplitude attenuation, which not only makes it difficult to effectively distinguish corrosion types, but is also more susceptible to environmental noise (wind vibration, electromagnetic interference), causing the false detection rate to rise; in addition, although deep learning-based damage identification models have shown some potential in laboratory environments, their generalization ability is severely limited by the single distribution characteristics of limited samples. When facing complex field conditions, the stability of the model output drops sharply, and there is a lack of quantitative characterization of the uncertainty of the prediction results, making it difficult to meet the power system's need for high-confidence decision-making.

[0003] Current technological systems also have shortcomings at the systemic level: most studies focus on single detection tasks, failing to achieve collaborative diagnosis of corrosion morphology, spatial location, and severity, resulting in fragmented operational decision-making information; research on adaptability to edge computing scenarios is particularly insufficient, existing models have a large number of parameters and high computational complexity, making them difficult to deploy on resource-constrained online monitoring terminals, and lack self-evolution capabilities in real-world environments, unable to continuously optimize model performance through incremental learning; more importantly, existing methods generally adopt threshold alarm mechanisms, which can only provide binary judgments, failing to reveal the gradual evolution of damage, and making it difficult to construct a correlation model between corrosion rate and the external environment, resulting in a lack of data support for preventive maintenance strategies; although some scholars have attempted to introduce multi-sensor fusion to improve reliability, its hardware costs are high, data heterogeneity is strong, and the optimization problems of cross-modal feature alignment and decision-level fusion have not yet been solved at the algorithm level, making large-scale promotion difficult. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is: how to address the difficulty of accurately extracting damage features from time-domain reflection waveforms due to the strong acoustic wave scattering characteristics of the multi-strand stranded structure of steel strands in traditional ultrasonic guided wave technology, the insufficient robustness of existing detection methods to environmental noise (such as wind vibration and electromagnetic interference), and the poor prediction stability of deep learning-based detection models under real working conditions (temperature fluctuations, icing, and salt spray corrosion). By using Bayesian theory to introduce probability distribution into deep learning methods, a probabilistic deep learning model is constructed to achieve real-time perception and risk warning of the corrosion status of overhead ground wires throughout their entire life cycle.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a ground wire damage identification method based on probabilistic learning, comprising the following steps:

[0007] Acquire ultrasonic guided wave reflection signal data of the ground wire, and obtain spectral data based on the ultrasonic guided wave reflection signal data; construct a probabilistic deep learning model based on the spectral data; obtain damage type identification results based on the probabilistic deep learning model and the spectral data, and determine the uncertainty quantification index of the damage type identification results; determine the ground wire damage diagnosis results based on the damage type identification results and the uncertainty quantification index.

[0008] As a preferred embodiment of the ground wire damage identification method based on probabilistic learning described in this invention, the step of acquiring ultrasonic guided wave reflection signal data of the ground wire and obtaining spectral data based on the ultrasonic guided wave reflection signal data includes: acquiring ultrasonic guided wave reflection signal data of the ground wire; performing noise reduction processing on the ultrasonic guided wave reflection signal data; and converting the noise-reduced ultrasonic guided wave reflection signal data into spectral data.

[0009] As a preferred embodiment of the ground wire damage identification method based on probabilistic learning described in this invention, the method involves: constructing a probabilistic deep learning model based on the spectral data, including: setting the weights in the probabilistic deep learning model as parameters that follow a probability distribution; approximating the posterior distribution of the weights; and constructing a loss function based on the variational evidence lower bound optimization principle. The beneficial effects of this preferred embodiment are that by setting the weights in the probabilistic deep learning model as parameters that follow a probability distribution, using variational inference to approximate the posterior distribution of the weights, and constructing a loss function based on the variational evidence lower bound optimization principle, the high-dimensional integral burden of marginal probability calculation in traditional Bayesian inference is effectively avoided, improving the computational efficiency of model training. Simultaneously, the KL divergence regularization term controls model complexity, effectively suppressing overfitting and ensuring the model's generalization ability and prediction stability in the field environment.

[0010] As a preferred embodiment of the ground wire damage identification method based on probabilistic learning described in this invention, the following steps are included: obtaining damage type identification results based on the probabilistic deep learning model and the spectral data, and determining the uncertainty quantification index of the damage type identification results, comprising: extracting features from the spectral data using the probabilistic deep learning model; identifying the damage type of the ground wire based on the extracted features to obtain the damage type identification result; and calculating the confidence level of the damage type identification result according to the probabilistic deep learning model to obtain the uncertainty quantification index of the damage type identification result.

[0011] As a preferred embodiment of the ground wire damage identification method based on probabilistic learning described in this invention, the method for determining ground wire damage diagnosis results based on the damage type identification results and in combination with the uncertainty quantification index includes: determining the type and location information of the damage based on the damage type identification results; and generating ground wire damage diagnosis results by combining the damage type and location information with the uncertainty quantification index.

[0012] As a preferred embodiment of the ground wire damage identification method based on probabilistic learning described in this invention, the training process of the probabilistic deep learning model includes: converting nodes into deterministic nodes using a reparameterization technique; eliminating noise by sampling a pseudo-random symbol matrix using a Flipout estimator; and updating the parameters of the probabilistic deep learning model using a Bayesian backpropagation method. The beneficial effects of this preferred embodiment are that by converting random nodes into deterministic nodes using a reparameterization technique, the gradient non-transferability problem caused by sampling in the probabilistic model is solved; the high variance noise in mini-batch training is effectively eliminated by sampling a pseudo-random symbol matrix using a Flipout estimator; and the parameter update is completed using a Bayesian backpropagation method. This not only maintains the uncertainty quantification capability of the probabilistic model but also improves the model training convergence speed and numerical stability.

[0013] As a preferred embodiment of the probabilistic learning-based ground wire damage identification method described in this invention, the method for generating a ground wire damage diagnosis result by combining the damage type and location information with the uncertainty quantification index includes: comparing the uncertainty quantification index with a preset threshold to determine whether a structured diagnostic conclusion is generated; and comparing the confidence level of the structured diagnostic conclusion with a preset alarm threshold to determine whether an automatic operation and maintenance response mechanism is triggered. The beneficial effect of this preferred embodiment is that by intelligently comparing the uncertainty quantification index with a preset threshold to determine whether a structured diagnostic conclusion is generated, and by performing a secondary comparison between the confidence level of the diagnostic conclusion and the alarm threshold to determine whether an automatic operation and maintenance response mechanism is triggered, the method effectively avoids false alarms caused by excessive model uncertainty, realizes a hierarchical early warning mechanism based on confidence assessment, provides scientific and reliable automated decision support for power system operation and maintenance, and improves the practicality and operability of ground wire damage detection.

[0014] Another objective of this invention is to provide a ground wire damage identification system based on probabilistic learning.

[0015] To address the aforementioned technical problems, this invention provides the following technical solution: a ground wire damage identification system based on probabilistic learning, comprising: a ground wire damage data module for acquiring ultrasonic guided wave reflection signal data of the ground wire and obtaining spectral data based on the ultrasonic guided wave reflection signal data; a probabilistic deep learning module for constructing a probabilistic deep learning model based on the spectral data; a damage identification module for obtaining damage type identification results based on the probabilistic deep learning model and the spectral data, and determining the uncertainty quantification index of the damage type identification results; and a damage diagnosis module for determining the ground wire damage diagnosis results based on the damage type identification results and the uncertainty quantification index.

[0016] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the ground wire damage identification method based on probabilistic learning.

[0017] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements the steps of the ground wire damage identification method based on probabilistic learning.

[0018] The beneficial effects of this invention are as follows: This invention effectively solves the strong acoustic wave scattering effect and dispersion problem caused by the multi-strand strand structure of steel strands through noise reduction processing and joint time-frequency domain characterization of ultrasonic guided wave reflection signal data. This improves the signal-to-noise ratio of the time-domain reflected wave packet, overcoming the shortcomings of low coverage and poor timeliness of traditional manual inspection. It achieves early and accurate detection of hidden corrosion such as inner strand breaks and intergranular cracks. By constructing a probabilistic deep learning model, the network weights are transformed into probability distribution parameters, and the calculation process is simplified based on the variational evidence lower bound optimization principle. This not only avoids the high-dimensional integral burden of marginal probability calculation in traditional Bayesian inference, improving computational efficiency, but also... Adaptive constraints on model parameters were achieved through the KL divergence regularization term, effectively suppressing overfitting and ensuring model stability under field conditions. By integrating deep feature extraction and uncertainty quantification mechanisms, two key damage fingerprints—abnormal guided wave reflection peaks and energy attenuation profiles in specific frequency bands—were accurately identified, enabling precise differentiation of multiple damage types such as pitting, cracks, and broken strands. At the same time, the asymmetric properties of KL divergence were utilized to ensure that the approximate distribution avoids covering the zero-probability region of the true posterior, improving the reliability of uncertainty estimation and providing high-confidence quantitative decision support for power system operation and maintenance, effectively reducing the false detection rate caused by environmental noise interference. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0020] Figure 1 This is an overall flowchart of a ground wire damage identification method based on probabilistic learning, provided as an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of a computer device provided in one embodiment of the present invention. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a ground wire damage identification method based on probabilistic learning, including:

[0026] S100: Acquire ultrasonic guided wave reflection signal data of the ground wire, and obtain spectrum data based on the ultrasonic guided wave reflection signal data.

[0027] S200: Construct a probabilistic deep learning model based on the map data.

[0028] S300: Based on probabilistic deep learning models and atlas data, damage type identification results are obtained, and the uncertainty quantification index of the damage type identification results is determined.

[0029] S400: Determine the grounding wire damage diagnosis result based on the damage type identification result and the uncertainty quantification index.

[0030] It should be noted that overhead ground wires, as key components for resisting lightning strikes and ensuring line insulation, are increasingly susceptible to damage caused by long-term exposure to complex meteorological environments, directly threatening the reliability and economic efficiency of power grid operation and maintenance. Traditional detection methods, mainly relying on periodic manual inspections and visual observation and local sampling analysis, suffer from fundamental defects such as low detection coverage and poor timeliness. They are unable to capture the early evolution characteristics of hidden corrosion, such as internal strand breakage and intergranular cracks, resulting in defect detection lagging behind the actual damage process. At the same time, the detection results are highly subjective, lacking quantitative indicators, and cannot accurately assess the degree of damage and remaining life. Although ultrasonic guided wave technology has been introduced into this field due to its advantages such as long-distance propagation and multimodal sensitivity, its practical application is still limited. Due to the strong acoustic scattering effect of the multi-strand stranded structure of steel strands, the guided wave signal repeatedly interacts with the strand interface during propagation, resulting in dispersion and mode conversion phenomena. This reduces the signal-to-noise ratio of the time-domain reflected wave packet. Existing methods mostly rely on manual extraction of empirical features such as time delay and amplitude attenuation, which not only makes it difficult to effectively distinguish corrosion types but is also susceptible to environmental noise such as wind vibration and electromagnetic interference, leading to a surge in false detection rates. In addition, although deep learning-based damage identification models have shown some potential in laboratory environments, their generalization ability is severely limited by the single distribution characteristics of a limited number of samples. When facing field conditions, the stability of the model output drops sharply, and there is a lack of quantitative representation of the uncertainty of the prediction results, making it difficult to meet the high-confidence decision-making requirements of power systems.

[0031] Therefore, to address the aforementioned issues of insufficient detection accuracy, poor environmental adaptability, and lack of uncertainty quantification, an intelligent damage identification method integrating ultrasonic guided wave technology and probabilistic deep learning is constructed through steps S100-S400. This method converts ultrasonic guided wave reflection signals into spectral data jointly characterized in the time and frequency domains, establishes a probabilistic deep learning model, models the probability distribution of weights, and obtains damage type identification results with uncertainty quantification capabilities. This enables accurate identification and confidence assessment of multiple damage types such as pitting, cracks, and fractures; real-time monitoring of grounding wire damage; quantification of the reliability of identification results; and early warning filtering for low-confidence identification results. Simultaneously, based on the uncertainty quantification mechanism within the probabilistic framework, highly reliable decision support for damage identification in real-world environments is achieved.

[0032] Example 2, refer to Figure 1 This is the second embodiment of the present invention. Based on the above embodiments, a ground wire damage identification method based on probabilistic learning is provided.

[0033] In this embodiment of the invention, step S100 involves acquiring ultrasonic guided wave reflection signal data of the ground wire and obtaining spectral data based on the ultrasonic guided wave reflection signal data, including the following steps A1-A3:

[0034] A1: Collect ultrasonic guided wave reflection signal data of the ground wire.

[0035] A2: Noise reduction processing is performed on the ultrasonic guided wave reflection signal data.

[0036] A3: Convert the noise-reduced ultrasonic guided wave reflection signal data into spectral data.

[0037] Specifically, in step A1, the acquisition of ultrasonic guided wave reflection signal data of the ground wire refers to the real-time acquisition of ultrasonic guided wave reflection signals through a distributed sensor network deployed on the overhead ground wire. This effectively addresses the strong acoustic wave scattering effect of the multi-strand stranded structure of the steel strand and avoids the complex dispersion and mode conversion phenomena caused by the repeated interaction between the guided wave signal and the strand interface during propagation.

[0038] For example, in step A1, the ultrasonic guided wave reflection signal data of the ground wire can be acquired through the following specific operations:

[0039] An ultrasonic thin-film sensor was installed near the grounding damage and connected to an oscilloscope via a bayonet for data acquisition in trigger mode. Only waveforms with a certain amplitude level were recorded to prevent interference from environmental noise. All signals were recorded and stored on a PC, which was wirelessly connected to an AE oscilloscope to simulate a field monitoring environment that requires remote control to pass the test. Each AE event was sampled at a length of 10,000 and a sampling frequency of 1 MHz, with 20% pre-recording deployed to ensure that the start time slot of each AE event was generally consistent, and a high-pass filter was used to eliminate low-frequency noise.

[0040] Specifically, the noise reduction process in step A2 refers to eliminating low-frequency noise by using a high-pass filter and suppressing environmental interference, including the effects of noise sources such as wind vibration and electromagnetic interference, by using an adaptive noise reduction module, thereby improving the signal-to-noise ratio of the time-domain reflected wave packet.

[0041] Specifically, in step A3, converting the noise-reduced ultrasonic guided wave reflection signal data into spectral data means converting the noise-reduced signal into spectral data that is jointly characterized in the time and frequency domains, effectively decoupling the frequency domain overlap between damage features and noise.

[0042] In one optional implementation, step S100 acquires ultrasonic guided wave reflection signal data of the ground wire and obtains spectral data based on the ultrasonic guided wave reflection signal data. Alternatively, a trigger mode data acquisition strategy can be used to set an amplitude threshold to filter valid signals and avoid recording invalid waveforms caused by environmental noise. At the same time, a wireless transmission method is used to realize remote data acquisition, adapting to the actual needs of the on-site monitoring environment. A pre-recording mechanism ensures the consistency of the start time slot of each acoustic emission event, providing a standardized data format for subsequent signal processing and feature extraction. Furthermore, sampling parameters are adjusted according to the structural characteristics of the steel strand to ensure that key damage feature information in the guided wave signal can be captured.

[0043] In another optional implementation, in step S100, ultrasonic guided wave reflection signal data of the ground wire is acquired, and spectral data is obtained based on the ultrasonic guided wave reflection signal data. Alternatively, ultrasonic guided wave signals from different locations can be acquired simultaneously using multi-sensor fusion technology to construct a spatially distributed monitoring network. Damage localization is performed using the propagation delay and attenuation characteristics of the signal. To address the problems of severe dispersion and mode aliasing of guided wave signals, a filtering algorithm is used to adjust the filtering parameters based on real-time signal characteristics. Machine learning algorithms are introduced to identify and separate guided wave signals of different modes. The one-dimensional time-domain signal is converted into a two-dimensional time-frequency spectrum using time-frequency analysis methods to highlight the frequency domain anomalies caused by damage, providing richer and more accurate input data for probabilistic deep learning models.

[0044] It should be noted that this invention, by acquiring ultrasonic guided wave reflection signal data of the ground wire and converting it into spectral data, solves the problems of traditional detection methods relying on visual observation and local sampling analysis, resulting in low detection coverage and poor timeliness. Compared with the existing technology that manually extracts empirical features such as time delay and amplitude attenuation, this invention solves the problems of traditional methods being unable to effectively distinguish corrosion types and being susceptible to environmental noise interference leading to a high false detection rate through standardized signal acquisition and spectral conversion processes. In particular, the adaptive noise reduction module and time-frequency domain joint characterization can effectively cope with the acoustic wave scattering challenge brought about by the multi-strand strand structure of the steel strand, ensuring the accuracy and reliability of the input data for the subsequent probabilistic deep learning model. This invention not only improves the accuracy of damage feature extraction but also provides a solid data foundation for realizing real-time perception and risk warning of the corrosion status of overhead ground wires throughout their entire life cycle, avoiding damage identification deviations caused by signal quality issues.

[0045] In this embodiment of the invention, step S200, which involves constructing a probabilistic deep learning model based on the map data, includes the following steps B1-B3:

[0046] B1: Set the weights in the probabilistic deep learning model to parameters that follow a probability distribution.

[0047] B2: Make an approximate inference of the posterior distribution of the weights.

[0048] B3: Construct a loss function based on the variational evidence lower bound optimization principle.

[0049] It should be noted that, compared to deterministic deep learning, probabilistic deep learning quantifies model uncertainty by introducing a probability distribution. In the probabilistic deep learning framework, network weights and biases are no longer fixed values, but follow a certain probability distribution, and their distribution parameters are updated through a data-driven approach.

[0050] Specifically, in step B1, the weights in the probabilistic deep learning model are set as parameters that follow a probability distribution. The specific operation can be as follows:

[0051] Given an observation dataset D = {(x1,y1),(x2,y2),…,(xn,yn)}, according to Bayesian theory, the posterior distribution of the weights in a probabilistic neural network can be represented by the following formula:

[0052]

[0053] Where w is the weight in the probabilistic neural network; p(w) is the prior distribution of the weight, usually set as a standard normal distribution; p(D|w) is the likelihood function, representing the probability distribution of the observed dataset given the model parameters; p(D) is the marginal likelihood function, representing the overall probability of the observed dataset, which can be expressed as ∫p(D|w)p(w)dw; and p(w|D) is the posterior distribution, characterizing the probability distribution of the updated parameters based on the observed dataset.

[0054] It should be noted that the computational process of probabilistic deep learning includes three core steps: prior distribution selection, posterior distribution inference, and prediction uncertainty estimation. In prior distribution selection, the Gaussian distribution is used as the prior belief parameter because it possesses excellent mathematical properties such as differentiability and closed-form solutions, making it convenient for computation and inference. The mean (μ) and variance (σ) of the Gaussian distribution are... 2 First, the parameters can be manually set or initialized based on empirical data. Second, the posterior distribution inference of the parameters can be achieved through variational inference methods, that is, by using the variational lower bound optimization principle, the posterior inference is transformed into a variational distribution q. θ (w) Approximation problem of the posterior distribution p(w|D).

[0055] Specifically, step B2 involves approximating the posterior distribution of the weights, including the following steps B21-B22:

[0056] B21: Transform posterior inference into variational distribution q θ (w) The approximation problem of the posterior distribution p(w|D) can be specifically expressed as:

[0057]

[0058] Where θ is the control variational distribution q θ (w) is the set of parameters for shape and position, designed to approximate the posterior distribution of the model weights w as accurately as possible within a Bayesian framework; μ is the mean of the Gaussian distribution; σ 2 denoted as the variance of the Gaussian distribution; p(w|D) is the posterior distribution, representing the probability distribution of the parameters updated based on the observed dataset.

[0059] B22: Using Kullback-Leibler divergence (KLD) to measure the variational distribution q θ The smaller the difference between the distribution p(w) and the prior distribution p(w), the lower the Kullback-Leibler divergence. This can be specifically reflected by the following formula:

[0060]

[0061] Where KLD(q||p) is the Kullback-Leibler divergence from distribution q to distribution p; q is the variational distribution q θ (w); p is the prior distribution p(w); E is the mathematical expectation. By optimizing the parameter set θ (i.e. minimizing the Kullback-Leibler divergence between the two distributions), the similarity between the two distributions can be maximized.

[0062] Specifically, in step B22, the similarity between the two is maximized by optimizing the parameter set θ, which can be represented by the following formula:

[0063]

[0064] Where, θ * The optimal variational distribution parameters are denoted as .

[0065] Specifically, step B3 involves constructing a loss function based on the variational evidence lower bound optimization principle, including the following steps B31-B33:

[0066] B31: In the optimization of probabilistic deep learning models, when the parameter set θ only affects the posterior distribution p(w|D) or the variational distribution q θ When (w), the marginal likelihood function p(D), as a constant term independent of the parameter set θ, can be ignored in the optimization objective, and can be specifically expressed as:

[0067]

[0068] in, For variational distribution q θ The expectation under (w); p(D|w) is the likelihood function, representing the probability distribution of the observed dataset given the model parameters.

[0069] B32: Based on the variational evidence lower bound (ELBO), the above equation is equivalent to maximizing the variational evidence lower bound, and it is also used as the loss function, which can be specifically expressed as:

[0070]

[0071] Where L(w,θ) is the loss function.

[0072] B33: Given a training set (x, y), then Equivalent to Where p(x) is a known distribution, the loss function can be further simplified to:

[0073]

[0074] in, For all parameters wi The sum of the logarithms of the prior probabilities; For the observed data y i Given input x i and parameter w i The sum of log-likelihoods under given conditions; Variational distribution The sum of log probabilities; i is the number of the i-th sample.

[0075] It should be noted that the above loss function can be divided into two parts. The first part is the Kullback-Leibler divergence (KLD) between the variational distribution and the prior distribution, which represents the cost of model complexity. The second part is related to the data samples and is called the likelihood cost, which represents the degree to which the network fits the data.

[0076] In an optional implementation, in step S200, a probabilistic deep learning model is constructed based on the spectral data. Furthermore, a prior distribution selection mechanism can be used to adjust the hyperparameters of the prior weights based on the characteristic distribution of the ground wire damage data, thereby improving the model's adaptability to different damage types. At the same time, a hierarchical variational inference method is adopted, using different approximation strategies for different network layers to optimize computational efficiency and inference accuracy. Additionally, a temperature parameter adjustment mechanism is introduced to adjust the sharpness of the probability distribution during training, balancing the model's exploratory ability and convergence stability, and ensuring that the probabilistic deep learning model can effectively handle the signal characteristics caused by the multi-strand stranded structure of the steel strand.

[0077] In another optional implementation, in step S200, a probabilistic deep learning model is constructed based on the map data. The prediction results of multiple probabilistic models can be integrated, and the Bayesian model averaging method can be used to improve the reliability and robustness of the prediction. To address the class imbalance problem in ground wire damage identification, a loss weight mechanism is designed to adjust the weight coefficients in the loss function according to the frequency of occurrence of different damage types. At the same time, an uncertainty-aware regularization term is introduced to add constraints on prediction uncertainty to the loss function, prompting the model to give more conservative predictions in areas with high uncertainty. An online learning mechanism is also established so that the model can continuously update the parameter distribution based on newly collected field data, thereby improving its adaptability to changes in actual working conditions.

[0078] In this embodiment of the invention, the training process of the probabilistic deep learning model in step S200 includes the following steps C1-C3:

[0079] C1: Use reparameterization techniques to convert nodes into deterministic nodes.

[0080] C2: Use the Flipout estimator to eliminate noise by sampling a pseudo-random symbol matrix.

[0081] C3: Combine Bayesian backpropagation to update the parameters of the probabilistic deep learning model.

[0082] It should be noted that during the forward computation, the weights are sampled from an approximate posterior distribution; deep learning architectures need to compute gradients through backpropagation to update the parameters in the network; however, gradients cannot be directly obtained at the sampling points, so these samples cannot be differentiated. To solve this problem, the reparameterization trick and the Flipout method can be used, where the reparameterization trick transforms nodes with randomness into deterministic nodes.

[0083] Specifically, in step C1, the reparameterization technique is used to transform the nodes into deterministic nodes. The specific operation can be as follows:

[0084] For reparameterization techniques, the weights are first sampled from a parameterless distribution and then transformed according to a specific rule, which can be expressed as:

[0085]

[0086] Here, ε is a random noise term sampled from the standard normal distribution N(0,1). After transforming it based on the mean and variance, a weight w is generated, which solves the problem of gradient non-transferability caused by sampling.

[0087] It should be noted that, since the network is trained by sharing mini-batch training samples with the same perturbation in this invention, and there is a high variance problem in the mini-batch training samples, the Flipout estimator is used, where Flipout is an efficient weight perturbation method that achieves nearly independent weight perturbation in mini-batch samples.

[0088] Specifically, in step C2, the Flipout estimator is used to eliminate noise by sampling a pseudo-random symbol matrix. The specific operation can be as follows:

[0089] Using the Flipout estimator, two assumptions are made about the weight distribution: first, the perturbations to the weights are independent; second, the distribution of the perturbations is symmetric about zero, which can be specifically expressed as:

[0090] w (j) =w′ (j) +Δw j ;

[0091] Where j is the j-th layer in the probabilistic deep network; w (j) The updated weights; w′ (j) The mean; Δw jThe Flipout estimator randomly flips the symmetric perturbation of the weights by sampling a series of pseudo-random symbol matrices to eliminate noise within mini-batches.

[0092] Furthermore, for the i-th sample in a batch, its random perturbation can be specifically expressed as:

[0093]

[0094] Where, r i and s i It is a random vector; This is a perturbation that is sampled only once for the entire batch; This is for element-wise multiplication.

[0095] It should be noted that, since deep learning networks use backpropagation for deterministic parameter updates, this invention combines Bayesian backpropagation (BBB) ​​with weight perturbation to update parameters in order to adapt to probabilistic deep learning networks.

[0096] Specifically, step C3 combines Bayesian backpropagation to update the parameters of the probabilistic deep learning model. The specific operations can be as follows:

[0097] Define the number of training rounds N, the initial weights ε0 and σ0.

[0098] The probability distribution parameters θ are constructed using the weight perturbation method. Based on the loss function L, the gradient of each parameter is obtained by taking the derivative of each parameter.

[0099] The gradient descent method is used to update the parameters based on the gradient descent index α.

[0100] It should be noted that this invention, by constructing a probabilistic deep learning model, transforms the fixed weights in deterministic deep learning into a probability distribution, thereby achieving a quantitative representation of model uncertainty. Compared with existing technologies that are limited by the single distribution characteristics of finite samples and have insufficient generalization ability, this invention, by introducing Bayesian theory and variational inference methods, solves the dilemma of traditional methods lacking a quantitative representation of the uncertainty of prediction results and failing to meet the high-confidence decision-making requirements of power systems. In particular, through the use of reparameterization techniques and Flipout estimators, it effectively solves the gradient propagation and high variance problems in probabilistic model training, not only improving the prediction stability of the model under various operating conditions, but also providing a solid theoretical foundation for subsequent damage identification and uncertainty assessment, avoiding the risk of misjudgment caused by poor model output stability.

[0101] In this embodiment of the invention, step S300 involves obtaining damage type identification results based on a probabilistic deep learning model and atlas data, and determining the uncertainty quantification index of the damage type identification results, including the following steps D1-D3:

[0102] D1: Use a probabilistic deep learning model to extract features from the graph data.

[0103] D2: Based on the extracted features, identify the damage type of the ground wire and obtain the damage type identification result.

[0104] D3: Calculate the confidence level of the damage type identification result based on the probabilistic deep learning model to obtain the uncertainty quantification index of the damage type identification result.

[0105] Specifically, in step D1, a probabilistic deep learning model is used to extract features from the map data. The specific operations can be as follows:

[0106] The probabilistic sensing layer, based on a probabilistic deep learning model, evaluates signal uncertainty through random weight distribution.

[0107] By combining the feature extraction capabilities of deep convolutional layers for time-frequency maps, feature mining is performed on the spectral data jointly represented in the time and frequency domains through a multi-layer convolutional neural network.

[0108] Specifically, the damage type identification results in step D2 may include:

[0109] Abnormal waveguide reflection peaks – typically correspond to strong reflection phenomena caused by broken strands.

[0110] Energy attenuation profiles for specific frequency bands – indicating scattering loss caused by pitting or cracks.

[0111] It should be noted that, through the fusion analysis of deep features and probability distribution, the probabilistic deep learning model can distinguish multiple damage types such as pitting, cracks, and broken strands. In this process, the probability perception layer continuously quantifies the reliability of features and filters out false alarm signals caused by sudden changes in wind speed or electromagnetic pulses, ensuring the accuracy of damage type identification.

[0112] For example, in step D3, the confidence level of the damage type identification result is calculated based on the probabilistic deep learning model. The specific operation can be as follows:

[0113] By fusing deep features and uncertainty quantification results through a probabilistic deep learning model, and utilizing the probability distribution characteristics of the weights in the probabilistic deep learning model, the confidence level of the identification results for each damage type is calculated.

[0114] In an optional implementation, in step S300, the damage type identification result is obtained based on the probabilistic deep learning model and the map data. In addition, the local detail features and global structural features of the map data can be extracted simultaneously through multi-scale feature fusion technology. An attention mechanism is used to highlight the feature representation of the damaged area, and the weight allocation of different frequency bands and time periods is adjusted. Different feature extraction strategies are designed for different damage types, and a quantitative assessment system for the severity of damage is established. The confidence interval of the damage degree is calculated through the statistical characteristics of the probability distribution. At the same time, a time series analysis method is introduced to track the evolution trend of the damage, and the development trajectory of the damage is predicted by combining historical data to provide decision support for preventive maintenance.

[0115] In another optional implementation, the damage type identification result is obtained based on the probabilistic deep learning model and the map data in step S300. The prediction results of multiple probabilistic deep learning sub-models can also be combined by ensemble learning methods. The robustness of identification can be improved by adopting a voting mechanism or weighted average method. A multi-label classification model is established for complex damage scenarios. At the same time, the coexistence of multiple damage types is identified. Domain adaptation technology is introduced to enable the model to adapt to different environmental conditions and equipment states. An uncertainty propagation mechanism is established to transfer the uncertainty in the feature extraction stage to the final identification result and provide end-to-end confidence assessment.

[0116] It should be noted that this invention achieves intelligent feature extraction and damage type identification of spectral data through a probabilistic deep learning model, transforming the traditional feature extraction rules, which rely on manual experience, into a data-driven learning process. Compared with existing deep learning methods that can only provide deterministic prediction results and lack uncertainty quantification, this invention solves the problems of traditional methods being unable to quantify prediction reliability and support high-confidence decisions by introducing a probabilistic framework. In particular, by identifying two key damage fingerprints—abnormal guided wave reflection peaks and energy attenuation profiles in specific frequency bands—it effectively distinguishes multiple damage types such as broken strands, pitting, and cracks. At the same time, it filters false alarm signals caused by environmental noise through a feature reliability quantification mechanism at the probabilistic layer. This not only improves the accuracy and reliability of damage identification but also provides a quantitative confidence assessment for power system operation and maintenance decisions, avoiding decision risks caused by the uncertainty of prediction results and ensuring the scientific and practical nature of ground wire damage diagnosis.

[0117] In this embodiment of the invention, step S400, which determines the ground wire damage diagnosis result based on the damage type identification result and in conjunction with the uncertainty quantification index, includes the following steps E1-E2:

[0118] E1: Determine the type and location information of the damage based on the damage type identification results.

[0119] E2: Combines information on the type and location of the damage with uncertainty quantification indicators to generate ground wire damage diagnosis results.

[0120] Specifically, in step E1, the type and location information of the damage are determined based on the damage type identification results. The specific operation can be as follows:

[0121] The abnormal guided wave reflection peak was identified as a broken strand type of damage, and the precise location coordinates of the broken strand were calculated based on the propagation delay of the reflected signal.

[0122] For the energy attenuation profile of a specific frequency band, the damage is determined to be either pitting or crack type, and the spatial distribution of the damage is determined by analyzing the frequency domain characteristics of the attenuation mode.

[0123] Specifically, step E2 combines the damage type and location information with uncertainty quantification indicators to generate ground wire damage diagnosis results, including the following steps E21-E22:

[0124] E21: Compare the uncertainty quantification index with the preset threshold to determine whether to generate a structured diagnostic conclusion.

[0125] E22: Compare the confidence level of the structured diagnostic conclusion with the preset alarm threshold to determine whether to automatically trigger the operation and maintenance response mechanism.

[0126] For example, determining whether to generate a structured diagnostic conclusion in step E21 can be done as follows:

[0127] The confidence level of the damage type identification result is compared with a preset threshold. If the confidence level of the damage type identification result is higher than the preset threshold, a triplet structured diagnostic conclusion containing damage type, location coordinates, and confidence assessment is generated.

[0128] If the uncertainty quantification index is lower than the preset threshold, no structured diagnostic conclusion will be generated to avoid false alarms due to excessive model uncertainty.

[0129] For example, in step E22, determining whether to automatically trigger the operation and maintenance response mechanism can be done as follows:

[0130] The confidence level assessment in the generated structured diagnostic conclusion is compared with the preset alarm threshold. If the confidence level of the structured diagnostic conclusion exceeds the preset alarm threshold, the operation and maintenance response mechanism is automatically triggered.

[0131] If the confidence level of the structured diagnostic conclusion does not reach the alarm threshold but is higher than the monitoring threshold, the operation and maintenance response mechanism will not be triggered.

[0132] In one optional implementation, step S400 determines the ground wire damage diagnosis result based on the damage type identification result and in conjunction with the uncertainty quantification index. Alternatively, a damage severity assessment system can be established to classify the identified damage types according to their impact on the safe operation of the ground wire. The risk assessment weight can be adjusted based on the criticality of the damage location (such as whether it is located near a tower or crosses an important facility). Statistical analysis of historical damage data can be introduced to establish a damage evolution trend prediction model, predict the future development trajectory based on the current damage status, and consider the impact of environmental factors (such as climate conditions and load levels) on damage development to adjust the urgency of the diagnosis conclusion, providing more comprehensive information support for operation and maintenance decisions.

[0133] In another optional implementation, in step S400, the ground wire damage diagnosis result is determined based on the damage type identification result and the uncertainty quantification index. Alternatively, by integrating multi-sensor data fusion technology, the ultrasonic guided wave detection result can be comprehensively analyzed with data from other monitoring methods (such as vibration monitoring, temperature monitoring, etc.) to improve the reliability of the diagnosis result. At the same time, the preset threshold and alarm threshold can be automatically adjusted based on the accuracy feedback of historical diagnosis results to establish a closed-loop optimization mechanism. The damage location, type and severity can be intuitively displayed in the form of charts, three-dimensional models, etc., to facilitate quick understanding and decision-making by maintenance personnel.

[0134] It should be noted that this invention generates comprehensive ground wire damage diagnosis results by combining damage type identification results and uncertainty quantification indicators, closely integrating the predictive capabilities of probabilistic deep learning models with actual operation and maintenance needs. Compared with existing technologies that only provide simple damage detection results and lack comprehensive diagnosis and decision support, this invention solves the problems of traditional methods being unable to quantify damage risks and support accurate operation and maintenance decisions by establishing structured diagnostic conclusions and a hierarchical early warning mechanism. In particular, through threshold comparison and alarm threshold mechanisms, it effectively balances the sensitivity and specificity of damage detection, avoiding resource waste caused by false alarms and safety risks caused by missed alarms. This not only improves the intelligence level of ground wire damage diagnosis but also provides a scientific decision-making basis for preventive maintenance of power systems, ensuring the safe and reliable operation of overhead ground wires and improving the efficiency and economy of power system operation and maintenance.

[0135] In summary, this invention effectively solves the strong acoustic wave scattering effect and dispersion problem caused by the multi-strand strand structure of steel strands through noise reduction processing of ultrasonic guided wave reflection signal data and joint time-frequency domain characterization and transformation. This improves the signal-to-noise ratio of the time-domain reflected wave packet and overcomes the shortcomings of low coverage and poor timeliness in traditional manual inspection, achieving early and accurate detection of hidden corrosion such as inner strand breaks and intergranular cracks. By constructing a probabilistic deep learning model, the network weights are transformed into probability distribution parameters, and the calculation process is simplified based on the variational evidence lower bound optimization principle. This not only avoids the high-dimensional integral burden of marginal probability calculation in traditional Bayesian inference, improving computational efficiency, but also achieves K... The L-divergence regularization term implements adaptive constraints on model parameters, effectively suppressing overfitting and ensuring model stability under field conditions. By integrating deep feature extraction and uncertainty quantification mechanisms, it accurately identifies two key damage fingerprints: abnormal guided wave reflection peaks and energy attenuation profiles in specific frequency bands. This enables precise differentiation of various damage types, such as pitting, cracks, and broken strands. At the same time, the asymmetric properties of KL divergence ensure that the approximate distribution avoids covering the zero-probability region of the true posterior, improving the reliability of uncertainty estimation. This provides high-confidence quantitative decision support for power system operation and maintenance, effectively reducing the false detection rate caused by environmental noise interference.

[0136] Example 3, the third embodiment of the present invention, provides a ground wire damage identification system based on probabilistic learning, comprising: a ground wire damage data module for acquiring ultrasonic guided wave reflection signal data of the ground wire and obtaining spectral data based on the ultrasonic guided wave reflection signal data; a probabilistic deep learning module for constructing a probabilistic deep learning model based on the spectral data; a damage identification module for obtaining damage type identification results based on the probabilistic deep learning model and the spectral data, and determining the uncertainty quantification index of the damage type identification results; and a damage diagnosis module for determining the ground wire damage diagnosis results based on the damage type identification results and the uncertainty quantification index.

[0137] Example 4 is the fourth embodiment of the present invention, which differs from the previous three embodiments in that: Figure 2As shown, if the function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0138] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0139] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0140] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0141] Example 5, referring to Table 1, is the fifth embodiment of the present invention, providing a ground wire damage identification method based on probabilistic learning. To verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0142] This embodiment constructs a comparative verification platform and uses 2000 sets of overhead ground wire damage test samples to conduct a comprehensive performance evaluation of traditional CNN, pure probabilistic models and the probabilistic deep learning model proposed in this invention, verifying the effectiveness of the technical solution of this invention. Specific data are shown in Table 1.

[0143] Table 1 Comparison data of different models

[0144]

[0145] As shown in Table 1, in terms of damage recognition performance, the probabilistic deep learning model of this invention achieves an average accuracy of 96.5% for typical steel strand damage, which is 13.6% higher than the traditional CNN method and 5.6% higher than the pure probabilistic model. Among them, the strand breakage recognition rate reaches 97.5%, which is 8.3% higher than the traditional CNN. This breakthrough is of great significance for preventing steel strand breakage accidents, because strand breakage in steel strands can easily lead to breakage accidents. This invention effectively quantifies the modal aliasing interference through the front-end probabilistic layer and accurately captures the strong reflection characteristics of ultrasonic guided waves at the strand breakage point.

[0146] In environmental adaptability tests, the recognition accuracy of this invention reached 92.7% and 89.5% in harsh environments such as strong winds (level 8, SNR=15dB) and substation electromagnetic interference (SNR=10dB), respectively, which is 16.3% and 20.6% higher than that of traditional CNNs. Its advantages stem from the active filtering mechanism of Bayesian weight distribution for environmental noise and the effective suppression of high-frequency vibration by Flipout perturbation. Moreover, the uncertainty increase is only 18% when the environment deteriorates, which is 38% lower than the 29% of the pure probability model, meeting the high confidence decision-making requirements of power systems.

[0147] In terms of computational efficiency, through the optimized architecture design of 1 probabilistic layer + 2 ordinary layers, the deployment size of the model of this invention on the edge device Jetson Nano is only 10.2MB, which is 76% smaller than that of the pure probabilistic model; the single-frame inference latency is 32.9ms, which is 52% better than that of the pure probabilistic model, meeting the requirements of real-time monitoring of power systems (≤50ms power standard); the peak power consumption is only 9.1W, which is 28% lower than that of the pure probabilistic model, making it suitable for outdoor solar power supply environments; the training speed reaches 24.7 epochs / hour, which is 75% better than that of the pure probabilistic model, indicating that this invention improves the model deployment efficiency.

[0148] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A ground wire damage identification method based on probabilistic learning, characterized in that: include, Acquire ultrasonic guided wave reflection signal data of the ground wire, and obtain spectral data based on the ultrasonic guided wave reflection signal data; Construct a probabilistic deep learning model based on the aforementioned map data; Based on the probabilistic deep learning model and the atlas data, the damage type identification result is obtained, and the uncertainty quantification index of the damage type identification result is determined. The grounding wire damage diagnosis result is determined based on the damage type identification result and the uncertainty quantification index.

2. The ground wire damage identification method based on probabilistic learning as described in claim 1, characterized in that: The process of acquiring ultrasonic guided wave reflection signal data of the ground wire and obtaining spectral data based on the ultrasonic guided wave reflection signal data includes: Collect ultrasonic guided wave reflection signal data from the ground wire; The ultrasonic guided wave reflection signal data is subjected to noise reduction processing; The noise-reduced ultrasonic guided wave reflection signal data is converted into spectral data.

3. The ground wire damage identification method based on probabilistic learning as described in claim 2, characterized in that: Constructing a probabilistic deep learning model based on the aforementioned map data includes: The weights in the probabilistic deep learning model are set to parameters that follow a probability distribution; An approximate inference is made about the posterior distribution of the weights; The loss function is constructed based on the variational evidence lower bound optimization principle.

4. The ground wire damage identification method based on probabilistic learning as described in claim 3, characterized in that: Based on the probabilistic deep learning model and the atlas data, damage type identification results are obtained, and the uncertainty quantification index of the damage type identification results is determined, including: The probabilistic deep learning model is used to extract features from the map data; Based on the extracted features, the damage type of the ground wire is identified, and the damage type identification result is obtained. The confidence level of the damage type identification result is calculated based on the probabilistic deep learning model to obtain the uncertainty quantification index of the damage type identification result.

5. The ground wire damage identification method based on probabilistic learning as described in claim 4, characterized in that: The grounding wire damage diagnosis result is determined based on the damage type identification result and the uncertainty quantification index, including: The type and location information of the damage are determined based on the damage type identification results; The grounding wire damage diagnosis result is generated by combining the damage type and location information with the uncertainty quantification index.

6. The ground wire damage identification method based on probabilistic learning as described in claim 5, characterized in that: The training process of the probabilistic deep learning model includes: The node is transformed into a deterministic node using a reparameterization technique; The Flipout estimator is used to eliminate noise by sampling a pseudo-random symbol matrix; The parameters of the probabilistic deep learning model are updated by combining the Bayesian backpropagation method.

7. The ground wire damage identification method based on probabilistic learning as described in claim 6, characterized in that: By combining the damage type and location information with the uncertainty quantification index, a ground wire damage diagnosis result is generated, including: The uncertainty quantification index is compared with a preset threshold to determine whether a structured diagnostic conclusion is generated. The confidence level of the structured diagnostic conclusion is compared with the preset alarm threshold to determine whether the operation and maintenance response mechanism should be automatically triggered.

8. A ground wire damage identification system based on probabilistic learning, employing the ground wire damage identification method based on probabilistic learning as described in any one of claims 1 to 7, characterized in that: include, The grounding wire damage data module is used to acquire ultrasonic guided wave reflection signal data of the grounding wire and obtain spectral data based on the ultrasonic guided wave reflection signal data; The probabilistic deep learning module is used to build probabilistic deep learning models based on graph data. The damage identification module is used to obtain damage type identification results based on probabilistic deep learning models and atlas data, and to determine the uncertainty quantification index of the damage type identification results. The damage diagnosis module is used to determine the ground wire damage diagnosis result based on the damage type identification result and the uncertainty quantification index.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the ground wire damage identification method based on probabilistic learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the ground wire damage identification method based on probabilistic learning as described in any one of claims 1 to 7.

Images