Urban underground cable operation and maintenance method and system based on virtual reality

Abstract

The invention relates to the field of electric digital data processing, in particular to an urban underground cable operation and maintenance method and system based on virtual reality. The method comprises the steps of collecting actual cable data to mark a corresponding operation state, training a data expansion model by adopting a generative adversarial network algorithm based on adaptive noise injection, training a classification model by using the actual cable data and synthetic cable data, and analyzing new actual cable data by using the trained classification model to obtain the operation state of a cable. And operation and maintenance work is completed. The existing urban underground cable operation and maintenance method and system have the problem of inaccurate judgment on the operation state of an underground cable. According to the urban underground cable operation and maintenance method and system based on virtual reality provided by the invention, the operation state of the underground cable can be accurately judged.

Description

Technical Field

[0001] The present invention relates to the field of electrical digital data processing, and more particularly to a method and system for operation and maintenance inspection of urban underground cables based on virtual reality. Background Art

[0002] A pipe gallery is an underground passage structure, whose main function is to provide accommodation space for various urban pipelines such as underground cables and play a protective role.

[0003] Once a problem occurs in the underground cable, it will affect the normal operation of the city. However, scientific management of the underground cable has always been a severe challenge because manual inspection is inefficient and it is not easy to inspect some places where the cables are intertwined. Moreover, the underground cable is relatively long, and it is difficult for the inspection personnel to grasp the operation status of the underground cable in real time.

[0004] With the gradual development of technologies such as artificial intelligence, machine vision, and the Internet of Things, new methods and systems for operation and maintenance inspection of urban underground cables that can implement all-weather remote monitoring of underground cables have emerged - the intelligent pipe gallery operation and maintenance system. The intelligent pipe gallery operation and maintenance system deploys sensing devices, video monitoring, gas monitoring, personnel intrusion management, cable joint temperature monitoring and other facilities in the cable pipe gallery (a pipe gallery specifically used to place underground cables). Combining with the intelligent pipe gallery operation and maintenance method, it restores the real scene of the pipe gallery through 3D modeling technology, establishes an interactive virtual reality environment, supports users to perform immersive browsing and operation through a head-mounted device or a projection system, realizes comprehensive collection and centralized monitoring of cable data in the pipe gallery. The intelligent pipe gallery operation and maintenance system also has functions such as alarm and trend analysis query. With the help of this platform, the operation and maintenance personnel can comprehensively master the internal environment of the pipe gallery and the operation status of the cables, and realize the informatization and intelligent management of the cables and the pipe gallery.

[0005] However, there are also some problems with the existing methods and systems for operation and maintenance inspection of urban underground cables. First, the basic generative adversarial network algorithm is mostly used to expand the collected cable data, lacking special optimization for rare state data. The diversity and distribution consistency of the generated virtual samples are poor, and it is difficult to improve the classification performance of low-frequency categories (such as fault states). Therefore, using the data expanded in this way to train the classification model will affect the training effect of the classification model. Second, when training the classification model, traditional classification algorithms (such as support vector machines, random forests, etc.) are not optimized for the class imbalance problem. Therefore, low-frequency categories are easily ignored, resulting in a decline in classification accuracy and insufficient processing ability for high-dimensional feature data, and it is difficult to effectively capture the complex non-linear relationships in the cable state attributes, making the training effect of the classification model poor, resulting in inaccurate judgments on the operation status of underground cables.

[0006] Therefore, the existing methods and systems for operation and maintenance inspection of urban underground cables have the problem of inaccurate judgments on the operation status of underground cables. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to overcome the above-mentioned disadvantages of the prior art and provide a virtual reality-based operation and maintenance inspection method and system for urban underground cables that can make a relatively accurate judgment on the operating status of underground cables.

[0008] To solve the above technical problem, the virtual reality-based operation and maintenance inspection method for urban underground cables provided by the present invention includes:

[0009] S1. Obtain the actual cable data of each underground cable, and manually label the operating status of each actual cable data;

[0010] The operating status includes normal, warning, and fault;

[0011] S2. Input the labeled actual cable data into a data augmentation model, train the data augmentation model using a generative adversarial network algorithm based on adaptive noise injection to obtain a trained data augmentation model, use the trained data augmentation model to generate synthetic cable data, and use the actual cable data and the synthetic cable data as training data;

[0012] S3. Input the training data into a classification model, train the classification model using the random forest algorithm based on relative entropy to obtain a trained classification model;

[0013] S4. Collect new actual cable data, input the new actual cable data into the trained classification model to obtain the operating status corresponding to the new actual cable data, the operating status includes normal, warning, and fault. When the operating status is a warning, conduct an inspection, and when the operating status is a fault, perform maintenance.

[0014] As a further improvement of the present invention: The training of the data augmentation model using the generative adversarial network algorithm based on adaptive noise injection in S2 includes:

[0015] S201. Initialize the parameters of the generator and discriminator of the generative adversarial network;

[0016] S202. Design a loss function according to the dynamic change of the cable data distribution during the training process, and the calculation formula is:

[0017]

[0018] In the formula, is the loss function of the generator; D c () is the discriminator function; G c() is a generator function; z is the input noise, specifically a random vector sampled from the latent space; is the loss function of the discriminator; x ce is the actual cable data sample;

[0019] S203. Dynamically adjust the weight parameters of the generator and the discriminator according to the feedback of the discriminator. At the same time, use the mechanism of topological structure difference and adaptive noise injection to control the learning rate and noise pattern, which is expressed as:

[0020]

[0021] In the formula, ΔT c is the topological difference; nc is the number of samples input to the generative adversarial network in the current batch;

[0022] Re() is the ReLU activation function used to extract the structural information of the sample extraction, G c () is the generator function, is the i z th noise vector, is the i z th actual cable data sample; i z is the cable data sample index and is the same as the noise vector index;

[0023] S204. Inject adaptive noise into the generator. Determine the noise intensity through the feedback of the discriminator on the synthetic cable data. The calculation formula is:

[0024] N c (z,α c ) = z + α c ·∈ c ,

[0025]

[0026] In the formula, N c (z,α c ) is the adaptive noise injection function; z is the input noise; α c is the noise intensity coefficient; ∈ c is the random noise of the normal distribution, is the normal distribution with a mean of 0 and a variance of the identity matrix. I is the identity matrix, represents the normal distribution;

[0027] S205. After obtaining the output of the generator, update the generator and the discriminator according to the dynamic learning rate. In the optimization of the generator and the discriminator, balance the diversity generated by the generator and the discrimination ability of the discriminator by dynamically adjusting the weight update strategy. The calculation formula is:

[0028]

[0029] In the formula, is the weight of the generator, and ← is the parameter update symbol; η c is the dynamic learning rate of the generative adversarial network; is the gradient of the generator loss function with respect to its weight, is the weight of the discriminator, is the gradient of the discriminator loss function with respect to its weight;

[0030] S206. Repeat iterations of S202 to S205 until a preset stop iteration condition is satisfied.

[0031] Preferably, in S202, the D c (G c (z)) is calculated as follows:

[0032]

[0033] In the formula, D c () is the discriminator function; G c () is the generator function; z is the input noise; Sig() is the Sigmoid activation function; is the weight of the discriminator, is the bias of the discriminator.

[0034] Preferably, in S204, the noise intensity coefficient α c is calculated as follows:

[0035]

[0036] In the formula, β c is the adjustment sensitivity parameter; τ ce is the target threshold.

[0037] As a further improvement of the present invention: The training of the classification model using the random forest algorithm based on relative entropy in S3 includes:

[0038] S301. Initialize the number and depth of decision trees required for the random forest;

[0039] S302. Screen features by using Gaussian curvature;

[0040] S303. In the process of constructing the decision tree, use a relative entropy loss function dynamically adjusted based on the distribution characteristics of cable data. The formula for the relative entropy loss function is:

[0041]

[0042] In the formula, is the relative entropy loss function; K ue is the number of classes; is the probability of the k-th class predicted by the model; is the true probability distribution of the k-th class, where k is the class index;

[0043] S304. Train each decision tree based on the adjusted loss function, and select the optimal splitting point that maximizes the inter-class separation when splitting nodes. Determine the optimal partitioning method during the growth of the decision tree. The method for optimizing node splitting is expressed as:

[0044]

[0045] In the formula, is the optimal splitting parameter; is the node splitting optimization function based on the adjusted loss function; θ ury is the set of candidate splitting parameters;

[0046] S305. Synthesize several independently trained decision trees into a random forest and perform global prediction by majority voting. The calculation formula for synthesizing the global model is:

[0047]

[0048] In the formula, is the final prediction result of the random forest; the mode() function is the majority voting function; Tu i is the i-th decision tree; is the input feature of the decision tree in the random forest; i is the decision tree index, is the number of decision trees in the random forest;

[0049] S306. Update its weight by analyzing the contribution of each feature to the output entropy of the model. The calculation formula for entropy weight adjustment of dynamic feature points is:

[0050]

[0051] In the formula, is the weight of the j-th feature, where j is the feature index; update() is the weight update function; is the contribution of the j-th feature to the change of the loss function;

[0052] S307. Repeat iterations of S302 - S306 until the preset stop iteration condition is met.

[0053] Preferably, in S304, the calculation formula is:

[0054]

[0055] In the formula, Left(θ ury ) is the data point in the left sub - decision tree of the decision tree under the candidate splitting parameter set; i w is the data point index; is the class probability of the i - th w data point predicted by the decision tree in the random forest; is the true class probability of the i - th w data point; Right(θ ury ) is the data point in the right sub - decision tree of the decision tree under the candidate splitting parameter set.

[0056] Preferably, the calculation formula of in S306 is:

[0057]

[0058] In the formula, update() is the weight update function; is the contribution of the j - th feature to the change of the loss function; is the weight of the j - th feature; is the natural constant; β uac is the Ricci curvature adjustment coefficient; is the Ricci curvature of the j - th feature; is the learning rate of the weight update at the t - th iteration, t is the number of iterations, is the partial derivative symbol, is the adjusted loss function, is the change amount of the adjusted loss function.

[0059] The present invention also provides a virtual - reality - based urban underground cable operation and maintenance inspection system, which applies the above - mentioned virtual - reality - based urban underground cable operation and maintenance inspection method, and includes:

[0060] A data acquisition unit, which is used to acquire actual cable data;

[0061] A data storage and expansion unit, which is used to store data and generate synthetic cable data based on the actual cable data through a data expansion model;

[0062] A data processing unit, which is used to label the corresponding operation status of the input cable data through a classification model.

[0063] The beneficial effects of the present invention are as follows: The virtual - reality - based urban underground cable operation and maintenance inspection method provided by the present invention can make a relatively accurate judgment on the operation status of underground cables.

[0064] This method uses a generative adversarial network algorithm based on adaptive noise injection to train a data augmentation model. Adaptive noise is injected into the generator to generate diverse synthetic cable data. The discriminator determines the noise intensity based on the quality feedback of the synthetic cable data, achieving dynamic regulation of the noise intensity. This optimizes the diversity exploration ability of the generator in the early stage and improves the authenticity and coverage of the synthetic cable data. In this way, higher-quality data for training the classification model is obtained. This method uses a random forest algorithm based on relative entropy to train the classification model. The loss function is dynamically adjusted to adapt to class-imbalanced data, strengthening the discriminative ability of the classification model for rare classes. This improves the classification ability of the classification model and ultimately enhances the accuracy of its judgment on the operating status of underground cables. Therefore, the virtual reality-based urban underground cable operation and maintenance inspection system applying this method also makes relatively accurate judgments on the operating status of underground cables. Description of the Drawings

[0065] Figure 1 It is a comparison graph of the convergence of the loss functions of the generator and the discriminator in the present invention;

[0066] Figure 2 It is a comparison graph of the distribution of synthetic cable data in the present invention;

[0067] Figure 3 It is a graph showing the influence of the number of decision trees on the model accuracy in the present invention;

[0068] Figure 4 It is a graph showing the influence of different feature selection methods on the model performance in the present invention. Detailed Embodiment

[0069] The following further elaborates on the detailed embodiment of the present invention with reference to the drawings.

[0070] The virtual reality-based urban underground cable operation and maintenance method provided by the present invention includes:

[0071] S1. Obtain the actual cable data of each underground cable and manually label the operating status of each actual cable data;

[0072] The actual cable data includes cable temperature for reflecting the thermal state of the cable during operation, cable humidity for reflecting the humidity level of the surrounding environment of the cable, cable voltage for reflecting the magnitude of the voltage carried by the cable, cable current for reflecting the current-carrying capacity of the cable, cable insulation status for reflecting whether the insulation performance of the cable is good, the abscissa of the cable position and the ordinate of the cable position for accurately positioning the cable on the urban map, depth information of the cable position for indicating the burial depth of the cable, cable type for reflecting whether the cable is high voltage or low voltage, and maintenance records for reflecting the historical maintenance information of the cable.

[0073] Manual annotation is carried out by technical experts in the field.

[0074] The operating status includes normal, warning, and fault; normal means the cable is in good condition; warning means there are potential risks in the cable, and it is recommended to check it in the near future; fault means the cable is damaged and needs to be repaired immediately;

[0075] S2. Input the annotated actual cable data into the data augmentation model, and use the generative adversarial network algorithm based on adaptive noise injection to train the data augmentation model to obtain the trained data augmentation model. Use the trained data augmentation model to generate synthetic cable data, and use the actual cable data and the synthetic cable data as training data;

[0076] The generative adversarial network algorithm based on adaptive noise injection means that during the training process, the noise intensity is dynamically regulated using the feedback of the discriminator. By improving the noise injection method, it can adaptively adjust the injection of noise according to data characteristics or model status, etc., to enhance the diversity exploration ability of the generator in the early stage of training, achieve more stable and efficient adversarial learning, and provide reliable support for generating high-quality cable data.

[0077] S201. Initialize the parameters of the generator and discriminator of the generative adversarial network; initialize the weights in the model using a random distribution to provide an expandable initial condition for subsequent optimization, expressed as:

[0078]

[0079] In the formula, is the weight of the generator, ~ follows a specific distribution, is the normal distribution, is the normal distribution with a mean of zero and a variance of the identity matrix, I is the identity matrix, is the weight of the discriminator;

[0080] S202. Design the loss function according to the dynamic changes in the cable data distribution during the training process, including the loss parts of the generator and the discriminator, to enhance the sensitivity and robustness of the network to the changes in the cable data distribution. The calculation formula is:

[0081]

[0082] In the formula, is the loss function of the generator, representing the optimization goal of the generator against the discriminator; D c () is the discriminator function, representing the probability of judging the input cable data as real or generated; G c () is the generator function, representing the mapping relationship of synthesizing cable data from noise; z is the input noise, specifically a random vector sampled from the latent space; is the loss function of the discriminator, representing the optimization goal for the discriminator to distinguish between actual cable data and synthetic cable data; x ce is a sample of actual cable data;

[0083] D c (G c (z)) is calculated as follows:

[0084] Based on the output of the discriminator, the probability estimation of the synthetic cable data is realized. The discrimination process of the discriminator for the synthetic cable data is expressed as:

[0085]

[0086] In the formula, D c () is the discriminator function; G c () is the generator function; z is the input noise; Sig() is the Sigmoid activation function, a non-linear function that maps the input to the interval (0,1); is the weight of the discriminator, is the bias of the discriminator.

[0087] The output of the generator is calculated based on adaptive noise. When synthesizing cable data, due to the diversity and complexity of cable data attributes (such as humidity, voltage, current), these attributes may vary depending on different cable operating states and environmental conditions. By dynamically controlling the noise intensity, the generated cable data can cover a wider distribution range. The calculation formula is:

[0088]

[0089] In the formula, is the generated cable data, representing the mapping output of the generator for the input noise; G c () is the generator function, z is the input noise, tanh() is the hyperbolic tangent activation function, is the weight of the generator; N c (z,α c ) is the adaptive noise injection function, representing additional random perturbations on the basis of the original noise; α c is the noise intensity coefficient, representing the dynamic weight superimposed on the original noise during the generation process; is the bias of the generator.

[0090] S203. To achieve self-balanced iterative updates, the weight parameters of the generator and discriminator are dynamically adjusted according to the discriminator feedback. At the same time, mechanisms for topological structure differences and adaptive noise injection are used to control the learning rate and noise pattern, thereby obtaining a more stable adversarial learning effect. The calculation of topological differences can effectively evaluate the distance between synthetic cable data and real data in the topological space (i.e., evaluate the consistency of the spatial attributes of synthetic cable data and real data using topological structure differences). By controlling the topological structure differences between synthetic cable data and real data, the consistency of the spatial attributes of synthetic cable data is ensured, and at the same time, the reliability of synthetic cable data in the cable status monitoring task is improved, expressed as:

[0091]

[0092] In the formula, ΔT c is the topological difference, representing the distance between synthetic cable data and actual cable data in the topological space; nc is the number of samples input to the generative adversarial network in the current batch; ∥∥ is the L2 norm, and Re() is the ReLU activation function used to extract the structural information of the samples. G c () is the generator function, is the i z th noise vector, is the i z th actual cable data sample; i z is the cable data sample index and is the same as the noise vector index;

[0093] The learning rate is dynamically adjusted according to the obtained topological difference to avoid overfitting or insufficient training that may be caused by a fixed learning rate. For example, cables with different maintenance records may have significant differences in voltage or insulation status. The dynamic learning rate can better adapt to these complexities and accelerate convergence. The calculation formula is:

[0094] η c = η 0 exp(-λ c ΔT c ),

[0095] In the formula, η c is the dynamic learning rate of the generative adversarial network, representing the control of the parameter update step size at different training times, set to 0.03; η 0 is the initial learning rate of the generative adversarial network; λ c is the adjustment intensity parameter, representing the sensitivity to topological differences, set to 0.3; ΔT c is the topological difference.

[0096] S204. To enhance the diversity exploration ability of the generator in the early stage of training and overcome the fluctuations in generation quality, adaptive noise is injected into the generator. The noise intensity is determined by the discriminator's feedback on the synthetic cable data, improving the restoration ability of the synthetic cable data at the detail level. For example, in the small fluctuations of the cable's thermal state, for continuous attributes such as temperature and humidity, the generator can better capture the fine distribution of real data features, improving the accuracy of reflecting the actual operating state. The calculation formula is as follows:

[0097] N c (z,α c )=z+α c ·∈ c ,

[0098]

[0099] In the formula, N c (z,α c ) is the adaptive noise injection function, representing additional random perturbations on the basis of the original noise; z is the input noise; α c is the noise intensity coefficient, representing the dynamic weight added to the original noise during the generation process; ∈ c is the random noise of the normal distribution, and ~ means following a specific distribution, is the normal distribution with a mean of 0 and a variance of the identity matrix, I is the identity matrix, represents the normal distribution;

[0100] The noise intensity coefficient α c dynamically adjusts the noise intensity according to the discriminator's output on the synthetic cable data. The discriminator can improve the quality control of the synthetic cable data through dynamically adjusted thresholds and sensitivity parameters. For example, for attributes with a high dynamic range such as current and voltage, it can ensure that the synthetic cable data has the distribution characteristics of real data and avoid generation distortion caused by excessive noise. The calculation formula is as follows:

[0101]

[0102] In the formula, exp() is the natural exponential function; β c is the adjustment sensitivity parameter, representing the amplification effect on the discriminator's recognition level, set to 0.1; τ ce is the target threshold, representing the probability level at which the expected synthetic cable data can be recognized as real by the discriminator, set to 0.5.

[0103] As Figure 1 shown, through the comparison of the convergence of the loss function, it shows the downward trend of the loss function of the adaptive noise injection generative adversarial network and the traditional generative adversarial network during the training process, reflecting the advantages of the adaptive noise injection technology with faster convergence speed and smaller fluctuations.

[0104] S205. After obtaining the output of the generator, update the generator and discriminator according to the dynamic learning rate to achieve balanced and effective adversarial training. In the optimization of the generator and discriminator, by dynamically adjusting the weight update strategy, balance the diversity generated by the generator and the discrimination ability of the discriminator. For example, for attributes with unbalanced characteristics (such as high-voltage and low-voltage cable types), the generator can better learn the distribution of rare high-voltage cable states, thus avoiding the singularity or class imbalance of synthetic cable data. The calculation formula is as follows:

[0105]

[0106] In the formula, is the weight of the generator, and ← is the parameter update symbol; η c is the dynamic learning rate of the generative adversarial network, which characterizes the control of the parameter update step size at different training times; is the gradient of the generator loss function with respect to its weight, is the weight of the discriminator, is the gradient of the discriminator loss function with respect to its weight;

[0107] S206. Repeat iterations of S202 - S205 until the preset stop iteration condition is met, which indicates that the data augmentation model training is completed. The preset stop iteration condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times. Suppose the actual cable data is 800 pieces, and the data augmentation model augments and generates 200 pieces of synthetic cable data, then the training data contains 1000 samples.

[0108] As Figure 2 shown, to verify the effectiveness, through the synthetic cable data distribution comparison chart, compare the distributions of real data, synthetic cable data generated by the adaptive noise injection generative adversarial network, and synthetic cable data generated by the traditional generative adversarial network, showing that the data distribution generated by the adaptive noise injection technology is closer to the real data.

[0109] S3. Input the training data into the classification model, and train the classification model using the random forest algorithm based on relative entropy to obtain the trained classification model;

[0110] The random forest algorithm based on relative entropy refers to adapting to class imbalance by using a dynamically adjusted relative entropy loss function, and at the same time using regularization to reduce the risk of overfitting. When splitting nodes, maximize the information gain to optimize the splitting method, and perform majority voting through the fusion of multiple decision tree models to improve classification stability.

[0111] S301. Initialize the number and depth of decision trees required for the random forest, and set the number of decision trees in the random forest to D u is the dimension of the cable dataset input into the random forest, representing the total number of input features. The number of decision trees is set to 200;

[0112] As Figure 3 shown, to illustrate the numerical rationality of the decision tree setting, it is verified through the model performance under different numbers of decision trees. At the same time, the random forest based on relative entropy has a faster performance improvement when the number of trees increases, reflecting the advantages of the optimized splitting point and dynamic adjustment of the loss function of the random forest based on relative entropy as a classification algorithm.

[0113] Randomly select a subset of cable data for each decision tree in the random forest to improve the diversity of the cable state training data. Let be the subset of cable data used by the i-th decision tree, representing the training samples of the i-th decision tree of the model; D u is the complete cable dataset, representing all available cable state training data.

[0114] By randomly selecting subsets of cable data for each decision tree, the diversity of the training data is increased. For example, for processing data with different spatial features (such as cable position attributes), regarding cables with different burial depths (the z-axis attribute of the cable position), the model can learn different state patterns of deeply buried and shallowly buried cables through subset sampling.

[0115] S302. Screen features by using Gaussian curvature. By calculating the local curvature of cable data features in a high-dimensional space, select features that are highly sensitive to the classification boundary to help the model better capture complex non-linear relationships in the cable data. Let the feature set selected by Gaussian curvature be The selection method is to calculate the Gaussian curvature of the feature vector. When it is greater than the preset threshold, it is selected. The selected features represent more discriminative features in the high-dimensional cable data;

[0116] Use the Gaussian curvature feature screening technology to select key features that are highly sensitive to the classification boundary (such as the combination of humidity and temperature), and improve the modeling ability for complex attribute relationships. For example, for the combined relationship between the cable temperature attribute and the cable humidity attribute of the data, it may be identified as an important feature through Gaussian curvature screening, revealing the unique impact of the humid and hot environment on the cable state.

[0117] As Figure 4 shown, to verify the effectiveness of screening features by using Gaussian curvature, by comparing the impact of different feature screening methods on performance, it shows that the features screened by Gaussian curvature significantly improve the model performance, highlighting the superiority of Gaussian curvature screening in complex non-linear relationships.

[0118] S303. During the construction of the decision tree, a relative entropy loss function dynamically adjusted based on the distribution characteristics of cable data is adopted to adapt to the inherent distribution characteristics of different categories of cable data. Especially in the case of class imbalance, the classifier can more accurately distinguish different categories during the learning process. For example, the occurrence frequency of low-voltage cables may be much higher than that of high-voltage cables. Through the dynamic adjustment of relative entropy, the model pays more attention to the status classification of high-voltage cables, improving the classification accuracy. The calculation formula of the relative entropy loss function is as follows:

[0119]

[0120] In the formula, is the relative entropy loss function, representing the overall classification error; K ue is the number of categories; is the probability of the k-th category predicted by the model, representing the output confidence of the model for the k-th category; log represents the logarithmic function, with the default base of 10; is the true distribution probability of the k-th category, representing the true category distribution of cable data; k is the category index;

[0121] To balance the complexity of the model structure and the risk of overfitting, a dynamically adjusted loss function is used to optimize the growth of the decision tree. The calculation formula is as follows:

[0122]

[0123] In the formula, is the adjusted loss function, representing the evaluation value considering both the classification error and the regularization term; α u is the first adjustment coefficient, is the relative entropy loss function; β u is the second adjustment coefficient, set to 0.1; R ur (u) is the regularization term of the random forest.

[0124] To realize the dynamicization of the loss function, the calculation formula of the first adjustment coefficient during the training process is as follows:

[0125]

[0126] In the formula, k u is the parameter for adjusting the slope, representing the sensitivity of the first adjustment coefficient to the change in imbalance degree, set to 0.3; is the entropy value of the cable dataset; δ u is the entropy value threshold, set to 0.2.

[0127] The model complexity is constrained by the regularization term to avoid overfitting to noisy data. For example, for the extreme value noise that may exist in Ia (cable current), the regularization mechanism will suppress its adverse effects on the model weights, thereby improving the generalization performance, the convergence speed and stability of the classification model. The calculation formula for the regularization term of the random forest is as follows:

[0128]

[0129] In the formula, R ur (u) is the regularization term of the random forest, and m u is the total number of sample features input into the random forest. is the regularization weight of the j-th feature, which characterizes the influence degree of the feature on the regularization term and is obtained by calculating the L2 norm of each feature vector in the cable dataset; is the weight of the j-th feature, which characterizes the importance of the feature in the model.

[0130] S304. During the growth process of the decision tree, the optimal partitioning method is determined by training each decision tree based on the adjusted loss function and selecting the optimal splitting point that maximizes the inter-class separation degree at node splitting, so as to fully exploit the information gain under the relative entropy criterion. The method for optimizing node splitting is expressed as:

[0131]

[0132] In the formula, is the optimal splitting parameter, which characterizes the splitting threshold or splitting dimension; is the node splitting optimization function based on the adjusted loss function, which characterizes the evaluation value of different splitting schemes; θ ury is the candidate splitting parameter set, which characterizes all feasible splitting methods; represents the candidate splitting parameter set in the case of obtaining the minimum value of the function;

[0133] To measure the classification error of the left and right sub-decision trees after node splitting, the calculation formula for the node splitting optimization function based on the adjusted loss function is:

[0134]

[0135] In the formula, Left(θ ury ) is the data point i in the left sub-decision tree of the decision tree under the candidate splitting parameter set w is the data point index; is the class probability of the i-th w data point predicted by the decision tree in the random forest, which characterizes the prediction confidence of the data point belonging to a certain class; is the i-th wThe true class probability of a data point, representing the true classification of the data point; Right(θ ury ) are the data points in the right sub-decision tree of the decision tree under the candidate split parameter set.

[0136] S305. By synthesizing several independently trained decision trees into a random forest and using the majority voting method for global prediction, the aggregation output result is improved in terms of classification accuracy and stability when multiple decision tree models are fused. The calculation formula for synthesizing the global model is:

[0137]

[0138] In the formula, is the final prediction result of the random forest, representing the final classification of the model for the input cable data; the mode() function is the majority voting function, representing selecting the class with the most occurrences from the outputs of each decision tree; Tu i is the i-th decision tree, representing the classification function of each decision tree; is the input feature, representing the cable data to be classified; i is the decision tree index, is the number of decision trees in the random forest;

[0139] S306. By analyzing the contribution of each feature to the output entropy of the model to update its weight, the feature sensitivity and classification accuracy are improved in the later stage of the random forest model training. The calculation formula for entropy weight adjustment of dynamic feature points is:

[0140]

[0141] In the formula, is the weight of the j-th feature, representing the influence of the feature on the model prediction; j is the feature index, and update() is the weight update function; is the contribution of the j-th feature to the change of the loss function, representing the importance of the feature in model tuning;

[0142] The weight update function represents the specific implementation method for modifying the feature weight. It dynamically updates the feature weight by combining Ricci curvature and the contribution of the feature to entropy, enabling the model to pay more attention to key features in the later stage of training. For example, for important historical maintenance information of cables, the model can gradually increase its weight during training to improve the recognition ability of the cable status with maintenance records. The calculation formula for the weight update function is:

[0143]

[0144] In the formula, update() is the weight update function; is the contribution of the j-th feature to the change of the loss function, representing the importance of the feature in model tuning; is the weight of the j-th feature, representing the impact of this feature on the model prediction; e is the natural constant; β uac is the Ricci curvature adjustment coefficient, representing the influence intensity of Ricci curvature on weight update; is the Ricci curvature of the j-th feature, representing the degree of bending of the data manifold of this feature under a specific category; is the learning rate of the weight update at the t-th iteration, is the partial derivative symbol, is the adjusted loss function, is the change amount of the adjusted loss function.

[0145] To adapt to the real-time changes in the model performance, the learning rate is dynamically adjusted. The learning rate is dynamically adjusted through real-time performance metrics to ensure that the model can quickly adapt to performance changes. For example, in the initial stage of model training, when the classification performance improves rapidly, the learning rate is relatively high; in the later stage, when the performance approaches convergence, the learning rate gradually decreases to improve the convergence stability of the model. The calculation formula is:

[0146]

[0147] In the formula, is the learning rate at the (t + 1)-th iteration, is the learning rate of the weight update at the t-th iteration, where t is the iteration number; δ ucg is the learning rate decay factor, representing the intensity of adjusting the learning rate according to the model performance, and is set to 0.3; perf(t) is the performance metric of the random forest model after the t-th iteration, and specifically uses the training accuracy metric.

[0148] S307. Repeat iterations S302 - S306 until the preset iteration stop condition is met, which means the training of the classification model is completed. The preset iteration stop condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0149] S4. Collect new actual cable data, input the new actual cable data into the trained classification model, and obtain the operating state corresponding to the new actual cable data. The operating state includes normal, warning, and fault. When the operating state is warning, an inspection is carried out, and when the operating state is fault, maintenance is carried out.

[0150] The urban underground cable operation and maintenance inspection system based on virtual reality provided by the present invention applies the above-mentioned urban underground cable operation and maintenance inspection method based on virtual reality, including:

[0151] A data acquisition unit, used to collect actual cable data;

[0152] A data storage and expansion unit for storing data and generating synthetic cable data based on actual cable data through a data expansion model;

[0153] A data processing unit for labeling the corresponding operating status of the input cable data through a classification model;

[0154] A control unit for receiving the corresponding operating status of the cable data, outputting an instruction according to the operating status, and sending the instruction to an execution unit;

[0155] An execution unit for receiving and executing instructions, including staff inspection and staff maintenance.

Claims

1. A method for urban underground cable operation, maintenance and inspection based on virtual reality, characterized in that: include: S1. Obtain the actual cable data of underground cables and manually mark the operating status of each of the actual cable data; The operating status includes normal, warning and fault; S2. Input the annotated actual cable data into the data expansion model, train the data expansion model using a generative adversarial network algorithm based on adaptive noise injection to obtain a trained data expansion model, generate synthetic cable data using the trained data expansion model, and use the actual cable data and the synthetic cable data as training data; S3. Inputting the training data into the classification model, training the classification model using a random forest algorithm based on relative entropy to obtain a trained classification model; S4. Collect new actual cable data, input the new actual cable data into the trained classification model, and obtain the operating status corresponding to the new actual cable data, the operating status including normal, warning and fault. When the operating status is a warning, an inspection is performed, and when the operating status is a fault, maintenance is performed.

2. The urban underground cable operation, maintenance and inspection method based on virtual reality according to claim 1 is characterized in that: The step of training the data augmentation model using a generative adversarial network algorithm based on adaptive noise injection in S2 includes: S201. Initialize the parameters of the generator and discriminator of the generative adversarial network; S202. Design a loss function based on the dynamic changes in cable data distribution during training, and the calculation formula is: In the formula, is the loss function of the generator; D c () is the discriminator function; G c () is the generator function; z is the input noise, specifically a random vector sampled from the latent space; is the loss function of the discriminator; x ce is the actual cable data sample; S203. Dynamically adjust the weight parameters of the generator and the discriminator according to the discriminator feedback, and use the mechanism of topological structure difference and adaptive noise injection to control the learning rate and noise mode, which is expressed as: In the formula, ΔT c is the topological difference; nc is the number of samples input to the generative adversarial network in the current batch; Re() is the ReLU activation function used to extract the structural information of the sample, G c () is the generator function, For the i z A noise vector, For the i z Actual cable data samples; i z is the cable data sample index and is the same as the noise vector index; S204. Inject adaptive noise into the generator, and determine the noise intensity through the feedback of the discriminator to the synthetic cable data. The calculation formula is: N c (z,α c )=z+α c ·∈ c , Where N c (z,α c ) is the adaptive noise injection function; z is the input noise; α c is the noise intensity coefficient; ∈ c is a normally distributed random noise, is a normal distribution with mean 0 and variance as the identity matrix, I is the identity matrix, represents normal distribution; S205. After obtaining the output of the generator, the generator and the discriminator are updated according to the dynamic learning rate. In the optimization of the generator and the discriminator, the weight update strategy is dynamically adjusted to balance the generation diversity of the generator and the discrimination ability of the discriminator. The calculation formula is: In the formula, is the weight of the generator, ← is the parameter update symbol; η c Dynamic learning rate for generative adversarial networks; is the gradient of the generator loss function with respect to its weights, is the weight of the discriminator, is the gradient of the discriminator loss function with respect to its weights; S206. Repeat iterations S202 to S205 until a preset stop iteration condition is met.

3. The urban underground cable operation, maintenance and inspection method based on virtual reality according to claim 2 is characterized in that: S202 D c (G c The calculation formula of (z) is: Where D c () is the discriminator function; G c () is the generator function; z is the input noise; Sig() is the Sigmoid activation function; is the weight of the discriminator, is the bias of the discriminator.

4. The urban underground cable operation, maintenance and inspection method based on virtual reality according to claim 2 is characterized in that: The noise intensity coefficient α in S204 c The calculation formula is: In the formula, β c To adjust the sensitivity parameter; τ ce is the target threshold.

5. The method for urban underground cable operation, maintenance and inspection based on virtual reality according to claim 1 is characterized in that: The step of using the random forest algorithm based on relative entropy to train the classification model in S3 includes: S301. Initialize the number and depth of decision trees required for the random forest; S302. Screening features by using Gaussian curvature; S303. In the process of building the decision tree, a relative entropy loss function dynamically adjusted based on the cable data distribution characteristics is used. The calculation formula of the relative entropy loss function is: In the formula, is the relative entropy loss function; K ue is the number of categories; is the probability of the kth category predicted by the model; is the true k-th class distribution probability, k is the class index; S304. Each decision tree is trained based on the adjusted loss function, and the optimal split point that maximizes the inter-class separation is selected when the node is split. The optimal division method is determined during the growth of the decision tree. The node splitting optimization method is expressed as: In the formula, is the optimal splitting parameter; is the node splitting optimization function based on the adjusted loss function; θ ury is the candidate split parameter set; S305. By combining several independently trained decision trees into a random forest and using majority voting to perform global prediction, the calculation formula of the synthesized global model is: In the formula, is the final prediction result of random forest; mode() function is the majority voting function; Tu i is the i-th decision tree; is the input feature of the decision tree in the random forest; i is the decision tree index, is the number of decision trees in the random forest; S306. Update the weight of each feature by analyzing its contribution to the model output entropy. The calculation formula for the entropy weight adjustment of the dynamic feature point is: In the formula, is the weight of the jth feature, j is the feature index; update() is the weight update function; is the contribution of the jth feature to the change in the loss function; S307. Repeat iterations S302 to S306 until a preset stop iteration condition is met.

6. The method for urban underground cable operation, maintenance and inspection based on virtual reality according to claim 5 is characterized in that: S304 The calculation formula is: In the formula, Left(θ urt ) is the data point in the left child decision tree under the candidate split parameter set; i w Index of data points; is the i-th prediction of the decision tree in the random forest w The class probability of each data point; For the i w The true category probability of data points; Right(θ ury ) is the data point in the right child decision tree under the candidate split parameter set.

7. The method for urban underground cable operation, maintenance and inspection based on virtual reality according to claim 5 is characterized in that: S306 The calculation formula is: Where update() is the weight update function; is the contribution of the jth feature to the change in the loss function; is the weight of the jth feature; e is a natural constant; β uac is the Ricci curvature adjustment coefficient; is the Ricci curvature of the jth feature; is the learning rate of the weight update for the tth iteration, t is the number of iterations, is the symbol of partial derivative, is the adjusted loss function, is the change in the adjusted loss function.

8. The urban underground cable operation and maintenance system based on virtual reality is characterized by: The method for urban underground cable operation, maintenance and inspection based on virtual reality according to any one of claims 1 to 7 comprises: A data acquisition unit, used to collect actual cable data; A data storage and expansion unit, used for storing data and generating synthetic cable data based on actual cable data through a data expansion model; The data processing unit is used to mark the corresponding operating status of the input cable data through the classification model.

Images