Artificial Intelligence-Based Method for Monitoring the Heating State and Predicting Faults of Power Cables

Abstract

The present invention discloses a method for monitoring the heating state and predicting faults of power cables based on artificial intelligence, which relates to the field of cable technology. It collects power cable data in the operating state and annotates the collected power cable data. The training framework of the state monitoring and fault prediction model is a fusion framework of meta-learning and transfer learning; it uses a neural network based on dynamic feature influence; and it uses an autoencoder based on feature path decoupling. The present invention optimizes feature independence through a feature path decoupling mechanism, evaluates and optimizes the independence between power cable data features through a regularization loss function. In addition, it calculates the linear dependence between power cable data features using covariance and strengthens the expression ability of independent power cable data features through a penalty mechanism, improving the model's processing ability for power cable data. Even in the case of missing power cable data features, the model can still maintain an accurate expression of the power cable data structure.

Description

Technical Field

[0001] The present invention relates to the technical field of cables, and particularly to a method for monitoring the heating state and predicting faults of power cables based on artificial intelligence. Background Art

[0002] In the tasks of monitoring the heating state and predicting faults of power cables, the real-time processing of high-dimensional sensor data and the effective extraction of features are the keys to achieving accurate monitoring and prediction. A large amount of sensing data generated during the operation of power cables is multi-dimensional and dynamic, including complex information such as temperature, load current, and environmental humidity. There is often a high degree of correlation and redundancy among these data, which poses challenges to fault identification and state assessment. At the same time, in practical applications, sensors may cause partial data loss due to environmental or equipment problems, which requires the monitoring and prediction models to have high robustness and be able to accurately capture key information under incomplete data conditions. In addition, the operating environment of the power system is complex, and the data feature distribution will change with time and working conditions. Existing technologies have significant deficiencies in coping with real-time adjustment of feature importance and optimizing fault prediction accuracy in a dynamic environment.

[0003] In the prior art, a Chinese invention patent with the application number disclosed a Chinese invention patent with the publication number CN118735141A, which proposed an optimization method for cable operation and maintenance strategies based on machine learning, belonging to the technical field of cable operation and maintenance. Specifically, it includes: setting a target area and an inspection cycle and obtaining the inspection data of all cables; dividing the inspection cycle into n operation and inspection cycles, selecting any cable and obtaining the number of failures of the cable in any operation and inspection cycle, and calculating the reference value J' of the cable; selecting any cable and calculating the adjacent failure time interval, determining the number of operation and inspection cycles in which the number of failures of any cable is greater than or equal to the reference value and calibrating it as the reference quantity; calculating the operation and inspection score of any cable; correcting the operation and inspection score of the cable according to the detection data, and taking the cable with the operation and inspection score greater than or equal to the preset risk threshold as the target cable that needs to be inspected. A Chinese invention patent with the publication number CN118708924A proposed a method, device and system for testing the electrical performance of robot cables. The method includes: based on each time period and a preset fixed value, confirming the data change stability coefficient of each sampling time point; based on the data change stability coefficient of each sampling time point, the total number of sampling time points in each electrical performance factor, the total number of sampling time points and the preset response sensitivity threshold of each electrical performance factor, obtaining the influence degree coefficient; based on the monitoring data, time difference and influence degree coefficient of each sampling time point, obtaining the efficiency influence trend coefficient of signal transmission for each electrical performance factor; based on the efficiency influence trend coefficient of signal transmission for each electrical performance factor, testing the electrical performance of the robot cable to obtain the test result. A Chinese invention patent with the publication number CN118228101A proposed a method and device for cable anomaly detection, which includes: obtaining historical anomaly monitoring data with cable anomalies and historical normal monitoring data with normal cables, comparing and analyzing them, determining anomaly parameters according to the comparison and analysis results, screening out the historical anomaly parameter data corresponding to the anomaly parameters from the historical anomaly monitoring data, analyzing it, extracting feature data from the historical anomaly parameter data, and preprocessing it. Based on the preprocessed multiple groups of feature data and a preset machine algorithm, a cable anomaly detection initial model is constructed, and the cable anomaly detection initial model is trained and tested to obtain a cable anomaly detection model; judging whether the cable is abnormal through the cable anomaly detection model.

[0004] The above technical solutions have the following problems: 1. In the tasks of power cable heating state monitoring and fault prediction, the existing feature extraction technologies fail to effectively solve the problems of redundancy and coupling among features, resulting in the difficulty for the power cable heating state monitoring and fault prediction model to maintain the integrity of the power cable data structure in the presence of missing features or redundant features. 2. In the tasks of power cable heating state monitoring and fault prediction, the existing models lack the ability to adjust feature weights in real time, leading to unstable performance of the power cable heating state monitoring and fault prediction model when the power cable data distribution changes. 3. In the tasks of power cable heating state monitoring and fault prediction, the current autoencoder methods usually only focus on the optimization of reconstruction error, ignoring the maximization of information gain during the power cable data compression process, which is prone to causing the loss of important information in the power cable data representation. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the above-mentioned disadvantages of the prior art and provide a method for monitoring the heating state and predicting faults of power cables based on artificial intelligence.

[0006] The technical solution adopted to solve the above technical problem is: A method for monitoring the heating state and predicting faults of power cables based on artificial intelligence, including the following steps:

[0007] S1, Collect power cable data in the operating state, and label the collected power cable data. The labeled categories include: normal, warning, and fault;

[0008] S2, The training framework of the state monitoring and fault prediction model is a fusion framework of meta-learning and transfer learning;

[0009] S3, Adopt a neural network based on feature dynamic influence, and the hidden layer of the neural network performs dynamic calculation using a multi-scale feature fusion and high-order feature interaction strategy;

[0010] S4, Adopt an autoencoder based on feature path decoupling, learn the compressed representation of power cable data through the autoencoder, and optimize the independence between features using a feature path decoupling mechanism. The feature path decoupling mechanism strengthens the independence between features through a regularization loss function, ensuring that even if some features are missing, the model can still capture the main structure of the power cable data.

[0011] Further, the power cable data in S1 includes the real-time current value Ra of the cable, the temperature value Da of the cable, the ambient temperature Ca, the ambient humidity Fa, the voltage value Ga of the cable, the wind speed Va, the solar radiation intensity Ma, the historical load record Ba of the cable, the aging degree Na of the insulating material, and the connection state Ka of the cable joint.

[0012] Further, the training process of the fusion framework in S2 includes a pre-training stage and a fine-tuning stage;

[0013] In the pre-training stage, power cable data samples are input into the feature extraction modules of transfer learning and meta-learning. Among them, power cable data samples with batches as the training unit are input into the transfer learning feature extraction module, and power cable data samples with tasks as the training unit are input into the meta-learning feature extraction module;

[0014] In the fine-tuning stage, the training set of the target domain power cable data is divided with batches and tasks as the basic units. The training set of the target domain power cable data is a set of power cable data with labels in the target domain power cable data. The pre-trained state monitoring and fault prediction models of meta-learning and transfer learning are respectively used to classify the target domain power cable data, and the scores of the classification results of the two state monitoring and fault prediction models are fused as the final classification result of the fusion framework.

[0015] Further, the meta-learning adopts the model weak association meta-learning strategy. The learning goal of the model weak association meta-learning strategy is to quickly learn and adapt to the optimal parameters required for new tasks through the parameter optimization algorithm, including two-layer loop modes of inner loop and outer loop;

[0016] The inner loop uses a base learner to extract features of a specific task, and the base learner is a neural network model; the outer loop uses the meta-learning strategy to update the initialization parameters of the base learner through gradient descent.

[0017] Further, the training process of the neural network algorithm based on feature dynamic influence in S3 includes the following steps:

[0018] S301. Initialize the structure and parameters of the neural network to adapt to the distribution characteristics of different power cable data sets, expressed as:

[0019] ,

[0020] In the formula, is the weight parameter of the th layer of the neural network, is initialized with a normal distribution with a mean of and a variance of , is the mean of the neural network parameter initialization, is the variance of the neural network parameter initialization, is an orthogonal perturbation matrix aligned with the input dimension to ensure that the initial parameters of the neural network are highly sensitive to the input features;

[0021] To optimize the sparsity of the neural network, the pruning threshold of the neural network is estimated according to the complexity of the input power cable data during initialization, which is expressed as:

[0022] ,

[0023] wherein, is the pruning threshold of the neural network; is the pruning adjustment coefficient of the neural network; is the feature variance of the input power cable data; is the variance function; are the features of the power cable data input to the neural network;

[0024] S302. For the power cable data input to the neural network, a method of suppressing feature redundancy is adopted. By dynamically adjusting the feature correlation weight parameters, the influence of redundant features is weakened. The calculation method of the correlation weight parameters is expressed as:

[0025] ,

[0026] wherein, is the correlation coefficient between the th feature and the th feature; is the covariance of the features and ; is the th feature input to the neural network; is the th feature input to the neural network; and are the variances of the features and respectively; is the index of the first neural network input feature; is the index of the second neural network input feature;

[0027] According to the feature correlation weight parameters, the weights of the input features are dynamically adjusted, which is expressed as:

[0028] ,

[0029] wherein, is the th feature input to the neural network after adjustment;

[0030] S303. The hidden layer of the neural network adopts a multi-scale feature fusion and high-order feature interaction strategy to realize the progressive abstraction of features from low order to high order and enhance the interactive representation ability between features. The multi-scale feature fusion method is expressed as:

[0031] ,

[0032] Wherein, is the fused feature input to the th layer of the neural network, is the index of the number of layers of the first neural network, is the number of multi - scales, s is the index of the scale, is the fusion weight of the sth scale, which is a training parameter and obtained through dynamic learning, is the non - linear activation function of the sth scale, is the weight parameter of the sth scale of the th layer of the neural network, is the fused feature input to the (l - 1)th layer of the neural network, is the bias parameter of the sth scale of the

[0033] The non - linear activation function adopted by the neural network can enhance the expression ability of the non - linear region, and ensure the smooth propagation of the gradient in the positive and negative regions through dynamic adjustment. The calculation method is expressed as:

[0034] ,

[0035] Wherein, is the non - linear activation function of the th layer of the neural network, is the input of the non - linear activation function, is the linear rectifier unit activation function, is the first dynamic adjustment factor, is the hyperbolic tangent activation function, is the second dynamic adjustment factor;

[0036] S304. Dynamically adjust the gradient update amplitude by calculating the importance coefficient of the features of each layer of the neural network. The calculation method is expressed as:

[0037] ,

[0038] Wherein, is the weight update increment of the th layer of the neural network, is the learning rate of the neural network, is the entropy value of the gradient of the th layer of the neural network, is the partial derivative symbol, is the loss function of the neural network, specifically the cross-entropy loss, which is calculated by using the preset Softmax function to calculate the class probabilities of the output features of the neural network, and then calculated by the cross-entropy loss function. is the weight parameter of the th layer of the neural network. is the importance factor of the features of the th layer of the neural network.

[0039] The feature importance factor dynamically measures the importance of features to the global output, ensuring more reasonable gradient allocation. The calculation method is expressed as:

[0040] ,

[0041] In the formula, is the variance function. is the fused feature input to the th layer of the neural network. is the total number of layers of the neural network. is the index of the number of layers of the second neural network.

[0042] The learning rate of the neural network is adjusted based on the rate of change of the gradient. The calculation method is expressed as:

[0043] ,

[0044] In the formula, is the base learning rate of the th layer of the neural network. is the learning rate adjustment coefficient of the th layer of the neural network, indicating the impact of the gradient change on the learning rate. is norm. is the updated gradient of the loss function in this iteration and the previous iteration.

[0045] During the training process of each layer, by calculating the entropy value of the gradient of the neural network, the learning speed of each layer is ensured to be coordinated. The calculation method is expressed as:

[0046] ,

[0047] In the formula, is the number of neurons in the th layer of the neural network. is the gradient of the th layer and the th neuron of the neural network. is the index of the neurons in the th layer of the neural network. is the logarithmic function.

[0048] Update the weight parameters of the neural network, and the update method is expressed as:

[0049] ,

[0050] In the formula, is the parameter update operation, is the regularization factor of the th layer of the neural network;

[0051] To avoid overfitting, the regularization factor of the neural network is related to the entropy value of the gradient. The regularization factor will decrease as the gradient increases, thereby reducing the sensitivity of the model to overfitting and effectively adjusting the learning progress of each layer in the network. The calculation method is expressed as:

[0052] ,

[0053] In the formula, is the hyperparameter that adjusts the influence degree of the regularization factor of the th layer of the neural network, is the weight update gradient of the neural network, is norm;

[0054] S305. Repeat the above steps iteratively until the preset stop iteration condition is satisfied. After the neural network model is trained, the output feature of the neural network is the feature vector after feature extraction, and this feature vector passes through the preset Softmax function to obtain the prediction score of the power cable data.

[0055] Furthermore, the training process of the autoencoder algorithm based on feature path decoupling in S4 includes the following steps:

[0056] S401. Initialize the encoder and decoder of the autoencoder. The encoder is responsible for mapping the high-dimensional input power cable data to a low-dimensional space, and the decoder attempts to reconstruct the original power cable data from the low-dimensional representation. The initialization method is expressed as:

[0057] ,

[0058] In the formula, and are the weight matrices of the th layers of the encoder and decoder respectively, and are the bias vectors of the th layers of the encoder and decoder respectively, is the dimension of the th layer of the encoder, is the dimension of the th layer of the encoder, and generate normally distributed random numbers and zero vectors respectively, is the layer index of the encoder and decoder;

[0059] S402. During the encoding process, decouple the dependency paths of the input features through a regularization method. The calculation method of the loss function of the autoencoder is expressed as:

[0060] ,

[0061] In the formula, is the total loss function of the autoencoder, including the reconstruction error and the regularization term. x represents the independent variable of the function, is the encoder function, is the decoder function, is the regularization weight, is the decoupling regularization function;

[0062] The role of the decoupling regularization function is to evaluate the independence of the features. The calculation method is expressed as:

[0063] ,

[0064] In the formula, z is the input of the decoupling regularization function, is the index of the input feature of the first decoupling function, is the index of the input feature of the second decoupling function, is the weight of the th feature in the t-th iteration, represents the feature and covariance, measuring the linear dependence between features, is the th feature of the input of the decoupling function, is the th feature of the input of the decoupling function, is the variance weight factor of the decoupling function, is the index of the input feature of the third decoupling function, is the variance of the feature , enabling the decoupling function to improve the information retention ability by suppressing features with small variances, is the th feature of the input of the decoupling function;

[0065] S403. Through the dynamic feature flow embedding mechanism, the model dynamically adjusts the weights of each feature in the hidden layer according to the reconstruction error and feature importance to adapt to the characteristics of power cable data and task requirements. The adjustment method is expressed as:

[0066] ,

[0067] In the formula, is the weight of the th feature in the (t + 1)-th iteration, is the weight of the th feature in the t-th iteration, is the learning rate of the dynamic feature stream embedding, is the gradient of the loss function with respect to the feature weight, is the adjustment parameter of the dynamic feature stream embedding, is the th information gain of the feature;

[0068] S404. The training process adopts a loop process of feedforward, feedback, and parameter update, and minimizes the loss function through the gradient descent method. The parameter update method of the encoder is expressed as:

[0069] ,

[0070] In the formula, is the parameter update operation, and are the gradients of the loss function of the autoencoder with respect to the weight and bias of the th layer respectively;

[0071] S405. Repeat the above iteration process, measure the model performance through comprehensive evaluation indicators, and optimize the reconstruction error and feature information content at the same time. The calculation method of the evaluation indicator is expressed as:

[0072] ,

[0073] In the formula, is the training evaluation indicator of the autoencoder, is the first adjustment weight of the autoencoder, is the second adjustment weight of the autoencoder, is the th information gain of the feature;

[0074] Repeat the iteration of S402 - S405 until the evaluation indicator is less than the preset threshold, then stop the iteration, which means the model training is completed. After the autoencoder model training is completed, the output features of the encoder are the low-dimensional feature vectors after feature extraction. The low-dimensional feature vectors pass through the preset Softmax function to obtain the prediction scores of the power cable data.

[0075] The beneficial effects of the present invention are as follows: (1) In the task of power cable heating state monitoring and fault prediction, the present invention optimizes feature independence through a feature path decoupling mechanism, evaluates and optimizes the independence between power cable data features through a regularization loss function. In addition, covariance is used to calculate the linear dependence between power cable data features, and the expression ability of independent power cable data features is strengthened through a penalty mechanism, improving the model's processing ability for power cable data. Even in the case of missing power cable data features, the model can still accurately represent the power cable data structure.

[0076] (2) In the task of power cable heating state monitoring and fault prediction, the present invention uses a dynamic feature flow embedding mechanism to adapt to the characteristics of power cable data, dynamically adjusts the feature weights of the hidden layer according to the importance of power cable data features, and optimizes the feature representation in real time through gradient update. In addition, combining the reconstruction error and information gain, the dynamic feature flow embedding enables the power cable heating state monitoring and fault prediction model to adaptively adjust the feature weight distribution under different data distributions and task requirements.

[0077] (3) In the task of power cable heating state monitoring and fault prediction, the present invention proposes a comprehensive evaluation index combining the reconstruction error and feature information gain, and guides the training of the power cable heating state monitoring and fault prediction model in a multi-objective optimization manner. By adjusting the parameters, the reconstruction ability and information retention ability of the model are balanced, effectively avoiding the problem of power cable data feature loss during the information compression process. Brief Description of the Drawings

[0078] Figure 1 is the principle block diagram of the fusion framework in the present invention. Detailed Embodiments

[0079] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0080] The method for monitoring the heating state and predicting the fault of a power cable based on artificial intelligence includes the following steps:

[0081] S1, collect power cable data in the operating state, and label the collected power cable data. The labeled categories include: Normal: The cable is operating in a normal state without overloading or abnormal heating. Warning: The cable may have slight heating or is about to enter the overloaded state. Fault: The cable has serious heating or has already failed.

[0082] Power cable data includes the real-time current value Ra of the cable, which reflects the cable load condition; the temperature value Da of the cable, which affects the safe operation of the cable; the ambient temperature Ca, which affects the cable temperature; the ambient humidity Fa, which is an important factor affecting the cable performance; the voltage value Ga of the cable, which reflects the power transmission efficiency; the wind speed Va, which affects the cable heat dissipation; the solar radiation intensity Ma, which has a direct impact on the cable temperature; the historical load record Ba of the cable, which is used for trend analysis; the aging degree Na of the insulation material, which measures the health status of the cable; and the connection status Ka of the cable joint, which causes local heating of the cable.

[0083] S2, the training framework of the state monitoring and fault prediction model is a fusion framework of meta-learning and transfer learning, as Figure 1 shown.

[0084] The training process of the fusion framework includes a pre-training stage and a fine-tuning stage:

[0085] In the pre-training stage, the power cable data samples are input into the feature extraction modules of transfer learning and meta-learning. Among them, the power cable data samples with batches as the training unit are input into the transfer learning feature extraction module, and the power cable data samples with tasks as the training unit are input into the meta-learning feature extraction module.

[0086] In the fine-tuning stage, the training set of the target domain power cable data is divided with batches and tasks as the basic units. The training set of the target domain power cable data is the set of power cable data with labels in the target domain power cable data. The pre-trained state monitoring and fault prediction models of meta-learning and transfer learning are respectively used to classify the target domain power cable data, and the scores of the classification results of the two state monitoring and fault prediction models are fused as the final classification result of the fusion framework.

[0087] Meta-learning learns the meta-knowledge of different tasks through a multi-task learning paradigm and uses the meta-knowledge to quickly learn in new tasks, enabling it to achieve better learning results with a small number of power cable data samples. Transfer learning relies on the knowledge transfer from the source domain power cable data to the target domain power cable data to solve the problems of small samples and cross-domain learning of power cable data, and requires a large number of labeled power cable data to pre-train the state monitoring and fault prediction model.

[0088] Meta-learning adopts the model weakly associated meta-learning strategy. The learning goal of the model weakly associated meta-learning strategy is to quickly learn and adapt to the optimal parameters required for new tasks through a parameter optimization algorithm, including two-layer loop modes of inner loop and outer loop. The inner loop uses a base learner to extract features of a specific task, and the base learner is a neural network model. The outer loop uses the meta-learning strategy to update the initialization parameters of the base learner through gradient descent.

[0089] The meta - learning strategy maps each sample in the support set and each sample in the query set to a high - dimensional feature space as the input of the neural network model. During the training process, several categories are randomly selected for state monitoring and fault prediction model training. After completing all batches of task learning, the optimal initialization parameters are obtained. When encountering a new task, the support set of the new task is used to fine - tune the optimal initialization parameters to obtain the parameters most suitable for the new task.

[0090] S3. Adopt a neural network based on the dynamic influence of features. The hidden layer of the neural network adopts a multi - scale feature fusion and high - order feature interaction strategy to achieve the progressive abstraction of features from low - order to high - order. It can not only effectively capture the diversity of high - dimensional power cable data, but also improve the training efficiency and inference speed of the neural network, while enhancing the generalization performance of the model and the adaptability to complex power cable data.

[0091] The training process of the neural network algorithm based on the dynamic influence of features includes the following steps:

[0092] S301. Initialize the structure and parameters of the neural network to adapt to the distribution characteristics of different power cable data sets, expressed as:

[0093] ,

[0094] In the formula, is the weight parameter of the th layer of the neural network, is initialized with a normal distribution with a mean of and a variance of , is the mean of the neural network parameter initialization, is the variance of the neural network parameter initialization, is an orthogonal perturbation matrix aligned with the input dimension to ensure that the initial parameters of the neural network are highly sensitive to the input features.

[0095] To optimize the sparsity of the neural network, estimate the pruning threshold of the neural network according to the complexity of the input power cable data during initialization, expressed as:

[0096] ,

[0097] In the formula, is the pruning threshold of the neural network, is the pruning adjustment coefficient of the neural network, is set to 0.2, is the feature variance of the input power cable data, is the variance function, is the feature of the power cable data input to the neural network.

[0098] When the pruning threshold of the neural network is greater than the preset threshold, a preset proportion of neurons are randomly inactivated to reduce redundant connections in the initialization stage while retaining the core feature pathways.

[0099] S302. For the power cable data input into the neural network, a method of suppressing feature redundancy is adopted. By dynamically adjusting the feature correlation weight parameters, the influence of redundant features is weakened. The calculation method of the correlation weight parameters is expressed as:

[0100] ,

[0101] In the formula, is the correlation coefficient between the th and the th features, is the covariance of features and , is the th feature input into the neural network, is the th feature input into the neural network, and are the variances of features and respectively, is the index of the first neural network input feature, is the index of the second neural network input feature.

[0102] According to the feature correlation weight parameters, the weights of the input features are dynamically adjusted, expressed as:

[0103] ,

[0104] In the formula, is the th feature input into the neural network after adjustment.

[0105] S303. The hidden layer of the neural network adopts a multi-scale feature fusion and high-order feature interaction strategy to achieve the progressive abstraction of features from low order to high order and enhance the interactive representation ability between features. The multi-scale feature fusion method is expressed as:

[0106] ,

[0107] In the formula, is the fused feature of the th layer input into the neural network, is the index of the first neural network layer, is the number of multi-scales, s is the index of the scale, is the fusion weight of the s-th scale, which is a training parameter obtained through dynamic learning. is the non-linear activation function of the s-th scale. is the weight parameter of the s-th scale in the layer of the neural network. is the fused feature input to the layer of the neural network.

[0108] The non-linear activation function adopted by the neural network can enhance the expression ability in the non-linear region. At the same time, through dynamic adjustment, the smoothness of gradient propagation in the positive and negative regions is ensured. The calculation method is expressed as:

[0109] ,

[0110] In the formula, is the non-linear activation function of the layer of the neural network. is the input of the non-linear activation function. is the linear rectifier unit activation function. is the first dynamic adjustment factor. is the hyperbolic tangent activation function. is the second dynamic adjustment factor.

[0111] S304. Dynamically adjust the gradient update amplitude by calculating the importance coefficient of the features of each layer of the neural network. The calculation method is expressed as:

[0112] ,

[0113] In the formula, is the weight update increment of the layer of the neural network. is the learning rate of the neural network. is the entropy value of the gradient of the layer of the neural network. is the partial derivative symbol. is the loss function of the neural network, specifically the cross-entropy loss. It is calculated by using the preset Softmax function to calculate the class probabilities of the output features of the neural network, and then obtained through the cross-entropy loss function. is the weight parameter of the layer of the neural network. is the importance factor of the features of the layer of the neural network.

[0114] The importance factor of the features dynamically measures the importance of the features to the global output, ensuring more reasonable gradient allocation. The calculation method is expressed as:

[0115] ,

[0116] wherein, is the variance function, is the fused feature input to the -th layer of the neural network, is the total number of layers of the neural network, is the index of the number of layers of the second neural network.

[0117] The learning rate of the neural network is adjusted based on the rate of change of the gradient. As the gradient change increases, the learning rate is dynamically adjusted to avoid unstable learning caused by an overly large update step size. The calculation method is expressed as:

[0118] ,

[0119] wherein, is the base learning rate of the -th layer of the neural network, is the learning rate adjustment coefficient of the -th layer of the neural network, indicating the influence of the gradient change on the learning rate, is norm, is the updated gradient of the loss function in the current iteration and the previous iteration.

[0120] During the training process of each layer, by calculating the entropy value of the neural network gradient, it is ensured that the learning speed of each layer is coordinated. The calculation method is expressed as:

[0121] ,

[0122] wherein, is the number of neurons in the -th layer of the neural network, is the gradient of the -th layer and the -th neuron of the neural network, is the index of the neurons in the -th layer of the neural network, is the logarithmic function, and the base of the logarithmic function is default set to 10.

[0123] Update the weight parameters of the neural network. The update method is expressed as:

[0124] ,

[0125] wherein, is the parameter update operation, is the regularization factor of the -th layer of the neural network.

[0126] To avoid overfitting, the regularization factor of the neural network is related to the entropy value of the gradient. The regularization factor decreases as the gradient increases, thereby reducing the model's sensitivity to overfitting and effectively regulating the learning progress of each layer in the network. The calculation method is expressed as:

[0127] ,

[0128] In the formula, is the hyperparameter that affects the degree of adjustment of the regularization factor for the th layer of the neural network. The hyperparameters that affect the degree of adjustment of the regularization factor for each layer of the neural network are all set to 0.1. is the weight update gradient of the neural network. is the L2 norm.

[0129] S305. Repeat the iteration of S302 - S304 until the preset stop iteration condition is met. 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, which means the model training is completed. After the neural network model training is completed, the output features of the neural network are the feature vectors after feature extraction. This feature vector passes through the preset Softmax function to obtain the prediction scores of the power cable data.

[0130] In the fusion framework, different from the meta - learning based on the weak model association strategy, transfer learning can mine the subtle features under the shallow features of the power cable data. The advantage of transfer learning is that it can fully utilize the source - domain power cable data samples to pre - train the state monitoring and fault prediction model, and specifically uses an auto - encoder model for transfer.

[0131] S4. Use an auto - encoder based on feature - path decoupling to learn the compressed representation of the power cable data through the auto - encoder, and use the feature - path decoupling mechanism to optimize the independence between features. The feature - path decoupling mechanism strengthens the independence between features through the regularization loss function, ensuring that even if some features are missing, the model can still capture the main structure of the power cable data.

[0132] The training process of the auto - encoder algorithm based on feature - path decoupling includes the following steps:

[0133] S401. Initialize the encoder and decoder of the auto - encoder. The encoder is responsible for mapping the high - dimensional input power cable data to a low - dimensional space, and the decoder attempts to reconstruct the original power cable data from the low - dimensional representation. The initialization method is expressed as:

[0134] ,

[0135] In the formula, and are the weight matrices of the th layer of the encoder and decoder respectively, and are the bias vectors of the th layer of the encoder and decoder respectively, is the dimension of the th layer of the encoder, is the dimension of the th layer of the encoder, and generate normally distributed random numbers and zero vectors respectively, is the layer number index of the encoder and decoder.

[0136] S402. During the encoding process, the dependence path of the input features is decoupled through a regularization method, and the features that can independently represent the power cable data structure even when other feature information is removed are discovered and strengthened. The feature path decoupling strengthens the independence between features through a regularization loss function, ensuring that the model can still capture the main structure of the power cable data even when some features are missing. The calculation method of the loss function of the autoencoder is expressed as:

[0137] ,

[0138] In the formula, is the total loss function of the autoencoder, including the reconstruction error and the regularization term. x represents the independent variable of the function, which is only the power cable data input to the autoencoder here, is the encoder function, is the decoder function, is the regularization weight, is set to 0.45, is the decoupling regularization function.

[0139] The role of the decoupling regularization function is to evaluate the independence of features, and its calculation method is expressed as:

[0140] ,

[0141] In the formula, z is the input of the decoupling regularization function, is the index of the input feature of the first decoupling function, is the index of the input feature of the second decoupling function, is the weight of the th feature in the t-th iteration, represents the feature and covariance, measuring the linear dependence between features, is the th feature input to the decoupling function, is the a feature is the decoupling function variance weight factor is the index of the input feature of the third decoupling function is the feature variance, enabling the decoupling function to enhance the information retention ability by suppressing features with small variances is the th feature input to the decoupling function

[0142] S403. Through the dynamic feature flow embedding mechanism, the model dynamically adjusts the weights of each feature in the hidden layer according to the reconstruction error and feature importance to adapt to the characteristics of power cable data and task requirements. The adjustment method is expressed as:

[0143] ,

[0144] In the formula, is the weight of the th feature in the (t + 1)-th iteration is the weight of the th feature in the t-th iteration is the learning rate of the dynamic feature flow embedding is set to 0.01 is the gradient of the loss function with respect to the feature weight is the adjustment parameter of the dynamic feature flow embedding is set to 0.4 is the th feature information gain

[0145] S404. The training process adopts a loop process of feedforward, feedback, and parameter update, and minimizes the loss function through the gradient descent method. The parameter update method of the encoder is expressed as:

[0146] ,

[0147] In the formula, is the parameter update operation and are the gradients of the loss function of the autoencoder with respect to the weight and bias of the th layer respectively

[0148] S405. Repeat the above iterative process, measure the model performance through comprehensive evaluation indicators, and optimize both the reconstruction error and feature information content simultaneously. The calculation method of the evaluation indicator is expressed as:

[0149] ,

[0150] In the formula, is the training evaluation indicator of the autoencoder is the first adjustment weight of the autoencoder, is the second adjustment weight of the autoencoder, is the information gain of the

[0151] Repeat the iteration of S402 - S405 until the iteration stops when the evaluation index is less than the preset threshold, which indicates that the model training is completed. After the autoencoder model training is completed, the output features of the encoder are the low - dimensional feature vectors after feature extraction. The low - dimensional feature vectors pass through the preset Softmax function to obtain the prediction scores of the power cable data.

[0152] After the pre - training is completed, fix the parameters of the two feature extraction networks and transfer them to the power cable data in the target domain for fine - tuning. Divide the data samples of the power cable data in the target domain into a support set and a query set according to the meta - learning strategy. Use the power cable data in the support set to fine - tune the state monitoring and fault prediction model of transfer learning. The fine - tuning strategy can adopt the gradient descent method.

[0153] After completing the fine - tuning of the state monitoring and fault prediction model, input the query set of the power cable data in the target domain into the neural network model and the autoencoder model respectively to obtain their corresponding prediction scores. Then use the Softmax function to normalize the prediction scores. Finally, fuse the two prediction scores and output them as the final prediction result. The classification category is the category corresponding to the maximum value of the prediction score.

[0154] The above is only the preferred embodiment of the present invention and is not used to limit the protection scope of the present invention.

Claims

1. A method for monitoring heating status and predicting faults of power cables based on artificial intelligence, characterized in that: The following steps are involved: S1, collecting power cable data in operation, and marking the collected power cable data, the marked categories include: normal, warning, and fault; S2, the training framework of the condition monitoring and fault prediction model is a fusion framework of meta-learning and transfer learning; S3, uses a neural network based on the dynamic influence of features. The hidden layer of the neural network uses multi-scale feature fusion and high-order feature interaction strategies for dynamic calculation; The multi-scale feature fusion method is expressed as: , In the formula, is the first input to the neural network The fusion features of the layers, is the index of the first neural network layer, S is the number of multi-scales, s is the index of the scale, is the fusion weight of the s-th scale, and the fusion weight of the s-th scale is a training parameter obtained through dynamic learning. is the nonlinear activation function of the s-th scale, The neural network The weight parameter of the s-th scale of the layer, is the fusion feature input to the l-1th layer of the neural network, The neural network Bias parameter of the sth scale of the layer; S4, adopts an autoencoder based on feature path decoupling, learns the compressed representation of power cable data through the autoencoder, and uses the feature path decoupling mechanism to optimize the independence between features. The feature path decoupling mechanism strengthens the independence between features through the regularized loss function, ensuring that even if some features are missing, the model can still capture the main structure of the power cable data; The loss function calculation method of the autoencoder is expressed as: ， In the formula, is the total loss function of the autoencoder, including reconstruction error and regularization term, x represents the independent variable of the function, is the encoder function, is the decoder function, is the regularization weight, is the decoupling regularization function; The role of the decoupling regularization function is to evaluate the independence of features, and the calculation method is expressed as: ， Where z is the input of the decoupling regularization function, The index of the input feature for the first decoupling function, The index of the input feature for the second decoupling function, is the tth iteration The weight of the feature, Representation characteristics and The covariance of , which measures the linear dependence between features, The first input of the decoupling function Features, The first input of the decoupling function Features, is the decoupling function variance weight factor, The index of the input feature for the third decoupling function, Features The variance of , enables the decoupling function to improve the information retention ability by suppressing small variance features, The first input of the decoupling function Features.

2. The method for monitoring heating status and predicting faults of power cables based on artificial intelligence according to claim 1 is characterized in that: The power cable data in S1 include the real-time current value Ra of the cable, the temperature value Da of the cable, the ambient temperature Ca, the ambient humidity Fa, the voltage value Ga of the cable, the wind speed Va, the solar radiation intensity Ma, the historical load record Ba of the cable, the aging degree Na of the insulating material, and the connection status Ka of the cable joint.

3. The method for monitoring heating status and predicting faults of power cables based on artificial intelligence according to claim 1 is characterized in that: The training process of the fusion framework in S2 includes a pre-training stage and a fine-tuning stage; In the pre-training stage, the power cable data samples are input into the feature extraction modules of transfer learning and meta-learning, wherein the power cable data samples with batches as training units are input into the transfer learning feature extraction module, and the power cable data samples with tasks as training units are input into the meta-learning feature extraction module; In the fine-tuning stage, the training set of the target domain power cable data is divided into batches and tasks as basic units. The training set of the target domain power cable data is the set of labeled power cable data in the target domain power cable data. The pre-trained condition monitoring and fault prediction models of meta-learning and transfer learning are used to classify the target domain power cable data, and the scores of the classification results of the two condition monitoring and fault prediction models are fused as the final classification results of the fusion framework.

4. The method for monitoring heating status and predicting faults of power cables based on artificial intelligence according to claim 3 is characterized in that: The meta-learning adopts a model weak correlation meta-learning strategy. The learning goal of the model weak correlation meta-learning strategy is to quickly learn and adapt to the optimal parameters required for new tasks through a parameter optimization algorithm, including a two-layer loop mode of inner loop and outer loop; The inner loop uses a basic learner to extract features of a specific task, and the basic learner is a neural network model; the outer loop uses a meta-learning strategy to update the initialization parameters of the basic learner through gradient descent.

5. The method for monitoring heating status and predicting faults of power cables based on artificial intelligence according to claim 1, characterized in that: The training process of the neural network algorithm based on dynamic influence of features in S3 includes the following steps: S301, initialize the structure and parameters of the neural network to adapt to the distribution characteristics of different power cable data sets, expressed as: ， In the formula, For the neural network The weight parameters of the layer, The mean is , the variance is Initialize with a normal distribution, is the mean of the neural network parameter initialization, is the variance of the neural network parameter initialization, An orthogonal perturbation matrix aligned with the input dimension ensures that the initial parameters of the neural network are highly sensitive to the input features; In order to optimize the sparseness of the neural network, the pruning threshold of the neural network is estimated according to the complexity of the input power cable data during initialization, which is expressed as: ， In the formula, is the pruning threshold of the neural network, is the pruning adjustment coefficient of the neural network, is the characteristic variance of the input power cable data, is the variance function, is the power cable data feature input to the neural network; S302: For the power cable data input into the neural network, a feature redundancy suppression method is adopted to weaken the influence of redundant features through dynamic adjustment based on feature correlation weight parameters. The calculation method of the correlation weight parameters is expressed as: ， In the formula, For the The first The correlation coefficient of the features, Features and The covariance of is the first input to the neural network Features, is the first input to the neural network Features, and Characteristics and The variance of The index of the input feature for the first neural network, The index of the input feature for the second neural network; According to the feature correlation weight parameter, the weight of the input feature is dynamically adjusted, which is expressed as: ， In the formula, is the adjusted input to the neural network Features S303. The hidden layer of the neural network adopts a multi-scale feature fusion and high-order feature interaction strategy to achieve progressive abstraction of features from low-order to high-order, and enhance the interactive representation capability between features; The nonlinear activation function used by the neural network can enhance the expression ability of nonlinear regions and ensure the smooth propagation of gradients in positive and negative regions through dynamic adjustment. The calculation method is expressed as: ， In the formula, The neural network The nonlinear activation function of the layer, is the input of the nonlinear activation function, is the activation function of the rectified linear unit, is the first dynamic adjustment factor, is the hyperbolic tangent activation function, is the second dynamic regulatory factor; S304, dynamically adjust the gradient update amplitude by calculating the importance coefficient of each layer feature of the neural network, and the calculation method is expressed as: , In the formula, For the neural network The weight update increment of the layer, is the learning rate of the neural network, For the neural network The entropy of the layer gradient, is the symbol of partial derivative, is the loss function of the neural network, specifically the cross entropy loss, which is obtained by calculating the category probability of the output features of the neural network using the preset Softmax function and then calculating it through the cross entropy loss function. The neural network The weight parameters of the layer, For the neural network Importance factor of layer features; The feature importance factor dynamically measures the importance of the feature to the global output to ensure a more reasonable gradient distribution. The calculation method is expressed as: ， In the formula, is the variance function, is the first input to the neural network The fusion features of the layers, is the total number of layers in the neural network, is the index of the second neural network layer; The learning rate of the neural network is adjusted based on the rate of change of the gradient, and the calculation method is expressed as: , In the formula, For the neural network The base learning rate of the layer, For the neural network The learning rate adjustment coefficient of the layer indicates the impact of gradient changes on the learning rate. for norm, is the updated gradient of the loss function in this iteration and the previous iteration; During the training process of each layer, the entropy value of the neural network gradient is calculated to ensure that the learning speed of each layer is coordinated. The calculation method is expressed as: , In the formula, The neural network The number of neurons in the layer, For the neural network Tier The gradient of a neuron, For the neural network The index of the layer neuron, is a logarithmic function; Update the neural network weight parameters, and the update method is expressed as: ， In the formula, is the parameter update operation, For the neural network Regularization factor of the layer; In order to avoid overfitting, the regularization factor of the neural network is related to the entropy value of the gradient. The regularization factor decreases as the gradient increases, thereby reducing the model's sensitivity to overfitting and effectively adjusting the learning progress of each layer in the network. The calculation method is expressed as: ， In the formula, The neural network The hyperparameters of the layer’s regularization factor affect the degree of influence. is the weight update gradient of the neural network, for norm; S305, repeat the above steps until the preset stop iteration condition is met. After the neural network model training is completed, the output feature of the neural network is the feature vector after feature extraction. The feature vector is passed through a preset Softmax function to obtain the prediction score of the power cable data.

6. The method for monitoring heating status and predicting faults of power cables based on artificial intelligence according to claim 1, characterized in that: The training process of the autoencoder algorithm based on feature path decoupling in S4 includes the following steps: S401, initialize the encoder and decoder of the autoencoder. The encoder is responsible for mapping the high-dimensional input power cable data to a low-dimensional space, and the decoder attempts to reconstruct the original power cable data from the low-dimensional representation. The initialization method is expressed as: ， In the formula, and The encoder and decoder are The weight matrix of the layer, and The encoder and decoder are The bias vector of the layer, For encoder The dimension of the layer, For encoder The dimension of the layer, and to generate normally distributed random numbers and zero vectors, respectively. is the layer index of the encoder and decoder; S402, during the encoding process, decoupling the dependency path of the input features by a regularization method; S403, through the dynamic feature flow embedding mechanism, the model dynamically adjusts the weight of each feature in the hidden layer according to the reconstruction error and feature importance to adapt to the power cable data characteristics and task requirements. The adjustment method is expressed as: ， In the formula, is the t+1th iteration The weight of the feature, is the tth iteration The weight of the feature, is the dynamic feature flow embedding learning rate, is the gradient of the loss function with respect to the feature weight, Embedding tuning parameters for dynamic feature flow, For the The information gain of each feature; S404, the training process adopts a cyclic process of feedforward, feedback and parameter update, and minimizes the loss function through the gradient descent method. The parameter update method of the encoder is expressed as: ， In the formula, is the parameter update operation, and are the loss functions of the autoencoder for the Gradients of layer weights and biases; S405. The model performance is measured by comprehensive evaluation indicators, while optimizing the reconstruction error and feature information. The calculation method of the evaluation indicators is expressed as: ， In the formula, is the training evaluation index of the autoencoder, is the first adjustment weight of the autoencoder, is the second adjustment weight of the autoencoder, For the The information gain of each feature; Repeat iteration S402-S405 until the evaluation index is less than the preset threshold, then stop the iteration, which means the model training is completed. After the autoencoder model training is completed, the output feature of the encoder is the low-dimensional feature vector after feature extraction. The low-dimensional feature vector is passed through the preset Softmax function to obtain the prediction score of the power cable data.

Images