Graphene cable monitoring system based on deep learning

Abstract

The invention relates to the technical field of cable monitoring, in particular to a graphene cable monitoring system based on deep learning, which is used for extracting time-frequency domain characteristics by acquiring current, voltage, electromagnetic wave signals and temperature distribution data and utilizing Fourier transform and wavelet transform to improve the data identification degree. Gaussian filtering noise reduction is carried out on temperature data, and an abnormal hot spot area is identified through image segmentation, so that the local overheating detection capability is improved. And the edge calculation module fuses multi-source data, eliminates acquisition delay by utilizing feature alignment, and improves data synchronism and fusion quality. Through a multi-head self-attention mechanism, time sequence characteristics of historical monitoring data are extracted, and change modes of current, voltage, electromagnetic wave and temperature distribution are learned. And calculating an attention weight matrix to extract correlation between time steps, and forming a time sequence feature matrix. The characteristic matrix is subjected to nonlinear transformation through a feedforward neural network, cable state parameters of a future time step are predicted, the cable state is evaluated in advance, and the fault risk is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of cable monitoring, and particularly to a graphene cable monitoring system based on deep learning. Background Art

[0002] As a new type of high-performance conductor material, graphene cables have broad application prospects in high-voltage power transmission and smart grids due to their excellent electrical conductivity, thermal conductivity, and corrosion resistance. However, due to the high sensitivity of graphene materials, their conductive state, temperature distribution, and electromagnetic environment are easily affected by working conditions, and traditional cable monitoring methods are difficult to meet the requirements of high-precision and real-time monitoring. Currently, traditional cable monitoring systems mainly rely on regular manual inspections, sensor monitoring, and rule-based threshold analysis. However, these methods have many limitations: Manual inspection lag: Regular inspections cannot achieve continuous monitoring of cables, easily miss early fault signals, resulting in the accumulation of faults and affecting the stability of the power system. The single-sensor monitoring plus traditional threshold alarm method cannot effectively distinguish noise from real faults, is prone to false alarms or missed alarms, affects the reliability of fault location, and lacks the ability to accurately predict the future state of cables, making it difficult to give early warnings. Summary of the Invention

[0003] To solve the above problems, the present invention provides a graphene cable monitoring system based on deep learning.

[0004] To achieve the above object, the technical solution adopted by the present invention is:

[0005] A graphene cable monitoring system based on deep learning, comprising: a working parameter acquisition module, a thermal imaging acquisition module, an edge computing module, and a deep learning monitoring module;

[0006] The working parameter acquisition module is used to acquire current, voltage, and electromagnetic wave signals in the working state of the graphene cable;

[0007] The thermal imaging acquisition module is used to acquire temperature distribution data in the working state of the graphene cable;

[0008] The edge computing module is used to preprocess the current, voltage, electromagnetic wave signals, and temperature distribution data in the working state of the graphene cable, and input the preprocessed data into the deep learning monitoring module. The preprocessing includes data denoising, feature extraction, and data fusion;

[0009] The deep learning monitoring module is used to predict the cable state parameters at a future time based on the Transformer architecture through the preprocessed data; generate a risk score based on the cable state parameters at the future time, and perform real-time warning monitoring according to the risk score.

[0010] Further, the current, voltage, and electromagnetic wave signals in the working state of the graphene cable are collected and obtained through a current sensor, a voltage sensor, and an electromagnetic wave sensor respectively.

[0011] Further, the current, voltage, electromagnetic wave signals, and temperature distribution data in the working state of the graphene cable are preprocessed. Among them, the preprocessing of the current, voltage, and electromagnetic wave signals in the working state of the graphene cable includes the following steps:

[0012] Sample the collected current, voltage, and electromagnetic signals at a preset frequency, and synchronize the data at uniform time intervals;

[0013] Use an adaptive filtering algorithm to denoise the synchronized current, voltage, and electromagnetic signals;

[0014] Extract the frequency-domain features and time-domain features in the signals through Fourier transform and wavelet transform, and perform normalization processing.

[0015] Further, the current, voltage, electromagnetic wave signals, and temperature distribution data in the working state of the graphene cable are preprocessed. Among them, the preprocessing of the temperature distribution data includes the following steps:

[0016] Collect temperature data on the surface and surrounding environment of the graphene cable through a thermal imaging sensor, and the sampling frequency is synchronized with the sampling frequency of the current, voltage, and electromagnetic wave signals;

[0017] Use a Gaussian filtering algorithm to denoise the thermal imaging data, and enhance the image features based on contrast stretching and histogram equalization algorithms;

[0018] Detect abnormal temperature hot spots based on an image segmentation algorithm and convert them into feature codes.

[0019] Further, based on the Transformer architecture, predicting the cable state parameters at future times through the preprocessed data includes the following steps:

[0020] Receive the normalized data output by the edge computing module, including current, voltage, electromagnetic wave signals, and temperature distribution data, and perform position encoding to generate time series feature data;

[0021] Calculate the correlation between data at different time steps in the time series feature data based on the multi-head self-attention mechanism to generate a time series feature matrix;

[0022] Input the time series feature matrix into a feed-forward neural network for non-linear transformation to calculate the cable state parameters at future time steps, including future predicted current values, voltage predicted values, electromagnetic wave signal predicted values, and temperature distribution predicted values.

[0023] Further, the calculation formula of the multi-head self-attention mechanism is as follows:

[0024]

[0025] Among them, Z is the time series feature matrix; X is the time series feature data; is the query weight matrix of the i-th attention head; is the key weight matrix of the i-th attention head; is the value weight matrix of the i-th attention head; d k is the dimension of the key matrix; W o is the projection transformation matrix after multi-head attention calculation; h is the number of multi-head attentions; softmax is the activation function.

[0026] Further, the feed-forward neural network is trained through the following steps:

[0027] Receive the time series feature matrix as the input data of the feed-forward neural network, and successively pass through the fully connected layer, activation function and batch normalization to generate intermediate feature representations;

[0028] Map the intermediate features, calculate the cable state parameters at future time steps, and calculate the prediction error through the loss function;

[0029] Based on the prediction error, adjust the network parameters through the backpropagation algorithm.

[0030] Further, generating the risk score based on the cable state parameters at future times includes the following steps:

[0031] Receive the predicted current value, predicted voltage value, predicted electromagnetic wave signal value and predicted temperature distribution value at future time steps, compare them with the historical operation reference values, and calculate the state deviation value;

[0032] Calculate the risk score by weighting the state deviation value.

[0033] Further, the real-time early warning monitoring based on the risk score includes:

[0034] Receive the calculated risk score, compare it with the preset risk threshold, and classify the cable state into normal, abnormal and faulty;

[0035] Send the cable state to the monitoring platform for voice or text alarm.

[0036] The beneficial effects of the present invention are as follows: The present invention obtains current, voltage, electromagnetic wave signal and temperature distribution data. For the current, voltage and electromagnetic wave signal data, the time-frequency domain features are extracted through Fourier transform and wavelet transform to improve the recognizability of the data. For the temperature distribution data, the Gaussian filtering algorithm is used for noise reduction, and the abnormal hot spot area is extracted through the image segmentation algorithm to improve the detection ability of local overheating phenomena. In addition, the edge computing module further fuses the information of different data sources, eliminates the acquisition delay and asynchronous problems by feature alignment, and improves the fusion quality of multi-modal data. The time series prediction ability is improved through the Transformer model. The traditional threshold-based monitoring method is difficult to identify the fault trend in advance, while this solution uses the multi-head self-attention mechanism to construct a Transformer time series prediction model, which can extract time series features from historical monitoring data and learn the change patterns of current, voltage, electromagnetic wave signal and temperature distribution. The Transformer model first performs position encoding on the input data to retain the time series information, then calculates the attention weight matrix, extracts the correlation between different time steps, and forms a time series feature matrix. This feature matrix undergoes a non-linear transformation through a feed-forward neural network to obtain the predicted values of the cable state parameters at future time steps, including the predicted values of future current, voltage, electromagnetic wave signal and temperature distribution, so as to realize the early evaluation of the cable state and reduce the possibility of faults. Description of the Drawings

[0037] Figure 1 It is a schematic structural diagram of a graphene cable monitoring system based on deep learning in the present invention.

[0038] Figure 2 It is a diagram of predicting the cable state parameters at future time based on the Transformer architecture through preprocessed data in the present invention. Detailed Embodiments

[0039] Please refer to Figure 1 - Figure 2 As shown, the present invention relates to a graphene cable monitoring system based on deep learning, including: a working parameter acquisition module, a thermal imaging acquisition module, an edge computing module and a deep learning monitoring module;

[0040] The working parameter acquisition module is used to acquire the current, voltage and electromagnetic wave signals in the working state of the graphene cable;

[0041] The thermal imaging acquisition module is used to acquire the temperature distribution data in the working state of the graphene cable;

[0042] The edge computing module is used to preprocess the current, voltage, electromagnetic wave signal and temperature distribution data in the working state of the graphene cable, and input the preprocessed data into the deep learning monitoring module. The preprocessing includes data denoising, feature extraction and data fusion;

[0043] The deep learning monitoring module is used to predict the cable state parameters at future times based on the Transformer architecture through the preprocessed data; generate risk scores based on the cable state parameters at future times, and conduct real-time warning monitoring according to the risk scores.

[0044] In some embodiments, first, the working parameter acquisition module obtains the current, voltage, and electromagnetic wave signal data of the graphene cable in real time. Specifically, this module uses high-precision sensors to perform multi-frequency sampling on the current, voltage, and electromagnetic wave signals, and synchronizes the data at uniform time intervals to ensure that the acquired signals have good temporal continuity. To reduce the interference of environmental noise on the data accuracy, the module is built with an adaptive filtering algorithm that can dynamically adjust the filtering parameters according to different noise characteristics, thereby significantly improving the signal-to-noise ratio of the acquired data. The processed signals contain the core features in the time domain and frequency domain, which can provide high-quality basic data for subsequent feature extraction and modeling. The thermal imaging acquisition module is responsible for obtaining the temperature distribution data of the cable surface and the surrounding environment in real time. A high-resolution thermal imaging sensor is used to collect thermal images of the cable surface, and the sampling frequency is synchronized with the working parameter data to ensure that the data in each dimension can achieve temporal matching. During the acquisition process, the Gaussian filtering algorithm is used to denoise the thermal image data to eliminate the random noise generated by external light or sensor hardware. To accurately identify possible local overheating phenomena, the system applies an image segmentation algorithm based on K-means clustering to divide the thermal image data into regions, extracts the hot spots with significant temperature anomalies, and further converts them into quantitative feature codes as input data for subsequent analysis. In the edge computing module, comprehensive preprocessing is performed on the working parameter data and the thermal imaging data. This module first aligns the data streams from different sensors in time, adjusts the timestamps of different source data through the Dynamic Time Warping (DTW) method, and eliminates the asynchronous problems during the acquisition process. Subsequently, feature extraction is performed on the aligned multi-source data. Fourier transform and wavelet transform are used for the working parameter data to extract frequency domain features and time domain features, while the thermal imaging data extracts the hot spot region information and the temperature gradient change information through image feature coding. To achieve data fusion, the module uses a feature fusion method based on weighted average to integrate the working parameter data and the thermal imaging data according to weights to generate a unified multi-modal input feature vector. The deep learning monitoring module models and analyzes the preprocessed multi-modal features based on the Transformer architecture. The system first performs time series annotation on the input multi-modal feature vector through the position encoding technique to retain the temporal information of the data. Subsequently, the multi-head self-attention mechanism is used to calculate the correlation between different time steps in the input features to generate a temporal feature matrix. After being processed by the feed-forward neural network, the temporal feature matrix outputs the cable state parameters at future time steps, including key parameters such as predicted current, voltage, electromagnetic wave signals, and temperature distribution. These prediction results can provide trend information about the cable operating state, especially under high load or extreme environmental conditions, and can effectively identify potential fault risks. Based on the predicted cable state parameters, the system generates a risk score through the risk assessment module.The risk assessment module calculates the deviation value between the actual state and the predicted state, and generates a comprehensive risk score by combining historical data and the set weight parameters. According to the magnitude of the risk score, the system triggers warning signals at different levels in real time. Combining with the fast response ability of the edge computing module, the warning information is sent to the operation and maintenance personnel or the monitoring platform in a timely manner to achieve early detection and proactive intervention of cable faults.

[0045] Further, the current, voltage, and electromagnetic wave signals in the working state of the graphene cable are respectively collected by a current sensor, a voltage sensor, and an electromagnetic wave sensor.

[0046] It should be noted that the collection of the current signal uses a Hall effect current sensor to perform non-contact measurement on the current in the cable. This sensor can sense the change of the current magnetic field in the cable and convert it into a corresponding current signal. The sensor has high linearity and low noise characteristics and can maintain stable performance within a wide temperature range. The collected current signal is digitally processed by an analog-to-digital conversion module (ADC), and the power frequency interference and high-frequency noise are removed through a band-pass filter to ensure signal purity. The voltage signal is collected by connecting a voltage divider to the cable line and combining a high-impedance voltage sensor to collect the voltage at both ends of the cable. To reduce the sampling error, the sensor is designed with a capacitive coupling and a resistance matching network and can operate stably in a high-voltage environment. The collected voltage signal is oversampled to improve the signal resolution, and at the same time, the Kalman filtering algorithm is used to eliminate the influence of external electromagnetic interference on the voltage signal. The electromagnetic wave signal is collected by a broadband electromagnetic wave sensor to monitor the electromagnetic wave radiation intensity and distribution around the cable. This sensor can capture the high-frequency electromagnetic radiation caused by the current change during the operation of the cable.

[0047] Further, the current, voltage, electromagnetic wave signals, and temperature distribution data in the working state of the graphene cable are preprocessed. Among them, the preprocessing of the current, voltage, and electromagnetic wave signals in the working state of the graphene cable includes the following steps:

[0048] Sample the collected current, voltage, and electromagnetic signals at a preset frequency, and synchronize the data at uniform time intervals;

[0049] Use an adaptive filtering algorithm to perform noise reduction processing on the synchronized current, voltage, and electromagnetic signals;

[0050] Extract the frequency-domain features and time-domain features in the signal through Fourier transform and wavelet transform, and perform normalization processing.

[0051] It should be noted that, first, the collected current, voltage, and electromagnetic wave signal data are sampled at a preset frequency to ensure that the signals have sufficient resolution and temporal continuity. The sampling frequency is determined according to the characteristics of signal changes. For example, the sampling frequency of current and voltage signals is set to more than ten times the power frequency (50 / 60 Hz) to capture possible harmonics and mutations. The sampling frequency of electromagnetic wave signals is dynamically adjusted according to the target frequency band coverage and signal bandwidth, usually using a sampling frequency in the range of several hundred kilohertz to megahertz. After sampling, a global timestamp allocation mechanism is used to synchronize the multi-source signals. This process uses the Dynamic Time Warping (DTW) method to correct the time deviation between sensors, thereby achieving the alignment of data in the time dimension. The synchronized signal data usually contains various noise components, such as power frequency interference and high-frequency random noise introduced by the external environment or the device itself. These noises may mask the characteristics of the true signal and affect the effect of subsequent feature extraction. Therefore, this solution uses an adaptive filtering algorithm to denoise the current, voltage, and electromagnetic wave signals. Specifically, by analyzing the power spectrum distribution of the signal, the adaptive filtering algorithm dynamically adjusts the parameters of the filter to suppress noise components in different frequency ranges. For example, for the power frequency interference of current and voltage signals, a notch filter is used; for the high-frequency noise of electromagnetic wave signals, a combination of a band-pass filter and a low-pass filter is used to achieve multi-level noise reduction. Further, Fourier transform and wavelet transform are used to extract the frequency domain features and time domain features of the signal. Fourier transform is used to convert the signal from the time domain to the frequency domain, analyze the frequency components and their amplitude distribution of the signal, so as to identify specific frequency anomalies that may occur during the operation of the cable, such as harmonic or electromagnetic interference frequency bands. Wavelet transform captures the local feature changes of the signal in different time and frequency ranges through multi-scale decomposition of the signal, and can effectively reflect the changes of short-term emergencies and low-frequency trends. The extracted features include the energy distribution of harmonic components, the frequency response of transient signals, and the energy ratio of the characteristic frequency band corresponding to a specific fault mode.

[0052] Furthermore, the current, voltage, electromagnetic wave signals, and temperature distribution data in the working state of the graphene cable are preprocessed. Among them, the preprocessing of the temperature distribution data includes the following steps:

[0053] The temperature data of the surface and surrounding environment of the graphene cable are collected by a thermal imaging sensor, and the sampling frequency is synchronized with the sampling frequency of the current, voltage, and electromagnetic wave signals;

[0054] The Gaussian filtering algorithm is used to denoise the thermal imaging data, and the image features are enhanced based on the contrast stretching and histogram equalization algorithms;

[0055] Based on the image segmentation algorithm, the abnormal temperature hot spot area is detected and converted into a feature code.

[0056] Specifically, first, the thermal imaging sensor is used to collect the temperature distribution data of the graphene cable surface and its surrounding environment in real time. The sensor adopts non-contact temperature measurement technology, which can capture the thermal radiation characteristics of the cable surface and generate thermal image data. During the collection process, the sampling frequency is consistent with the current, voltage, and electromagnetic wave signals to ensure the time synchronization between multi-modal data. This synchronization mechanism is achieved through global timestamp allocation, which guarantees the consistency of different data sources in the time dimension, thus facilitating subsequent data fusion and modeling. The collected thermal image data usually contains random noise and interference caused by non-ideal characteristics of the sensor, such as environmental light changes and sensor sensitivity fluctuations. Therefore, this solution uses the Gaussian filtering algorithm to denoise the thermal image data. Specifically, by applying the Gaussian kernel function to the image, the weighted average of each pixel is performed to smooth the high-frequency noise in the image while retaining the temperature change characteristics of the edges and local regions. The kernel size of the Gaussian filter is dynamically adjusted according to the thermal imaging resolution to ensure the optimality of the denoising effect. To further improve the feature distinguishability of the thermal image, this solution applies contrast stretching and histogram equalization algorithms to the denoised thermal image for image enhancement processing. In the contrast stretching stage, by adjusting the dynamic range of the image pixel values, the detailed information in the low-contrast region is significantly amplified, making the boundaries of the abnormal temperature regions clearer. Subsequently, histogram equalization redistributes the probability density of the pixel values, evenly distributing the pixel values in the bright region concentration throughout the entire dynamic range, thereby optimizing the overall contrast of the image. In the feature extraction stage, an image segmentation algorithm based on K-means clustering is used to process the thermal image to detect abnormal temperature hotspots. Specifically, the thermal image pixels are divided into several categories, and the central value of each category is used to represent the average temperature characteristics of the corresponding region. The segmentation algorithm can identify hotspots with significant temperature anomalies, such as overheating regions that may be caused by local overload or insulation aging. To improve the accuracy of the segmentation results, morphological processing methods (such as erosion and dilation operations) are combined to further optimize the segmentation boundary, removing isolated noise points and enhancing the coherence of the target region. Finally, the features of the detected hotspot regions are standardized and encoded to generate quantitative feature vectors. These features include the area of the hotspot region, the maximum temperature value, the temperature gradient distribution, and the spatial position of the hotspot region. All features are output in a fixed format for input to the subsequent deep learning model. The standardization process uses linear normalization to ensure the consistency of different feature dimensions and improve the comparability of the features.

[0057] Furthermore, the steps for predicting the cable state parameters at future times based on the data after preprocessing by the Transformer architecture are as follows:

[0058] Receive the normalized data output by the edge computing module, including current, voltage, electromagnetic wave signal, and temperature distribution data, and perform position encoding to generate time series feature data;

[0059] Calculate the correlation between data at different time steps in the time series feature data based on the multi-head self-attention mechanism to generate a time series feature matrix;

[0060] Input the time series feature matrix into a feed-forward neural network for non-linear transformation to calculate the cable state parameters at future time steps, including predicted values of future current, voltage, electromagnetic wave signal, and temperature distribution.

[0061] Specifically, first, receive the normalized data output by the edge computing module, including current, voltage, electromagnetic wave signal, and temperature distribution data. These data have undergone preprocessing in the early stage to complete time series synchronization, noise reduction, and normalization to ensure the standardization and consistency of the input data. To retain the time series information of the data, the system encodes the normalized data through position encoding technology, corresponding each time step of the data with its time position. The position encoding uses sine and cosine functions to generate a vector of fixed length and adds this vector to the input data to form time series feature data, thereby retaining the time position information in the subsequent model. Next, the multi-head self-attention mechanism based on the Transformer architecture calculates the correlation of the time series feature data to extract key time series features. The self-attention mechanism first calculates the query vector (Query), key vector (Key), and value vector (Value) of the input data, and then generates a similarity matrix of the query and the key through inner product operation to measure the correlation between different time steps. After normalization processing, the similarity matrix is used to weight the value vector to obtain the time series weighted result. To improve the expression ability of the model, the multi-head self-attention mechanism independently executes the above process in multiple subspaces, and each subspace focuses on different time series feature dimensions. Finally, the results of the multi-head mechanism are concatenated and projected into the target dimension to generate a time series feature matrix containing global time series information. Input the generated time series feature matrix into a feed-forward neural network for non-linear transformation. The feed-forward neural network consists of multiple fully connected layers and can further extract high-order features. Specifically, each fully connected layer is followed by an activation function (such as ReLU) and batch normalization (BatchNormalization) to enhance the non-linear expression ability of the model and improve the stability of training. The output result is mapped to the target dimension through a projection layer, and finally the cable state parameters at future time steps are generated, including predicted values of current, voltage, electromagnetic wave signal, and temperature distribution. Each predicted value output by the model is associated with the time step and can clearly reflect the change trend of the cable state in the future for a period of time.

[0062] Furthermore, the calculation formula of the multi-head self-attention mechanism is as follows:

[0063]

[0064] where Z is the time series feature matrix; X is the time series feature data; is the query weight matrix of the i-th attention head; is the key weight matrix of the i-th attention head; is the value weight matrix of the i-th attention head; d k is the dimension of the key matrix; W O is the projection transformation matrix after multi-head attention calculation; h is the number of multi-head attention; softmax is the activation function.

[0065] Specifically, it captures the global correlation between time steps and generates high-quality time series features, providing basic support for the prediction of subsequent deep learning models.

[0066] The multi-head self-attention mechanism takes the normalized time series feature data as input, specifically including current, voltage, electromagnetic wave signal, and temperature distribution data. These data are linearly transformed to generate three vectors: query, key, and value. The query vector represents the information requirement of the target time step, the key vector represents the content features of other time steps, and the value vector contains the specific information of this content. Next, the inner product between the query vector and the key vector is calculated to measure the correlation between the target time step and other time steps. To avoid the situation of unstable gradients caused by too large inner product values, the correlation values are normalized by the square root of the key vector dimension. Subsequently, the softmax function is applied to convert the normalized correlation values into a weight probability distribution to ensure the normalization of the total weight. This normalization process can highlight the time positions most relevant to the target time step while weakening the influence of irrelevant parts. The calculated weight values are used to perform weighted summation on the value vectors, thereby obtaining the important feature information of each time step in the time series. Through this weighting mechanism, the multi-head self-attention can dynamically adjust the influence of different time steps in the model, making the features of key time points more prominent, thus improving the model's ability to capture important time series features. The significant feature of the multi-head self-attention mechanism lies in its parallel computing ability. Through multiple independent attention heads, the system can simultaneously calculate multiple time series features in different subspaces, capturing multi-dimensional information in the data. The features generated by each attention head are concatenated and then further fused through a linear projection matrix to form a unified time series feature matrix. This matrix reflects the global correlation of the input data in the time dimension and is an important input for subsequent deep learning models.

[0067] Furthermore, the feed-forward neural network is trained through the following steps:

[0068] Receive the timing feature matrix as the input data of the feedforward neural network, and sequentially pass through the fully connected layer, activation function, and batch normalization to generate the intermediate feature representation;

[0069] Map the intermediate features, calculate the cable state parameters at future time steps, and calculate the prediction error through the loss function;

[0070] Based on the prediction error, adjust the network parameters through the backpropagation algorithm.

[0071] Specifically, first, the time-series feature matrix serves as the input data for the feed-forward neural network. The time-series feature matrix is generated by the multi-head self-attention mechanism and contains the time-series features of multi-modal data such as current, voltage, electromagnetic wave signals, and temperature distribution. The input data passes through the first fully connected layer of the neural network, which maps the high-dimensional feature data into the latent feature space. The fully connected layer extracts the preliminary relationships of the features by applying linear transformations and weighting operations to the input data. This step is achieved through trainable weight and bias parameters and incorporates the interaction information between multi-dimensional features. Next, the features output by the fully connected layer are fed into the activation function for non-linear transformation. Commonly used activation functions include ReLU (Rectified Linear Unit) or GELU (Gaussian Error Linear Unit) to increase the non-linear representation ability of the model, thereby capturing the complex non-linear relationships in the input data. Through the action of the activation function, the model can learn higher-order feature representations, which is crucial for capturing the dynamic changes of complex cable states. Subsequently, to further stabilize the training process of the model and improve the convergence speed, batch normalization operations are added to each subsequent layer of the model. Batch normalization normalizes the feature distribution of each batch of data to a range with a mean of zero and a variance of one, reducing the dimensionality differences between different features and suppressing the occurrence of gradient vanishing or gradient explosion phenomena. After completing feature extraction, the generated intermediate feature representations are input into the last fully connected layer of the network to be mapped to the target output dimension, thereby calculating the cable state parameters for future time steps. Specifically, the state parameters output by the model include predicted values of future current, voltage, electromagnetic wave signals, and temperature distribution. The output results are compared with the true state parameter labels, and the loss function is used to calculate the prediction error. Commonly used loss functions include mean squared error (MSE) or mean squared logarithmic error (MSLE), and an appropriate loss function is selected according to the task requirements to measure the deviation between the predicted value and the true value. Based on the prediction error calculated by the loss function, the system updates the network parameters through the backpropagation algorithm. The backpropagation algorithm combines the gradient information of the error and calculates the partial derivatives of the loss function with respect to the network weights and biases layer by layer through the chain rule. Subsequently, optimization algorithms (such as Adam or SGD) are used to adjust the model parameters to gradually reduce the loss function value in each iteration, thereby improving the prediction accuracy of the model. The learning rate and weight decay parameters of the optimization algorithm are dynamically adjusted during the training process to balance the convergence speed and the generalization ability of the model.

[0072] Furthermore, generating the risk score from the cable state parameters at future times includes the following steps:

[0073] Receive the predicted current value, voltage value, electromagnetic wave signal value, and temperature distribution value at future time steps, compare them with the historical operation reference values, and calculate the state deviation value;

[0074] Calculate the risk score by weighting the state deviation value.

[0075] Specifically, first, the system receives the cable state parameters predicted by the deep learning model at future time steps, including the predicted current value, voltage value, electromagnetic wave signal value, and temperature distribution value. These predicted values describe the possible state changes of the cable in the future. At the same time, the system calls the corresponding operation reference values from the historical database. The reference values are obtained by long-term monitoring and statistical analysis of the cable state data under normal operating conditions and are of reference significance. Next, the system compares each type of predicted parameter with its corresponding historical reference value one by one to calculate the state deviation value. The magnitude and direction of the state deviation value indicate the difference between the future cable operating state and the normal state. For example, when the predicted current value is much higher than the reference value, it may mean that the cable load exceeds the safe range, indicating a possible overload risk; when the predicted temperature distribution value shows a significant deviation from the reference value, it may point to a potential problem of local overheating. These deviation values directly reflect whether the cable operating state deviates from the safe range and the severity of the deviation. To further quantify the impact of these state deviations on the cable operating safety, the system weights the deviation values of different parameters. The weighting coefficients for the weighted calculation are determined based on the importance of each parameter and the statistical analysis of historical fault data. For example, in the high-voltage power transmission application scenario, the current and temperature parameters usually play a key role in the safe operation of the cable, so their weights are relatively high. In this way, the system comprehensively considers the risk contributions of multiple parameters and generates an overall risk score. The risk score is a quantitative indicator that can accurately reflect the safety level of the future cable operating state. The risk score is divided into different levels to indicate the severity of the cable operating state.

[0076] Furthermore, the real-time warning monitoring based on the risk score includes:

[0077] Receive the calculated risk score, compare it with the preset risk threshold, and classify the cable state as normal, abnormal, and faulty;

[0078] Send the cable state to the monitoring platform for voice or text warning.

[0079] Specifically, according to the comparison result of the risk score and the threshold, the system classifies the cable state into the following three levels:

[0080] Normal: When the risk score is lower than the first threshold, it indicates that the cable is in good operating condition, and all monitoring parameters are within the safe range, and no intervention measures are required.

[0081] Abnormal: When the risk score is between the first threshold and the second threshold, it indicates that the cable may have a slight abnormality. For example, the current or temperature value is close to the safety limit but does not exceed the standard. At this time, the system will prompt the operation and maintenance personnel to strengthen monitoring.

[0082] Fault: When the risk score is higher than the second threshold, it indicates that the operating state of the cable has reached or exceeded the dangerous level, and there may be serious overload, local overheating or other potential fault hazards, and emergency treatment measures need to be taken immediately.

[0083] After the classification is completed, the system sends the cable status to the monitoring platform in the form of a data packet. The data packet contains the following information: risk score, cable status level, triggered threshold information, and metadata such as relevant timestamps. This information is used by the monitoring platform for further display and recording. At the same time, the system uses the built-in voice and text warning mechanism to inform the operation and maintenance personnel of the cable status in real time. For example, when the cable status is rated as "fault", the system will generate a high-priority voice warning to notify the relevant personnel to take emergency treatment; for the "abnormal" status, the system sends a warning message in text form to prompt the operation and maintenance personnel to check for potential hidden dangers.

[0084] The above embodiments are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A graphene cable monitoring system based on deep learning, characterized in that: include: Work parameter acquisition module, thermal imaging acquisition module, edge computing module and deep learning monitoring module; The industrial parameter collection module is used to collect current, voltage and electromagnetic wave signals in the working state of the graphene cable; The thermal imaging acquisition module is used to collect temperature distribution data of the graphene cable in a working state; The edge computing module is used to pre-process the current, voltage, electromagnetic wave signal and temperature distribution data in the working state of the graphene cable, and input the pre-processed data into the deep learning monitoring module, wherein the pre-processing includes data denoising, feature extraction and data fusion; The deep learning monitoring module is used to predict the cable status parameters in the future through preprocessed data based on the Transformer architecture; generate risk scores through the cable status parameters in the future, and perform real-time early warning monitoring based on the risk scores.

2. A graphene cable monitoring system based on deep learning according to claim 1, characterized in that: The current, voltage and electromagnetic wave signals of the graphene cable in the working state are collected and acquired by a current sensor, a voltage sensor and an electromagnetic wave sensor respectively.

3. A graphene cable monitoring system based on deep learning according to claim 1, characterized in that: The preprocessing of the current, voltage, electromagnetic wave signal and temperature distribution data in the working state of the graphene cable comprises the following steps: The collected current, voltage and electromagnetic signals are sampled at a preset frequency and the data are synchronized at uniform time intervals; Adopt adaptive filtering algorithm to reduce noise of synchronized current, voltage and electromagnetic signals; The frequency domain features and time domain features of the signal are extracted through Fourier transform and wavelet transform, and then normalized.

4. A graphene cable monitoring system based on deep learning according to claim 1, characterized in that: The preprocessing of the current, voltage, electromagnetic wave signal and temperature distribution data in the working state of the graphene cable comprises the following steps: The temperature data of the graphene cable surface and the surrounding environment are collected through thermal imaging sensors, and the sampling frequency is synchronized with the sampling frequency of current, voltage and electromagnetic wave signals; Gaussian filtering algorithm is used to reduce the noise of thermal imaging data, and image features are enhanced based on contrast stretching and histogram equalization algorithms; Abnormal temperature hot spots are detected based on image segmentation algorithm and converted into feature codes.

5. The graphene cable monitoring system based on deep learning according to claim 1, characterized in that: The method of predicting the cable state parameters at a future time by using the preprocessed data based on the Transformer architecture includes the following steps: Receive normalized data output by the edge computing module, including current, voltage, electromagnetic wave signal and temperature distribution data, perform position encoding, and generate time series feature data; Based on the multi-head self-attention mechanism, the correlation between data at different time steps in the time series feature data is calculated to generate a time series feature matrix; The time series feature matrix is ​​input into the feedforward neural network for nonlinear transformation to calculate the cable state parameters of the future time step, including the future current prediction value, voltage prediction value, electromagnetic wave signal prediction value and temperature distribution prediction value.

6. A graphene cable monitoring system based on deep learning according to claim 5, characterized in that: The calculation formula of the multi-head self-attention mechanism is as follows: Among them, Z is the time series feature matrix; X is the time series feature data; is the query weight matrix of the i-th attention head; is the key weight matrix of the i-th attention head; is the value weight matrix of the i-th attention head; d k is the dimension of the bond matrix; W O is the projection transformation matrix after multi-head attention calculation; h is the number of multi-head attention; softmax is the activation function.

7. A graphene cable monitoring system based on deep learning according to claim 5, characterized in that: The feedforward neural network is trained by the following steps: Receive the time series feature matrix as the input data of the feedforward neural network, and generate intermediate feature representation through the fully connected layer, activation function and batch normalization in sequence; Map the intermediate features, calculate the cable state parameters for future time steps, and calculate the prediction error through the loss function; Based on the prediction error, the network parameters are adjusted through the back-propagation algorithm.

8. The graphene cable monitoring system based on deep learning according to claim 1, characterized in that: Generating a risk score by using the cable status parameters at a future time comprises the following steps: Receive the current prediction value, voltage prediction value, electromagnetic wave signal prediction value and temperature distribution prediction value of the future time step, compare them with the historical operation reference value, and calculate the state deviation value; The risk score is calculated by weighting the state deviation values.

9. The graphene cable monitoring system based on deep learning according to claim 1, characterized in that: The real-time early warning monitoring according to the risk score includes: Receive the calculated risk score and compare it with the preset risk threshold to classify the cable status into normal, abnormal and faulty; Send the cable status to the monitoring platform for voice or text alarm.

Images