Cable insulation aging defect positioning evaluation self-diagnosis method and system

Abstract

The invention discloses a cable insulation aging defect positioning evaluation self-diagnosis method, and belongs to the technical field of power system detection. According to the method, cable multi-dimensional data are obtained through a multi-band broadband dielectric spectrometer, a high-frequency pulse reflection tester and an optical time domain reflectometer, and features are extracted by applying wavelet transform, Fourier transform and a thermal diffusion model after digital processing. A deep convolutional neural network and a long and short term memory network are used for feature fusion and time sequence analysis, and a self-adaptive threshold algorithm and a fuzzy inference system are used for realizing defect accurate positioning and reliability scoring. According to the method, the accuracy and reliability of cable insulation aging defect diagnosis are improved, and intelligent self-diagnosis is realized.

Description

Technical Field

[0001] The invention relates to cable technology, and in particular to a cable insulation aging defect positioning assessment self-diagnosis method and system. Background Art

[0002] Power cables are an important part of the power system, and their insulation performance is directly related to the safety and reliability of power transmission. As the service time of cables increases, the problem of insulation aging becomes increasingly prominent. The existing methods for detecting and locating cable insulation aging defects have the following main defects:

[0003] Single data collection, incomplete information: Traditional methods usually only use a single detection method, such as dielectric loss measurement or partial discharge detection, which is difficult to fully reflect the multi-faceted characteristics of cable insulation. The limitation of this single data source leads to an incomplete and inaccurate assessment of the insulation status, and it is easy to overlook certain types of insulation defects.

[0004] The data analysis method is simple and lacks deep mining: Existing technologies mostly use simple threshold judgment or statistical analysis methods, which cannot fully utilize the complex relationships and hidden features contained in multidimensional data. This shallow analysis method is difficult to deal with the nonlinear and time-varying characteristics of cable insulation aging, reducing the accuracy of defect identification.

[0005] Lack of intelligence and self-adaptation: Current insulation aging assessment methods mostly rely on fixed diagnostic models and empirical rules, and lack self-learning and self-adaptation capabilities. This results in low reliability and universality of diagnostic results when facing cables of different types and operating environments, making it difficult to meet the needs of increasingly complex power systems for intelligent diagnosis.

[0006] These defects seriously restrict the accurate positioning and reliable assessment of cable insulation aging defects, affecting the preventive maintenance and safe operation of the power system. Therefore, there is an urgent need for a cable insulation aging defect positioning and assessment method that can integrate multi-source data, deeply analyze features, and have intelligent self-diagnosis capabilities. Summary of the invention

[0007] The purpose of the present invention is to solve the problems of the prior art, such as single data collection, incomplete information, simple data analysis method, lack of deep mining, lack of intelligence and self-adaptation, which seriously restrict the precise positioning and reliable evaluation of cable insulation aging defects and affect the preventive maintenance and safe operation of the power system.

[0008] In order to solve the above problems, the present invention provides a cable insulation aging defect location assessment self-diagnosis method, comprising:

[0009] A multi-band broadband dielectric spectrometer is used to perform frequency sweep measurement on the cable under test to obtain the dielectric parameters of the cable at different frequencies; a high-frequency pulse reflection tester is used to perform pulse reflection test on the cable under test to obtain the time domain reflection waveform of the cable; an optical time domain reflectometer is used to perform optical fiber distributed temperature measurement on the cable under test to obtain the temperature distribution data of the entire cable; the obtained dielectric parameters, time domain reflection waveform and temperature distribution data are digitized and noise filtered to form a standardized multidimensional data set;

[0010] Wavelet transform is applied to the standardized multidimensional data set to extract the time-frequency characteristics of the cable dielectric properties. Fourier transform is used to analyze the time-domain reflection waveform to obtain the cable impedance characteristics. Based on the temperature distribution data, the thermal characteristic parameters of the cable are calculated using the heat diffusion model. The extracted time-frequency characteristics, impedance characteristics and thermal characteristic parameters are input into the pre-trained deep convolutional neural network, and a comprehensive feature vector of the cable insulation state is generated through multi-layer feature extraction and fusion.

[0011] The comprehensive feature vector is input into the timing analysis model based on the long short-term memory network to dynamically evaluate the degree of cable insulation aging. The adaptive threshold algorithm is used to detect anomalies in the timing analysis results and identify potential insulation defect locations. The detected anomalies are evaluated in multiple dimensions through the fuzzy inference system, and the severity, development trend and environmental factors of the defects are comprehensively considered to generate accurate positioning results and reliability scores for cable insulation aging defects. The positioning results and scoring information are visualized, and the parameters of the diagnostic model are continuously optimized through a self-learning algorithm.

[0012] In the preferred mode, wavelet transform is applied to the standardized multidimensional data set to extract the time-frequency characteristics of the cable dielectric characteristics; the time domain reflection waveform is analyzed by Fourier transform to obtain the cable impedance characteristics; based on the temperature distribution data, the thermal characteristic parameters of the cable are calculated by using the heat diffusion model, including:

[0013] Wavelet transform is applied to the standardized multidimensional data set to extract the time-frequency characteristics of the cable dielectric properties: Daubechies wavelet is selected as the basis function, and the standardized dielectric parameter data set is decomposed into five layers of wavelets to obtain an approximate coefficient and five detail coefficients; the energy distribution of each layer of wavelet coefficients is calculated; the wavelet entropy index is introduced to quantify the complexity of the cable dielectric properties by calculating the negative logarithm sum of the ratio of each layer of energy to the total energy, which is used as an indicator to judge the degree of insulation aging;

[0014] The cable impedance characteristics are obtained by analyzing the time domain reflection waveform using Fourier transform: the time domain reflection waveform is Fourier transformed to convert it into the frequency domain; the power spectrum density is estimated using the Welch method, which includes dividing the time domain signal into multiple overlapping segments, applying the Hanning window function to each segment of data, calculating the periodogram of each segment of data, and averaging all periodograms to obtain the power spectrum density estimate; the main frequency, three-decibel bandwidth and spectrum slope are extracted from the power spectrum density to form the cable impedance characteristic vector;

[0015] Based on the temperature distribution data, the thermal characteristic parameters of the cable are calculated using the thermal diffusion model: a one-dimensional heat diffusion equation is established and discretized using the finite difference method; the thermal diffusion coefficient is estimated by fitting the measured temperature data and the model prediction value using the least squares method; the concepts of thermal resistance and heat capacity are introduced, and the thermal resistance and heat capacity are calculated based on the thermal diffusion coefficient, cable material density, specific heat capacity and volume per unit length.

[0016] In a preferred manner, the extracted time-frequency features, impedance features and thermal characteristic parameters are input into a pre-trained deep convolutional neural network, and through multi-layer feature extraction and fusion, a comprehensive feature vector of the cable insulation state is generated, including:

[0017] Constructing a deep convolutional neural network with a multi-branch parallel structure, wherein the deep convolutional neural network includes a time-frequency feature branch, an impedance feature branch, and a thermal characteristic parameter branch;

[0018] Inputting the time-frequency feature into the time-frequency feature branch, the time-frequency feature branch includes a first one-dimensional convolution layer, a maximum pooling layer, a second one-dimensional convolution layer and a global average pooling layer connected in sequence; inputting the impedance feature into the impedance feature branch, the impedance feature branch includes a first fully connected layer and a second fully connected layer connected in sequence; inputting the thermal characteristic parameter into the thermal characteristic parameter branch, the structure of the thermal characteristic parameter branch is the same as that of the impedance feature branch;

[0019] The outputs of the time-frequency feature branch, the impedance feature branch and the thermal characteristic parameter branch are connected in a feature fusion layer to generate a fused feature vector; the fused feature vector is input into a fully connected layer sequence, wherein the fully connected layer sequence includes two fully connected layers, each of which is followed by a dropout layer; the outputs of the fully connected layer sequence are connected to an output layer, wherein the output layer is a dense layer with a linear activation function, to generate a comprehensive feature vector of the cable insulation state.

[0020] Preferably, the method further comprises:

[0021] The deep convolutional neural network is pre-trained using an autoencoder structure, the input of the autoencoder structure is the original features, and the output is the reconstructed features; after the pre-training is completed, the encoder part of the autoencoder structure is retained as the initial weight of the deep convolutional neural network;

[0022] The transfer learning method is adopted to freeze the weights of the first few layers of the deep convolutional neural network and only fine-tune the subsequent fully connected layers; the mean square error including the L2 regularization term is used as the loss function, and the coefficient of the L2 regularization term is determined through grid search and cross-validation; the Adam optimizer is used for model training, and a learning rate decay strategy is implemented;

[0023] Add a batch normalization layer after each convolutional layer and fully connected layer; enhance the input data, including random cropping and rotation of time-frequency features, and adding Gaussian noise to impedance features and thermal characteristic parameters;

[0024] After training, the SHAP method is used to analyze the importance of features and calculate the SHAP value of each feature; the trained model is quantized into an 8-bit fixed-point model; a sliding window mechanism is implemented to process the incoming sensor data in real time; preprocessing and preliminary feature extraction are implemented on the sensor node, including simple filtering and Fourier transform.

[0025] In the preferred method, the comprehensive feature vector is input into the time series analysis model based on the long short-term memory network to dynamically evaluate the degree of cable insulation aging; the adaptive threshold algorithm is used to detect anomalies in the time series analysis results to identify potential insulation defect locations including:

[0026] A 64-dimensional comprehensive feature vector sequence generated by a deep convolutional neural network is received, wherein the sequence length is 30, corresponding to 30 days of cable insulation status data; the 64-dimensional comprehensive feature vector sequence is input into the input layer of the long short-term memory network; and sequentially passes through the first long short-term memory layer, the first dropout layer, the second long short-term memory layer, the fully connected layer and the output layer, wherein:

[0027] The first LSTM layer contains 128 hidden units and returns the complete sequence; the first dropout layer has a dropout rate of 0.3; the second LSTM layer contains 64 hidden units and returns only the output of the last time step; the fully connected layer contains 32 neurons and uses the ReLU activation function; the output layer contains 1 neuron and uses the Sigmoid activation function; an attention mechanism is introduced after the LSTM layer, and the attention score is calculated through a learnable weight matrix and bias vector, and a context vector is generated;

[0028] Binary cross entropy is used as the loss function, and L2 regularization term is introduced, where the regularization coefficient is determined by cross-validation; Adam optimizer is used for model training, the initial learning rate is set to 0.001, and the learning rate decay strategy is implemented; the early stopping strategy is used to stop training when there is no improvement in 5 consecutive training cycles on the validation set; the training data is enhanced by adding Gaussian noise and random time offset; the sliding window method is used to predict the continuous comprehensive feature vector sequence to obtain a continuous aging degree time series;

[0029] Initialize the adaptive threshold to the mean of the aging degree sequence plus two times the standard deviation; use a sliding window of size 7 days to traverse the aging degree sequence, and calculate the local mean, local standard deviation and local trend for each window; update the adaptive threshold based on local statistics, with the smoothing factor set to 0.1 and the sensitivity parameter set to 2.5; if the aging degree of the last point in the window exceeds the updated adaptive threshold and the local trend is positive, mark the point as a potential defect point;

[0030] Perform DBSCAN clustering on the marked potential defect points, and classify points with close distances as the same defect; use three sliding windows of different sizes, 3 days, 7 days, and 14 days, to perform multi-scale analysis and integrate the detection results of multiple scales; assign a confidence score to each potential defect point based on the degree and duration of exceeding the threshold; compare the current detection results with historical data to identify new or accelerated defects; consider the impact of environmental factors such as temperature and humidity on the degree of aging, obtain the environmental impact function through regression analysis, and correct the original degree of aging;

[0031] The detected potential defect locations are combined with the cable wiring diagram to generate a visual report containing a trend chart of the overall cable aging degree, a heat map of the potential defect locations, detailed defect information and a priority list of repair recommendations.

[0032] In the preferred mode, the detected abnormal points are evaluated in multiple dimensions through the fuzzy inference system, and the severity, development trend and environmental factors of the defects are comprehensively considered to generate the precise positioning results and reliability scores of the cable insulation aging defects, including:

[0033] Receive cable insulation aging abnormal point data detected by an adaptive threshold algorithm;

[0034] Construct a fuzzy inference system, including the following input variables:

[0035] The severity of defects is divided into three fuzzy sets: low, medium, and high; the development trend is divided into three fuzzy sets: slow, stable, and accelerated; the ambient temperature is divided into three fuzzy sets: low temperature, normal temperature, and high temperature; the ambient humidity is divided into three fuzzy sets: dry, normal, and humid;

[0036] Design output variables:

[0037] Defect reliability scores are divided into five fuzzy sets: low, medium-low, medium, medium-high, and high; position correction coefficients are divided into four fuzzy sets: fine adjustment, minor adjustment, medium adjustment, and major adjustment; a fuzzy rule base is defined, containing at least 50 rules, covering various input variable combinations;

[0038] Select the Mamdani reasoning method and use the minimum-maximum synthesis rule for fuzzy reasoning; for each detected abnormal point, calculate the fuzzy membership according to its characteristic value: use the Gaussian membership function to calculate the membership of the severity of the defect; use the triangle membership function to calculate the membership of the development trend; use the trapezoidal membership function to calculate the membership of the ambient temperature and humidity; apply fuzzy rules for reasoning and obtain the fuzzy set of the output variable;

[0039] The output fuzzy set is defuzzified using the centroid method to obtain an accurate defect reliability score and position correction coefficient; the original abnormal point position is adjusted based on the position correction coefficient to obtain an accurate positioning result; a defect feature vector is established, including the defect reliability score, accurate position, severity, development trend and environmental factors; a self-organizing map neural network is used to perform cluster analysis on the defect feature vector to identify similar defect patterns; based on the clustering results, a combination analysis is performed on defects of the same category to evaluate their impact on the overall performance of the cable; a defect impact propagation model is constructed to simulate the diffusion process of defects in the cable:

[0040] Use the heat conduction equation to describe the diffusion of defects; consider the cable material characteristics and structural factors; set boundary conditions and initial conditions; use the finite element method to solve the defect impact propagation model to obtain the defect impact range and severity distribution; based on the defect impact propagation results, update the defect reliability score and precise location; establish a multi-layer perceptron neural network, use the updated defect characteristics as input, and predict the defect development trend and remaining service life;

[0041] The Monte Carlo simulation method is used to generate a variety of possible defect development scenarios and evaluate the uncertainty of the prediction results. A defect risk assessment report is generated by comprehensively considering the defect reliability score, precise location, impact range, development trend and remaining life prediction. Based on the risk assessment results, a cable maintenance priority list and specific maintenance recommendations are formulated. The assessment results are compared with historical data, and the fuzzy rule base and model parameters are continuously optimized to improve the system's adaptability and assessment accuracy.

[0042] Cable insulation aging defect location assessment self-diagnosis system, including:

[0043] The first unit is used to perform frequency sweep measurement on the cable under test using a multi-band broadband dielectric spectrometer to obtain the dielectric parameters of the cable at different frequencies; perform pulse reflection test on the cable under test using a high-frequency pulse reflection tester to obtain the time domain reflection waveform of the cable; perform optical fiber distributed temperature measurement on the cable under test using an optical time domain reflectometer to obtain the temperature distribution data of the entire cable; digitize and filter the obtained dielectric parameters, time domain reflection waveform and temperature distribution data to form a standardized multi-dimensional data set;

[0044] The second unit is used to apply wavelet transform to the standardized multidimensional data set to extract the time-frequency characteristics of the cable dielectric properties; use Fourier transform to analyze the time-domain reflection waveform to obtain the cable impedance characteristics; based on the temperature distribution data, the thermal diffusion model is used to calculate the thermal characteristic parameters of the cable; the extracted time-frequency characteristics, impedance characteristics and thermal characteristic parameters are input into the pre-trained deep convolutional neural network, and through multi-layer feature extraction and fusion, a comprehensive feature vector of the cable insulation state is generated;

[0045] The third unit is used to input the comprehensive feature vector into the timing analysis model based on the long short-term memory network to dynamically evaluate the degree of cable insulation aging; use the adaptive threshold algorithm to detect anomalies in the timing analysis results and identify potential insulation defect locations; use the fuzzy reasoning system to conduct a multi-dimensional evaluation of the detected anomalies, comprehensively consider the severity of the defects, development trends and environmental factors, and generate accurate positioning results and reliability scores for cable insulation aging defects; visualize the positioning results and scoring information, and continuously optimize the parameters of the diagnostic model through a self-learning algorithm.

[0046] An electronic device, comprising:

[0047] processor;

[0048] a memory for storing processor-executable instructions;

[0049] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0050] A computer-readable storage medium stores computer program instructions, which implement the aforementioned method when executed by a processor.

[0051] Beneficial effects of the present invention:

[0052] Improved the comprehensiveness and accuracy of cable insulation aging defect diagnosis:

[0053] The present invention uses a variety of detection methods such as a multi-band broadband dielectric spectrometer, a high-frequency pulse reflection tester, and an optical time-domain reflectometer to obtain the dielectric parameters, time-domain reflection waveform, and temperature distribution data of the cable to form a multidimensional data set. Through a variety of analysis methods such as wavelet transform, Fourier transform, and thermal diffusion model, the time-frequency characteristics, impedance characteristics, and thermal characteristic parameters of the cable are fully extracted. This method of multi-source data fusion and multi-dimensional feature extraction overcomes the limitations of traditional single detection methods and greatly improves the comprehensiveness and accuracy of insulation aging defect diagnosis.

[0054] Realizes intelligent self-diagnosis of cable insulation aging defects:

[0055] The present invention adopts a deep learning model combining a deep convolutional neural network and a long short-term memory network to perform comprehensive feature extraction and time series analysis on the cable insulation status. Through an adaptive threshold algorithm and a fuzzy inference system, intelligent identification, precise positioning and reliability scoring of insulation defects are achieved. This intelligent self-diagnosis method effectively overcomes the limitations of traditional empirical rules and fixed models, and improves the reliability and adaptability of diagnostic results.

[0056] Possessing self-learning capability to continuously improve diagnostic performance:

[0057] The present invention continuously optimizes the parameters of the diagnostic model through a self-learning algorithm, so that the system can continuously improve its performance based on new data and diagnostic results. This adaptive learning mechanism enables the diagnostic system to adapt to different types of cables and changing operating environments, and continuously improves the accuracy and universality of defect identification. At the same time, the visual presentation of diagnostic results also facilitates intuitive understanding and decision-making by operation and maintenance personnel, improving the efficiency and pertinence of preventive maintenance of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A schematic diagram of a process of a cable insulation aging defect location assessment self-diagnosis method according to an embodiment of the present invention;

[0059] Figure 2 It is a structural schematic diagram of a cable insulation aging defect location assessment self-diagnosis system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0062] Figure 1 FIG. 1 is a flow chart of a method for locating and evaluating cable insulation aging defects according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0063] S101. Use a multi-band broadband dielectric spectrometer to perform frequency sweep measurement on the cable under test to obtain the dielectric parameters of the cable at different frequencies; use a high-frequency pulse reflection tester to perform pulse reflection test on the cable under test to obtain the time domain reflection waveform of the cable; use an optical time domain reflectometer to perform optical fiber distributed temperature measurement on the cable under test to obtain the temperature distribution data of the entire cable; digitize and filter the obtained dielectric parameters, time domain reflection waveform and temperature distribution data to form a standardized multidimensional data set;

[0064] S102. Apply wavelet transform to the standardized multidimensional data set to extract the time-frequency characteristics of the cable dielectric characteristics; use Fourier transform to analyze the time-domain reflection waveform to obtain the cable impedance characteristics; based on the temperature distribution data, use the heat diffusion model to calculate the thermal characteristic parameters of the cable; input the extracted time-frequency characteristics, impedance characteristics and thermal characteristic parameters into the pre-trained deep convolutional neural network, and generate a comprehensive feature vector of the cable insulation state through multi-layer feature extraction and fusion;

[0065] S103. Input the comprehensive feature vector into the timing analysis model based on the long short-term memory network to dynamically evaluate the degree of cable insulation aging; use the adaptive threshold algorithm to detect anomalies in the timing analysis results and identify potential insulation defect locations; use the fuzzy reasoning system to perform multi-dimensional evaluation on the detected anomalies, comprehensively consider the severity of the defects, development trends and environmental factors, and generate accurate positioning results and reliability scores for cable insulation aging defects; visualize the positioning results and scoring information, and continuously optimize the parameters of the diagnostic model through the self-learning algorithm.

[0066] The specific implementation method of the cable insulation aging defect location assessment self-diagnosis method is as follows:

[0067] First, a multi-band broadband dielectric spectrometer is used to perform frequency sweep measurement on the cable to be tested, and the dielectric parameters of the cable at different frequencies are obtained. Specifically, a dielectric spectrometer with a frequency range of 100Hz-1MHz is used to select 50 frequency points in a logarithmic interval to obtain the complex dielectric constant ε*=ε'-jε". Among them, ε' is the real part of the dielectric constant, which characterizes the energy storage characteristics of the cable; ε" is the imaginary part, which characterizes the loss characteristics of the cable. For a section of 1km long cross-linked polyethylene insulated power cable, ε'=2.3, ε"=0.0023 were measured at a frequency of 100Hz, while ε'=2.2, ε"=0.022 were measured at a frequency of 1MHz, showing typical frequency dispersion characteristics.

[0068] Next, the high-frequency pulse reflection tester is used to perform a pulse reflection test on the cable to obtain the time domain reflection waveform of the cable. Specifically, a Gaussian pulse signal with a rise time of 2ns and an amplitude of 5V is used as the excitation, the sampling rate is 10GS / s, and the recording length is 100μs. For the above 1km long cable, an obvious reflection wave is observed at 500m from the beginning of the cable, and the reflection coefficient is about 0.1, indicating that there is an impedance discontinuity at this location.

[0069] At the same time, an optical time domain reflectometer was used to measure the fiber-optic distributed temperature of the cable under test to obtain the temperature distribution data of the entire cable. A 1550nm wavelength laser was used as the light source, with a spatial resolution of 1m, a temperature resolution of 0.1℃, and a measurement time of 10min. The measurement results showed that the temperature along the cable fluctuated between 20-25℃, but a hot spot of 27℃ appeared 600m away from the starting point, which may be related to local overheating.

[0070] The acquired dielectric parameters, time-domain reflection waveforms, and temperature distribution data are digitized and noise-filtered to form a standardized multidimensional data set. The dielectric parameters are logarithmically transformed to make their distribution more uniform; wavelet denoising is applied to the time-domain reflection waveform to reduce the impact of random noise; and moving average filtering is performed on the temperature data to eliminate mutation points. The processed data are uniformly normalized to the [0,1] interval for subsequent analysis.

[0071] Wavelet transform is applied to the standardized multidimensional data set to extract the time-frequency characteristics of the cable dielectric properties. The db4 wavelet basis is selected and a 5-layer decomposition is performed to obtain the multi-scale characteristics of the dielectric parameters varying with frequency. For the above cables, obvious mutations are observed in the high-frequency scale coefficients, indicating an abnormal increase in high-frequency dielectric loss.

[0072] The time domain reflection waveform is analyzed by Fourier transform to obtain the cable impedance characteristics. The amplitude spectrum and phase spectrum of the reflection waveform are calculated to extract the characteristic frequency components. The results show that a resonance peak appears around 300MHz, which corresponds to the local resonance phenomenon in the cable, which may be related to insulation defects.

[0073] Based on the temperature distribution data, the thermal diffusion model is used to calculate the thermal characteristic parameters of the cable. Assuming that the cable is a cylindrical structure, a one-dimensional radial heat conduction equation is established, and the temperature field distribution is solved by the finite difference method. The calculated equivalent thermal conductivity is 0.3W / (m·K) and the thermal time constant is 30min, which is lower than that of normal cables, indicating that the heat dissipation performance is reduced.

[0074] The extracted time-frequency features, impedance features and thermal characteristic parameters are input into the pre-trained deep convolutional neural network, and the comprehensive feature vector of the cable insulation status is generated through multi-layer feature extraction and fusion. The network structure contains 3 convolutional layers, 2 pooling layers and 2 fully connected layers, using the ReLU activation function and dropout to prevent overfitting. Finally, a 256-dimensional feature vector is obtained, which encodes the key information of the cable insulation status.

[0075] The comprehensive feature vector is input into the time series analysis model based on the long short-term memory network to dynamically evaluate the degree of cable insulation aging. The LSTM network contains 64 hidden units, and the activation functions of the input gate, forget gate, and output gate are all sigmoid functions. By learning from historical data, the model can capture the trend of insulation performance over time and predict future development. For the above cable, the model predicts that its insulation aging index will increase by 20% in the next 3 months.

[0076] The adaptive threshold algorithm is used to detect anomalies in the timing analysis results and identify potential insulation defect locations. The local anomaly factor (LOF) algorithm based on sliding windows is adopted, with a window size of 20 and an initial anomaly threshold of 2.5. Through iterative optimization, the final anomaly threshold is 2.8, and three potential defect points are detected, which are located at 300m, 500m and 600m from the starting point.

[0077] The detected abnormal points are evaluated in multiple dimensions through the fuzzy reasoning system, and the severity, development trend and environmental factors of the defects are comprehensively considered to generate the precise positioning results and reliability scores of the cable insulation aging defects. The fuzzy rule base contains 20 if-then rules, with the input variables being the defect amplitude, change rate and ambient temperature, and the output variable being the defect reliability score. After fuzzy reasoning, the defect scores at 300m are 0.6, 500m are 0.8, and 600m are 0.9, all exceeding the warning threshold of 0.5 and requiring special attention.

[0078] Finally, the positioning results and scoring information are visualized, and the parameters of the diagnostic model are continuously optimized through a self-learning algorithm. A heat map is used to display the insulation status of the entire length of the cable, and different colors are used to mark the defect location and severity. At the same time, the incremental learning method is used to dynamically update the model parameters based on the newly collected data to improve the diagnostic accuracy. After three months of continuous optimization, the average diagnostic accuracy of the model has increased from 85% to 92%.

[0079] In an optional implementation, wavelet transform is applied to the standardized multidimensional data set to extract the time-frequency characteristics of the cable dielectric characteristics; Fourier transform is used to analyze the time domain reflection waveform to obtain the cable impedance characteristics; based on the temperature distribution data, the thermal characteristic parameters of the cable are calculated using a heat diffusion model, including:

[0080] Wavelet transform is applied to the standardized multidimensional data set to extract the time-frequency characteristics of the cable dielectric properties: Daubechies wavelet is selected as the basis function, and the standardized dielectric parameter data set is decomposed into five layers of wavelets to obtain an approximate coefficient and five detail coefficients; the energy distribution of each layer of wavelet coefficients is calculated; the wavelet entropy index is introduced to quantify the complexity of the cable dielectric properties by calculating the negative logarithm sum of the ratio of each layer of energy to the total energy, which is used as an indicator to judge the degree of insulation aging;

[0081] The cable impedance characteristics are obtained by analyzing the time domain reflection waveform using Fourier transform: the time domain reflection waveform is Fourier transformed to convert it into the frequency domain; the power spectrum density is estimated using the Welch method, which includes dividing the time domain signal into multiple overlapping segments, applying the Hanning window function to each segment of data, calculating the periodogram of each segment of data, and averaging all periodograms to obtain the power spectrum density estimate; the main frequency, three-decibel bandwidth and spectrum slope are extracted from the power spectrum density to form the cable impedance characteristic vector;

[0082] Based on the temperature distribution data, the thermal characteristic parameters of the cable are calculated using the thermal diffusion model: a one-dimensional heat diffusion equation is established and discretized using the finite difference method; the thermal diffusion coefficient is estimated by fitting the measured temperature data and the model prediction value using the least squares method; the concepts of thermal resistance and heat capacity are introduced, and the thermal resistance and heat capacity are calculated based on the thermal diffusion coefficient, cable material density, specific heat capacity and volume per unit length.

[0083] In a specific implementation, the acquired multi-dimensional cable data set is first standardized to make data of different dimensions comparable. The standardization method uses Z-score standardization, that is, subtracting the mean of each feature and dividing it by the standard deviation. For example, for the dielectric constant data, the mean is calculated to be 2.3 and the standard deviation is 0.2, then the standardized data is (original value - 2.3) / 0.2.

[0084] Next, wavelet transform is applied to the standardized multidimensional data set to extract the time-frequency characteristics of the cable dielectric properties. Daubechies 4th-order wavelet (db4) is selected as the basis function, and the standardized dielectric parameter data is subjected to a 5-layer wavelet decomposition. Taking the dielectric loss tangent data with a sampling frequency of 1MHz and a data length of 1024 points as an example, after 5-layer wavelet decomposition, one approximate coefficient (length is 32) and five detail coefficients (lengths are 32, 64, 128, 256, and 512, respectively) are obtained.

[0085] Calculate the energy distribution of each layer of wavelet coefficients, that is, the sum of the squares of each coefficient. For example, the energy of the detail coefficient of the 5th layer is 0.35, the 4th layer is 0.28, the 3rd layer is 0.20, the 2nd layer is 0.12, the 1st layer is 0.03, and the approximate coefficient energy is 0.02. The wavelet entropy index is introduced to quantify the complexity of the cable dielectric properties, and the negative logarithm sum of the ratio of each layer's energy to the total energy is calculated. In this example, the wavelet entropy is 1.37, which can be used as an indicator to judge the degree of insulation aging.

[0086] Then, the time domain reflection waveform is analyzed by Fourier transform to obtain the cable impedance characteristics. The time domain reflection waveform with a sampling rate of 500MHz and a length of 10,000 points is subjected to fast Fourier transform (FFT) to obtain the frequency domain representation. The power spectral density is estimated by the Welch method, and the time domain signal is divided into 8 segments with 50% overlap, each with a length of 2048 points. The Hanning window function is applied to each segment of data, and the periodogram is calculated and averaged to obtain the power spectral density estimate.

[0087] Extract characteristic parameters from the power spectrum density: the main frequency is the frequency corresponding to the maximum value of the power spectrum density, which is 15MHz in this case; the decibel bandwidth is the frequency range when the power spectrum density drops by 3dB, which is 10-20MHz in this case; the spectrum tilt is the linear fitting slope of the power spectrum density in the frequency range of 20-100MHz, which is -0.2dB / MHz in this case. These three parameters constitute the cable impedance characteristic vector.

[0088] Finally, the thermal characteristic parameters of the cable are calculated based on the temperature distribution data using the heat diffusion model. A one-dimensional heat diffusion equation is established, and the central difference format is used for spatial discretization, and the forward Euler format is used for temporal discretization. The spatial step size is 0.01m, and the time step size is 0.1s. Taking the case of a cable with a length of 10m, an initial temperature of 20℃, and one end heated to 80℃ as an example, the measured temperature data and the model prediction value are fitted by the least squares method, and the thermal diffusion coefficient is estimated to be 1.2×10^-7m^2 / s.

[0089] The concepts of thermal resistance and thermal capacity are introduced. Based on the thermal diffusion coefficient, cable material density (1200kg / m^3), specific heat capacity (1000J / (kg·K)) and unit length volume (7.85×10^-5m^3 / m), the thermal resistance per unit length is calculated to be 2.54K·m / W and the heat capacity per unit length is 94.2J / (m·K). These thermal characteristic parameters can be used to evaluate the heat dissipation performance and temperature rise characteristics of the cable.

[0090] Through the above steps, a comprehensive analysis of the cable's multi-dimensional data is achieved, and key parameters such as dielectric properties, impedance characteristics and thermal characteristics are extracted, providing an important basis for cable status assessment and life prediction.

[0091] In an optional embodiment, the extracted time-frequency features, impedance features and thermal characteristic parameters are input into a pre-trained deep convolutional neural network, and a comprehensive feature vector of the cable insulation state is generated through multi-layer feature extraction and fusion, including:

[0092] Constructing a deep convolutional neural network with a multi-branch parallel structure, wherein the deep convolutional neural network includes a time-frequency feature branch, an impedance feature branch, and a thermal characteristic parameter branch;

[0093] Inputting the time-frequency feature into the time-frequency feature branch, the time-frequency feature branch includes a first one-dimensional convolution layer, a maximum pooling layer, a second one-dimensional convolution layer and a global average pooling layer connected in sequence; inputting the impedance feature into the impedance feature branch, the impedance feature branch includes a first fully connected layer and a second fully connected layer connected in sequence; inputting the thermal characteristic parameter into the thermal characteristic parameter branch, the structure of the thermal characteristic parameter branch is the same as that of the impedance feature branch;

[0094] The outputs of the time-frequency feature branch, the impedance feature branch and the thermal characteristic parameter branch are connected in a feature fusion layer to generate a fused feature vector; the fused feature vector is input into a fully connected layer sequence, wherein the fully connected layer sequence includes two fully connected layers, each of which is followed by a dropout layer; the outputs of the fully connected layer sequence are connected to an output layer, wherein the output layer is a dense layer with a linear activation function, to generate a comprehensive feature vector of the cable insulation state.

[0095] In a specific implementation, the extracted time-frequency features, impedance features and thermal characteristic parameters are input into a pre-trained deep convolutional neural network, and the detailed steps of generating a comprehensive feature vector of the cable insulation state through multi-layer feature extraction and fusion are as follows:

[0096] Firstly, a multi-branch parallel deep convolutional neural network is constructed. The network includes three parallel branches: time-frequency feature branch, impedance feature branch and thermal characteristic parameter branch.

[0097] The time-frequency feature branch is used to process time-frequency features, including the first one-dimensional convolution layer, the maximum pooling layer, the second one-dimensional convolution layer, and the global average pooling layer connected in sequence. The convolution kernel size of the first one-dimensional convolution layer is 3, the number of convolution kernels is 32, the step length is 1, and the ReLU activation function is used. The pooling window size of the maximum pooling layer is 2, and the step length is 2. The convolution kernel size of the second one-dimensional convolution layer is 3, the number of convolution kernels is 64, the step length is 1, and the ReLU activation function is also used. The global average pooling layer performs average pooling on the entire feature map and outputs a fixed-size feature vector.

[0098] The impedance feature branch is used to process the impedance feature, including the first fully connected layer and the second fully connected layer connected in sequence. The number of neurons in the first fully connected layer is 64, and the ReLU activation function is used. The number of neurons in the second fully connected layer is 32, and the ReLU activation function is also used.

[0099] The structure of the thermal characteristic parameter branch is the same as that of the impedance characteristic branch, and also includes two fully connected layers for processing thermal characteristic parameters.

[0100] Next, the extracted time-frequency features are input into the time-frequency feature branch, the impedance features are input into the impedance feature branch, and the thermal characteristic parameters are input into the thermal characteristic parameter branch. For example, suppose the extracted time-frequency feature dimension is (100, 1), the impedance feature dimension is (10, 1), and the thermal characteristic parameter dimension is (5, 1).

[0101] Then, the outputs of the three branches are connected in the feature fusion layer to generate a fused feature vector. Assuming that the time-frequency feature branch outputs 64-dimensional features, the impedance feature branch outputs 32-dimensional features, and the thermal characteristic parameter branch outputs 32-dimensional features, the dimension of the fused feature vector is 128.

[0102] The fused feature vector is input into a sequence of fully connected layers, which includes two fully connected layers, each followed by a dropout layer. The number of neurons in the first fully connected layer is 64, the ReLU activation function is used, and the dropout rate is 0.5. The number of neurons in the second fully connected layer is 32, and the ReLU activation function is also used, and the dropout rate is 0.5.

[0103] Finally, the output of the fully connected layer sequence is connected to the output layer. The output layer is a dense layer with a linear activation function and 16 neurons, generating a 16-dimensional comprehensive feature vector of the cable insulation status.

[0104] Through the above steps, the deep convolutional neural network can effectively extract and fuse the key information in the time-frequency features, impedance features and thermal characteristic parameters, and generate a comprehensive feature vector that can fully characterize the insulation state of the cable. This multi-branch parallel structure network design can make full use of the complementarity of different types of features and improve the feature expression ability and classification performance.

[0105] In practical applications, the network structure and parameters can be adjusted appropriately according to the specific data characteristics and task requirements. For example, the number of convolutional layers and fully connected layers can be increased or decreased, the size and number of convolution kernels can be adjusted, and the pooling strategy can be changed. At the same time, batch normalization, residual connection and other technologies can be used to further improve network performance. Through repeated training and optimization, a deep learning model that can accurately evaluate the insulation status of cables can be obtained.

[0106] In an optional implementation, the method further includes:

[0107] The deep convolutional neural network is pre-trained using an autoencoder structure, the input of the autoencoder structure is the original features, and the output is the reconstructed features; after the pre-training is completed, the encoder part of the autoencoder structure is retained as the initial weight of the deep convolutional neural network;

[0108] The transfer learning method is adopted to freeze the weights of the first few layers of the deep convolutional neural network and only fine-tune the subsequent fully connected layers; the mean square error including the L2 regularization term is used as the loss function, and the coefficient of the L2 regularization term is determined through grid search and cross-validation; the Adam optimizer is used for model training, and a learning rate decay strategy is implemented;

[0109] Add a batch normalization layer after each convolutional layer and fully connected layer; enhance the input data, including random cropping and rotation of time-frequency features, and adding Gaussian noise to impedance features and thermal characteristic parameters;

[0110] After training, the SHAP method is used to analyze the importance of features and calculate the SHAP value of each feature; the trained model is quantized into an 8-bit fixed-point model; a sliding window mechanism is implemented to process the incoming sensor data in real time; preprocessing and preliminary feature extraction are implemented on the sensor node, including simple filtering and Fourier transform.

[0111] In this embodiment, the deep convolutional neural network is pre-trained and fine-tuned to improve the model performance. First, the deep convolutional neural network is pre-trained using an autoencoder structure. The input of the autoencoder is the original features, including time-frequency features, impedance features, and thermal characteristic parameters. Through multi-layer convolution and pooling operations, the input features are encoded into a low-dimensional representation, and then the low-dimensional representation is reconstructed into the original feature space through deconvolution operations. The goal of pre-training is to minimize the reconstruction error between the input and the reconstructed output. After the pre-training is completed, the encoder part of the autoencoder is retained as the initial weight of the deep convolutional neural network.

[0112] Next, the transfer learning method is used to fine-tune the model. The weights of the first few convolutional layers of the network are frozen, and only the subsequent fully connected layers are fine-tuned. This is because the first few convolutional layers usually extract low-level general features, while the later layers extract more task-related high-level features. During the fine-tuning process, the mean square error containing the L2 regularization term is used as the loss function. The L2 regularization term can prevent the model from overfitting, and its coefficient is determined by grid search and cross-validation. Specifically, the candidate values ​​of the L2 regularization coefficient are set to [0.001, 0.01, 0.1, 1], and a 5-fold cross-validation is performed on each candidate value to select the coefficient value with the best performance on the validation set. The Adam optimizer is used for model training, and the initial learning rate is set to 0.001. The learning rate is reduced to 0.1 times the original every 50 epochs, and a total of 200 epochs are trained.

[0113] In order to further improve the generalization ability of the model, a batch normalization layer is added after each convolutional layer and fully connected layer. Batch normalization can accelerate model convergence and play a certain regularization role. In addition, the input data is enhanced, including random cropping and rotation of time-frequency features, and adding Gaussian noise to impedance features and thermal characteristic parameters. Specifically, the 80%-100% area of ​​the time-frequency features is randomly cropped, and the rotation angle is randomly selected between -10° and 10°. Gaussian noise with a mean of 0 and a standard deviation of 0.01 is added to the impedance features and thermal characteristic parameters. These data enhancement techniques can improve the robustness of the model to input changes.

[0114] After the model training is completed, the SHAP (SHapley Additive exPlanations) method is used to analyze the importance of features. The SHAP method is based on the Shapley value in game theory and can explain the contribution of each feature to the model prediction. Specifically, 1,000 samples are randomly selected and the SHAP value of each feature is calculated. The larger the absolute value of the SHAP value, the greater the impact of the feature on the prediction result. Based on the analysis results, the most important features can be identified, such as the energy in certain specific frequency bands, the real and imaginary parts of the impedance, etc.

[0115] In order to deploy the model on resource-constrained embedded devices, the trained floating-point model is quantized to an 8-bit fixed-point model. The quantization process involves determining the dynamic range of each layer and mapping the weights and activation values ​​to an integer range of [-128, 127]. The size of the quantized model can be reduced to 1 / 4 of the original, and the inference speed is also improved, while ensuring that the accuracy loss is within an acceptable range (usually no more than 1%).

[0116] In practical applications, a sliding window mechanism is implemented to process incoming sensor data in real time. The window size is set to 1 second and the step size is 0.1 second. For the data in each time window, features are extracted and input into the model for prediction. This mechanism can realize real-time analysis of continuous data streams.

[0117] In order to reduce the computational burden of the central processing unit, preprocessing and preliminary feature extraction are implemented on the sensor nodes. Specifically, a simple median filter is performed on the sensor node to remove outliers. Then the filtered data is fast Fourier transformed to calculate the energy characteristics of different frequency bands. These preliminary extracted features can greatly reduce the amount of data that needs to be transmitted to the central processing unit, thereby reducing communication overhead and energy consumption.

[0118] Through the above steps, an efficient and robust deep learning model is implemented to analyze and predict sensor data. The model combines multiple technologies such as pre-training, transfer learning, regularization, and data enhancement to achieve good performance with limited training data. At the same time, through model quantization and distributed processing, the model can run efficiently on resource-constrained edge devices.

[0119] In an optional implementation, the comprehensive feature vector is input into a time series analysis model based on a long short-term memory network to dynamically evaluate the degree of cable insulation aging; an adaptive threshold algorithm is used to detect anomalies in the time series analysis results to identify potential insulation defect locations including:

[0120] A 64-dimensional comprehensive feature vector sequence generated by a deep convolutional neural network is received, wherein the sequence length is 30, corresponding to 30 days of cable insulation status data; the 64-dimensional comprehensive feature vector sequence is input into the input layer of the long short-term memory network; and sequentially passes through the first long short-term memory layer, the first dropout layer, the second long short-term memory layer, the fully connected layer and the output layer, wherein:

[0121] The first LSTM layer contains 128 hidden units and returns the complete sequence; the first dropout layer has a dropout rate of 0.3; the second LSTM layer contains 64 hidden units and returns only the output of the last time step; the fully connected layer contains 32 neurons and uses the ReLU activation function; the output layer contains 1 neuron and uses the Sigmoid activation function; an attention mechanism is introduced after the LSTM layer, and the attention score is calculated through a learnable weight matrix and bias vector, and a context vector is generated;

[0122] Binary cross entropy is used as the loss function, and L2 regularization term is introduced, where the regularization coefficient is determined by cross-validation; Adam optimizer is used for model training, the initial learning rate is set to 0.001, and the learning rate decay strategy is implemented; the early stopping strategy is used to stop training when there is no improvement in 5 consecutive training cycles on the validation set; the training data is enhanced by adding Gaussian noise and random time offset; the sliding window method is used to predict the continuous comprehensive feature vector sequence to obtain a continuous aging degree time series;

[0123] Initialize the adaptive threshold to the mean of the aging degree sequence plus two times the standard deviation; use a sliding window of size 7 days to traverse the aging degree sequence, and calculate the local mean, local standard deviation and local trend for each window; update the adaptive threshold based on local statistics, with the smoothing factor set to 0.1 and the sensitivity parameter set to 2.5; if the aging degree of the last point in the window exceeds the updated adaptive threshold and the local trend is positive, mark the point as a potential defect point;

[0124] Perform DBSCAN clustering on the marked potential defect points, and classify points with close distances as the same defect; use three sliding windows of different sizes, 3 days, 7 days, and 14 days, to perform multi-scale analysis and integrate the detection results of multiple scales; assign a confidence score to each potential defect point based on the degree and duration of exceeding the threshold; compare the current detection results with historical data to identify new or accelerated defects; consider the impact of environmental factors such as temperature and humidity on the degree of aging, obtain the environmental impact function through regression analysis, and correct the original degree of aging;

[0125] The detected potential defect locations are combined with the cable wiring diagram to generate a visual report containing a trend chart of the overall cable aging degree, a heat map of the potential defect locations, detailed defect information and a priority list of repair recommendations.

[0126] In a specific implementation, a 64-dimensional comprehensive feature vector sequence generated by a deep convolutional neural network is first received. The length of the sequence is 30, corresponding to 30 days of cable insulation status data. This 64-dimensional comprehensive feature vector sequence is input into the input layer of the long short-term memory network to start time series analysis.

[0127] The structure of the LSTM network includes the following layers: the first LSTM layer contains 128 hidden units and returns the complete sequence; the first dropout layer has a dropout rate of 0.3; the second LSTM layer contains 64 hidden units and only returns the output of the last time step; the fully connected layer contains 32 neurons and uses the ReLU activation function; the output layer contains 1 neuron and uses the Sigmoid activation function. The attention mechanism is introduced after the LSTM layer, and the attention score is calculated through the learnable weight matrix and bias vector, and the context vector is generated.

[0128] In terms of model training, binary cross entropy is used as the loss function, and the L2 regularization term is introduced, where the regularization coefficient is determined by cross-validation. The Adam optimizer is used for model training, the initial learning rate is set to 0.001, and the learning rate decay strategy is implemented. The early stopping strategy is used to stop training when there is no improvement in 5 consecutive training cycles on the validation set. The training data is enhanced by adding Gaussian noise and random time offset to improve the generalization ability of the model.

[0129] The sliding window method is used to predict the continuous comprehensive feature vector sequence to obtain a continuous aging degree time series. For example, for a certain section of cable, the 30-day prediction result may be: [0.12, 0.15, 0.18, 0.22, 0.25, 0.28, 0.32, 0.35, 0.39, 0.42, 0.46, 0.49, 0.53, 0.56, 0.60, 0.63, 0.67, 0.70, 0.74, 0.77, 0.81, 0.84, 0.88, 0.91, 0.95, 0.98, 1.02, 1.05, 1.09, 1.12]. This sequence represents the trend of cable insulation aging over time.

[0130] Next, anomaly detection is performed. First, the adaptive threshold is initialized to the mean of the aging degree sequence plus two times the standard deviation. Taking the above sequence as an example, the initial threshold is about 1.02. Then a sliding window of size 7 days is used to traverse the aging degree sequence, and the local mean, local standard deviation and local trend are calculated for each window. The adaptive threshold is updated based on these local statistics, with the smoothing factor set to 0.1 and the sensitivity parameter set to 2.5. If the aging degree of the last point in the window exceeds the updated adaptive threshold and the local trend is positive, the point is marked as a potential defect point.

[0131] DBSCAN clustering is performed on the marked potential defect points, and points with close distances are classified as the same defect. At the same time, three sliding windows of different sizes, 3 days, 7 days, and 14 days, are used for multi-scale analysis to integrate the detection results of multiple scales. A confidence score is assigned to each potential defect point based on the degree and duration of exceeding the threshold. For example, if the aging degree of a point is 1.15 and it exceeds the threshold for 3 days, it may be assigned a confidence score of 0.8.

[0132] Compare current test results with historical data to identify new or accelerating defects. Consider the impact of environmental factors such as temperature and humidity on the degree of aging, obtain the environmental impact function through regression analysis, and correct the original degree of aging. For example, in a high temperature and high humidity environment, it may be necessary to multiply the original degree of aging by a correction factor of 1.2.

[0133] Finally, the detected potential defect locations are combined with the cable wiring diagram to generate a visual report. The report includes a trend chart of the overall cable aging degree, showing the changing trend of the aging degree within 30 days; a heat map of the potential defect location, marking the severity of the defect with different colors; detailed defect information, listing the location, aging degree, confidence level, etc. of each defect; and a maintenance recommendation priority list, sorted by the severity and development speed of the defect. Such a visual report can intuitively display the cable insulation status and provide decision support for maintenance personnel.

[0134] In an optional implementation, a fuzzy inference system is used to perform a multi-dimensional evaluation of the detected abnormal points, comprehensively considering the severity of the defect, development trend and environmental factors, and generating an accurate positioning result and reliability score of the cable insulation aging defect, including:

[0135] Receive cable insulation aging abnormal point data detected by an adaptive threshold algorithm;

[0136] Construct a fuzzy inference system, including the following input variables:

[0137] The severity of defects is divided into three fuzzy sets: low, medium, and high; the development trend is divided into three fuzzy sets: slow, stable, and accelerated; the ambient temperature is divided into three fuzzy sets: low temperature, normal temperature, and high temperature; the ambient humidity is divided into three fuzzy sets: dry, normal, and humid;

[0138] Design output variables:

[0139] Defect reliability scores are divided into five fuzzy sets: low, medium-low, medium, medium-high, and high; position correction coefficients are divided into four fuzzy sets: fine adjustment, minor adjustment, medium adjustment, and major adjustment; a fuzzy rule base is defined, containing at least 50 rules, covering various input variable combinations;

[0140] Select the Mamdani reasoning method and use the minimum-maximum synthesis rule for fuzzy reasoning; for each detected abnormal point, calculate the fuzzy membership according to its characteristic value: use the Gaussian membership function to calculate the membership of the severity of the defect; use the triangle membership function to calculate the membership of the development trend; use the trapezoidal membership function to calculate the membership of the ambient temperature and humidity; apply fuzzy rules for reasoning and obtain the fuzzy set of the output variable;

[0141] The output fuzzy set is defuzzified using the centroid method to obtain an accurate defect reliability score and position correction coefficient; the original abnormal point position is adjusted based on the position correction coefficient to obtain an accurate positioning result; a defect feature vector is established, including the defect reliability score, accurate position, severity, development trend and environmental factors; a self-organizing map neural network is used to perform cluster analysis on the defect feature vector to identify similar defect patterns; based on the clustering results, a combination analysis is performed on defects of the same category to evaluate their impact on the overall performance of the cable; a defect impact propagation model is constructed to simulate the diffusion process of defects in the cable:

[0142] Use the heat conduction equation to describe the diffusion of defects; consider the cable material characteristics and structural factors; set boundary conditions and initial conditions; use the finite element method to solve the defect impact propagation model to obtain the defect impact range and severity distribution; based on the defect impact propagation results, update the defect reliability score and precise location; establish a multi-layer perceptron neural network, use the updated defect characteristics as input, and predict the defect development trend and remaining service life;

[0143] The Monte Carlo simulation method is used to generate a variety of possible defect development scenarios and evaluate the uncertainty of the prediction results. A defect risk assessment report is generated by comprehensively considering the defect reliability score, precise location, impact range, development trend and remaining life prediction. Based on the risk assessment results, a cable maintenance priority list and specific maintenance recommendations are formulated. The assessment results are compared with historical data, and the fuzzy rule base and model parameters are continuously optimized to improve the system's adaptability and assessment accuracy.

[0144] In order to accurately locate and evaluate the reliability of cable insulation aging defects, the cable insulation aging abnormal point data detected by the adaptive threshold algorithm are first received. These data include the initial position coordinates of the abnormal point, the detection time, and the relevant parameter values ​​reflecting the defect characteristics.

[0145] Next, a fuzzy inference system is constructed to perform multi-dimensional evaluation of the detected abnormal points. The input variables of the system include defect severity, development trend, ambient temperature and ambient humidity. The defect severity is divided into three fuzzy sets: low, medium and high, which can be represented by a value of 0-10, 0-3 for low, 4-7 for medium, and 8-10 for high. The development trend is divided into three fuzzy sets: slow, stable and accelerated, which can be represented by the rate of change of the abnormal point parameter value. For example, a monthly change rate of less than 5% is slow, 5%-15% is stable, and greater than 15% is accelerated. The ambient temperature is divided into three fuzzy sets: low temperature, normal temperature and high temperature, corresponding to 0-15℃, 15-30℃ and 30-50℃ respectively. The ambient humidity is divided into three fuzzy sets: dry, normal and humid, corresponding to relative humidity of 0-40%, 40%-70% and 70%-100% respectively.

[0146] The output variables of the system include defect reliability score and position correction coefficient. The defect reliability score is divided into five fuzzy sets: low, medium-low, medium, medium-high, and high, expressed by a score of 0-100, 0-20 for low, 20-40 for medium-low, 40-60 for medium, 60-80 for medium-high, and 80-100 for high. The position correction coefficient is divided into four fuzzy sets: fine adjustment, minor adjustment, medium adjustment, and major adjustment, expressed by an adjustment range of ±0-5cm, ±0-1cm for fine adjustment, ±1-2cm for minor adjustment, ±2-3cm for medium adjustment, and ±3-5cm for major adjustment.

[0147] Then define a fuzzy rule base containing at least 50 rules covering various combinations of input variables. For example, "If the defect severity is high, the development trend is accelerated, the ambient temperature is high, and the ambient humidity is humid, then the defect reliability score is low and the position correction factor is large." Select the Mamdani reasoning method and use the minimum-maximum synthesis rule for fuzzy reasoning.

[0148] For each detected abnormal point, the fuzzy membership is calculated according to its characteristic value. The membership of the defect severity is calculated using the Gaussian membership function, with the center values ​​of 1.5, 5.5, and 9.5, and the standard deviation is 1. The membership of the development trend is calculated using the triangular membership function, with the vertices of 2.5%, 10%, and 20%. The membership of the ambient temperature and humidity is calculated using the trapezoidal membership function, with the four inflection points of temperature being 0℃, 10℃, 20℃, and 30℃, and the four inflection points of humidity being 20%, 35%, 55%, and 75%.

[0149] Fuzzy rules are applied for reasoning to obtain the fuzzy set of the output variable. The centroid method is used to defuzzify the output fuzzy set, obtaining the accurate defect reliability score and the position correction coefficient. For example, the defect reliability score of a certain abnormal point is 75 points, and the position correction coefficient is +2.5 cm. The original abnormal point position is adjusted based on the position correction coefficient to obtain the precise positioning result.

[0150] Next, a defect feature vector is established, including the defect reliability score, precise position, severity, development trend, and environmental factors. The self-organizing map neural network is used to perform clustering analysis on the defect feature vector to identify similar defect patterns. The number of nodes in the input layer of the network is the same as the dimension of the feature vector, and the output layer is a 10×10 two-dimensional grid. The competitive learning algorithm is used to train the network, with the initial learning rate being 0.1 and gradually decreasing with the number of training rounds.

[0151] Based on the clustering results, combined analysis is performed on defects of the same category to evaluate their impact on the overall performance of the cable. For example, it is found that multiple adjacent high-severity defects are clustered in a certain area, which may lead to a significant decrease in the insulation performance of this area.

[0152] A defect impact propagation model is constructed to simulate the diffusion process of defects in the cable. The heat conduction equation is used to describe the defect diffusion, considering the cable material properties and structural factors. The boundary condition is set as the heat exchange between the outer surface of the cable and the environment, and the initial condition is the position and severity of the known defect. The finite element method is used to solve the defect impact propagation model. The cable is divided into 1000 grid cells, the time step is set to 1 hour, and the diffusion process for 100 days is simulated. The defect impact range and severity distribution are obtained.

[0153] Based on the defect impact propagation results, the defect reliability score and precise positioning are updated. For example, if the impact range of a certain defect expands to 5 cm around it and the severity increases by 20%, then its reliability score is reduced by 10 points, and the position range expands to ±2.5 cm.

[0154] A multi-layer perceptron neural network is established, taking the updated defect features as the input to predict the defect development trend and remaining service life. The network structure consists of an input layer (the number of nodes is the same as the dimension of the feature vector), two hidden layers (each layer has 20 nodes), and an output layer (2 nodes, representing the development trend and remaining life respectively). The backpropagation algorithm is used to train the network, with the learning rate being 0.01, the momentum factor being 0.9, and the number of training rounds being 10000.

[0155] Using the Monte Carlo simulation method, a variety of possible defect development scenarios are generated to evaluate the uncertainty of the prediction results. 1,000 simulations are performed, and the input characteristic value is randomly adjusted by ±10% each time to obtain the distribution range of the prediction results. For example, the remaining service life prediction result of a defect is 3-5 years, and the development trend is stable to slowly accelerating.

[0156] A defect risk assessment report is generated by comprehensively considering the defect reliability score, precise location, impact range, development trend and remaining life prediction. The report contains detailed information for each defect and an analysis of its impact on the overall performance of the cable. Based on the risk assessment results, a cable maintenance priority list and specific maintenance recommendations are developed. For example, for defects with a reliability score of less than 50 points and a remaining life of less than 2 years, it is recommended to prioritize local repair or replacement.

[0157] Finally, the evaluation results are compared with historical data, and the fuzzy rule base and model parameters are continuously optimized. For example, if the actual development speed of a certain type of defect is 50% faster than the predicted value, the corresponding fuzzy rules are adjusted to increase the weight of the development trend. By continuously accumulating data and optimizing models, the system's adaptive ability and evaluation accuracy are improved.

[0158] Figure 2 FIG. 1 is a schematic diagram of the structure of a cable insulation aging defect location assessment self-diagnosis system according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0159] The first unit is used to perform frequency sweep measurement on the cable under test using a multi-band broadband dielectric spectrometer to obtain the dielectric parameters of the cable at different frequencies; perform pulse reflection test on the cable under test using a high-frequency pulse reflection tester to obtain the time domain reflection waveform of the cable; perform optical fiber distributed temperature measurement on the cable under test using an optical time domain reflectometer to obtain the temperature distribution data of the entire cable; digitize and filter the obtained dielectric parameters, time domain reflection waveform and temperature distribution data to form a standardized multi-dimensional data set;

[0160] The second unit is used to apply wavelet transform to the standardized multidimensional data set to extract the time-frequency characteristics of the cable dielectric properties; use Fourier transform to analyze the time-domain reflection waveform to obtain the cable impedance characteristics; based on the temperature distribution data, the thermal diffusion model is used to calculate the thermal characteristic parameters of the cable; the extracted time-frequency characteristics, impedance characteristics and thermal characteristic parameters are input into the pre-trained deep convolutional neural network, and through multi-layer feature extraction and fusion, a comprehensive feature vector of the cable insulation state is generated;

[0161] The third unit is used to input the comprehensive feature vector into the timing analysis model based on the long short-term memory network to dynamically evaluate the degree of cable insulation aging; use the adaptive threshold algorithm to detect anomalies in the timing analysis results and identify potential insulation defect locations; use the fuzzy reasoning system to conduct a multi-dimensional evaluation of the detected anomalies, comprehensively consider the severity of the defects, development trends and environmental factors, and generate accurate positioning results and reliability scores for cable insulation aging defects; visualize the positioning results and scoring information, and continuously optimize the parameters of the diagnostic model through a self-learning algorithm.

[0162] According to a third aspect of the embodiments of the present invention,

[0163] An electronic device is provided, comprising:

[0164] processor;

[0165] a memory for storing processor-executable instructions;

[0166] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0167] According to a fourth aspect of the embodiments of the present invention,

[0168] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0169] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0170] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A self-diagnosis method for locating and evaluating cable insulation aging defects, characterized in that: include: A multi-band broadband dielectric spectrometer is used to perform frequency sweep measurement on the cable to be tested, and the dielectric parameters of the cable at different frequencies are obtained; The high-frequency pulse reflection tester is used to perform pulse reflection test on the cable to be tested, and the time domain reflection waveform of the cable is obtained; the optical time domain reflectometer is used to perform optical fiber distributed temperature measurement on the cable to be tested, and the temperature distribution data of the entire cable is obtained; the obtained dielectric parameters, time domain reflection waveform and temperature distribution data are digitized and noise filtered to form a standardized multi-dimensional data set; Apply wavelet transform to the standardized multidimensional data set to extract the time-frequency characteristics of the cable dielectric properties; use Fourier transform to analyze the time-domain reflection waveform to obtain the cable impedance characteristics; Based on the temperature distribution data, the thermal characteristic parameters of the cable are calculated using the heat diffusion model. The extracted time-frequency features, impedance features and thermal characteristic parameters are input into the pre-trained deep convolutional neural network, and a comprehensive feature vector of the cable insulation state is generated through multi-layer feature extraction and fusion. The comprehensive feature vector is input into the timing analysis model based on the long short-term memory network to dynamically evaluate the degree of cable insulation aging. The adaptive threshold algorithm is used to detect anomalies in the timing analysis results and identify potential insulation defect locations. The detected anomalies are evaluated in multiple dimensions through the fuzzy inference system, and the severity, development trend and environmental factors of the defects are comprehensively considered to generate accurate positioning results and reliability scores for cable insulation aging defects. The positioning results and scoring information are visualized, and the parameters of the diagnostic model are continuously optimized through a self-learning algorithm.

2. The cable insulation aging defect location assessment self-diagnosis method according to claim 1 is characterized in that: Wavelet transform is applied to the standardized multidimensional data set to extract the time-frequency characteristics of the cable dielectric properties; Use Fourier transform to analyze the time domain reflection waveform and obtain the cable impedance characteristics; Based on the temperature distribution data, the thermal characteristic parameters of the cable are calculated using the heat diffusion model, including: Wavelet transform is applied to the standardized multidimensional data set to extract the time-frequency characteristics of the cable dielectric properties: Daubechies wavelet is selected as the basis function, and the standardized dielectric parameter data set is decomposed into five layers of wavelets to obtain an approximate coefficient and five detail coefficients; the energy distribution of each layer of wavelet coefficients is calculated; the wavelet entropy index is introduced to quantify the complexity of the cable dielectric properties by calculating the negative logarithm sum of the ratio of each layer of energy to the total energy, which is used as an indicator to judge the degree of insulation aging; The cable impedance characteristics are obtained by analyzing the time domain reflection waveform using Fourier transform: the time domain reflection waveform is Fourier transformed to convert it into the frequency domain; the power spectrum density is estimated using the Welch method, which includes dividing the time domain signal into multiple overlapping segments, applying the Hanning window function to each segment of data, calculating the periodogram of each segment of data, and averaging all periodograms to obtain the power spectrum density estimate; the main frequency, three-decibel bandwidth and spectrum slope are extracted from the power spectrum density to form the cable impedance characteristic vector; Based on the temperature distribution data, the thermal characteristic parameters of the cable are calculated using the thermal diffusion model: a one-dimensional heat diffusion equation is established and discretized using the finite difference method; the thermal diffusion coefficient is estimated by fitting the measured temperature data and the model prediction value using the least squares method; the concepts of thermal resistance and heat capacity are introduced, and the thermal resistance and heat capacity are calculated based on the thermal diffusion coefficient, cable material density, specific heat capacity and volume per unit length.

3. The cable insulation aging defect location assessment self-diagnosis method according to claim 1, characterized in that: The extracted time-frequency features, impedance features and thermal characteristic parameters are input into the pre-trained deep convolutional neural network. Through multi-layer feature extraction and fusion, a comprehensive feature vector of the cable insulation status is generated, including: Constructing a deep convolutional neural network with a multi-branch parallel structure, wherein the deep convolutional neural network includes a time-frequency feature branch, an impedance feature branch, and a thermal characteristic parameter branch; Inputting the time-frequency feature into the time-frequency feature branch, the time-frequency feature branch includes a first one-dimensional convolution layer, a maximum pooling layer, a second one-dimensional convolution layer and a global average pooling layer connected in sequence; inputting the impedance feature into the impedance feature branch, the impedance feature branch includes a first fully connected layer and a second fully connected layer connected in sequence; inputting the thermal characteristic parameter into the thermal characteristic parameter branch, the structure of the thermal characteristic parameter branch is the same as that of the impedance feature branch; The outputs of the time-frequency feature branch, the impedance feature branch and the thermal characteristic parameter branch are connected in a feature fusion layer to generate a fused feature vector; the fused feature vector is input into a fully connected layer sequence, wherein the fully connected layer sequence includes two fully connected layers, each of which is followed by a dropout layer; the outputs of the fully connected layer sequence are connected to an output layer, wherein the output layer is a dense layer with a linear activation function, to generate a comprehensive feature vector of the cable insulation state.

4. The cable insulation aging defect location assessment self-diagnosis method according to claim 3 is characterized in that: The method further comprises: The deep convolutional neural network is pre-trained using an autoencoder structure, the input of the autoencoder structure is the original features, and the output is the reconstructed features; after the pre-training is completed, the encoder part of the autoencoder structure is retained as the initial weight of the deep convolutional neural network; The transfer learning method is adopted to freeze the weights of the first few layers of the deep convolutional neural network and only fine-tune the subsequent fully connected layers; the mean square error including the L2 regularization term is used as the loss function, and the coefficient of the L2 regularization term is determined through grid search and cross-validation; the Adam optimizer is used for model training, and a learning rate decay strategy is implemented; Add a batch normalization layer after each convolutional layer and fully connected layer; enhance the input data, including random cropping and rotation of time-frequency features, and adding Gaussian noise to impedance features and thermal characteristic parameters; After training, the SHAP method is used to analyze the importance of features and calculate the SHAP value of each feature; the trained model is quantized into an 8-bit fixed-point model; a sliding window mechanism is implemented to process the incoming sensor data in real time; preprocessing and preliminary feature extraction are implemented on the sensor node, including simple filtering and Fourier transform.

5. The cable insulation aging defect location assessment self-diagnosis method according to claim 1, characterized in that: The comprehensive feature vector is input into the time series analysis model based on long short-term memory network to dynamically evaluate the degree of cable insulation aging; An adaptive threshold algorithm is used to detect anomalies in the timing analysis results and identify potential insulation defect locations including: A 64-dimensional comprehensive feature vector sequence generated by a deep convolutional neural network is received, wherein the sequence length is 30, corresponding to 30 days of cable insulation status data; the 64-dimensional comprehensive feature vector sequence is input into the input layer of the long short-term memory network; and sequentially passes through the first long short-term memory layer, the first dropout layer, the second long short-term memory layer, the fully connected layer and the output layer, wherein: The first LSTM layer contains 128 hidden units and returns the complete sequence; the first dropout layer has a dropout rate of 0.3; the second LSTM layer contains 64 hidden units and returns only the output of the last time step; the fully connected layer contains 32 neurons and uses the ReLU activation function; the output layer contains 1 neuron and uses the Sigmoid activation function; an attention mechanism is introduced after the LSTM layer, and the attention score is calculated through a learnable weight matrix and bias vector, and a context vector is generated; Binary cross entropy is used as the loss function, and L2 regularization term is introduced, where the regularization coefficient is determined by cross-validation; Adam optimizer is used for model training, the initial learning rate is set to 0.001, and the learning rate decay strategy is implemented; the early stopping strategy is used to stop training when there is no improvement in 5 consecutive training cycles on the validation set; the training data is enhanced by adding Gaussian noise and random time offset; the sliding window method is used to predict the continuous comprehensive feature vector sequence to obtain a continuous aging degree time series; Initialize the adaptive threshold to the mean of the aging degree sequence plus two times the standard deviation; use a sliding window of size 7 days to traverse the aging degree sequence, and calculate the local mean, local standard deviation and local trend for each window; update the adaptive threshold based on local statistics, with the smoothing factor set to 0.1 and the sensitivity parameter set to 2.5; if the aging degree of the last point in the window exceeds the updated adaptive threshold and the local trend is positive, mark the point as a potential defect point; Perform DBSCAN clustering on the marked potential defect points, and classify points with close distances as the same defect; use three sliding windows of different sizes, 3 days, 7 days, and 14 days, to perform multi-scale analysis and integrate the detection results of multiple scales; assign a confidence score to each potential defect point based on the degree and duration of exceeding the threshold; compare the current detection results with historical data to identify new or accelerated defects; consider the impact of environmental factors such as temperature and humidity on the degree of aging, obtain the environmental impact function through regression analysis, and correct the original degree of aging; The detected potential defect locations are combined with the cable wiring diagram to generate a visual report containing a trend chart of the overall cable aging degree, a heat map of the potential defect locations, detailed defect information and a priority list of repair recommendations.

6. The cable insulation aging defect location assessment self-diagnosis method according to claim 1, characterized in that: The fuzzy inference system is used to conduct a multi-dimensional evaluation of the detected abnormal points, comprehensively consider the severity of the defects, development trends and environmental factors, and generate accurate positioning results and reliability scores for cable insulation aging defects, including: Receive cable insulation aging abnormal point data detected by an adaptive threshold algorithm; Construct a fuzzy inference system, including the following input variables: The severity of defects is divided into three fuzzy sets: low, medium, and high; the development trend is divided into three fuzzy sets: slow, stable, and accelerated; the ambient temperature is divided into three fuzzy sets: low temperature, normal temperature, and high temperature; the ambient humidity is divided into three fuzzy sets: dry, normal, and humid; Design output variables: Defect reliability scores are divided into five fuzzy sets: low, medium-low, medium, medium-high, and high; position correction coefficients are divided into four fuzzy sets: fine adjustment, minor adjustment, medium adjustment, and major adjustment; a fuzzy rule base is defined, containing at least 50 rules, covering various input variable combinations; Select the Mamdani reasoning method and use the minimum-maximum synthesis rule for fuzzy reasoning; for each detected abnormal point, calculate the fuzzy membership according to its characteristic value: use the Gaussian membership function to calculate the membership of the severity of the defect; use the triangle membership function to calculate the membership of the development trend; use the trapezoidal membership function to calculate the membership of the ambient temperature and humidity; apply fuzzy rules for reasoning and obtain the fuzzy set of the output variable; The output fuzzy set is defuzzified using the centroid method to obtain an accurate defect reliability score and position correction coefficient; the original abnormal point position is adjusted based on the position correction coefficient to obtain an accurate positioning result; a defect feature vector is established, including the defect reliability score, accurate position, severity, development trend and environmental factors; a self-organizing map neural network is used to perform cluster analysis on the defect feature vector to identify similar defect patterns; based on the clustering results, a combination analysis is performed on defects of the same category to evaluate their impact on the overall performance of the cable; a defect impact propagation model is constructed to simulate the diffusion process of defects in the cable: Use the heat conduction equation to describe the diffusion of defects; consider the cable material characteristics and structural factors; set boundary conditions and initial conditions; use the finite element method to solve the defect impact propagation model to obtain the defect impact range and severity distribution; based on the defect impact propagation results, update the defect reliability score and precise location; establish a multi-layer perceptron neural network, use the updated defect characteristics as input, and predict the defect development trend and remaining service life; The Monte Carlo simulation method is used to generate a variety of possible defect development scenarios and evaluate the uncertainty of the prediction results. A defect risk assessment report is generated by comprehensively considering the defect reliability score, precise location, impact range, development trend and remaining life prediction. Based on the risk assessment results, a cable maintenance priority list and specific maintenance recommendations are formulated. The assessment results are compared with historical data, and the fuzzy rule base and model parameters are continuously optimized to improve the system's adaptability and assessment accuracy.

7. A cable insulation aging defect location assessment self-diagnosis system, used to implement the method described in any one of claims 1 to 6, characterized in that it comprises: The first unit is used to perform frequency sweep measurement on the cable under test using a multi-band broadband dielectric spectrometer to obtain the dielectric parameters of the cable at different frequencies; perform pulse reflection test on the cable under test using a high-frequency pulse reflection tester to obtain the time domain reflection waveform of the cable; perform optical fiber distributed temperature measurement on the cable under test using an optical time domain reflectometer to obtain the temperature distribution data of the entire cable; digitize and filter the obtained dielectric parameters, time domain reflection waveform and temperature distribution data to form a standardized multi-dimensional data set; The second unit is used to apply wavelet transform to the standardized multidimensional data set to extract the time-frequency characteristics of the cable dielectric properties; use Fourier transform to analyze the time-domain reflection waveform to obtain the cable impedance characteristics; based on the temperature distribution data, the thermal diffusion model is used to calculate the thermal characteristic parameters of the cable; the extracted time-frequency characteristics, impedance characteristics and thermal characteristic parameters are input into the pre-trained deep convolutional neural network, and through multi-layer feature extraction and fusion, a comprehensive feature vector of the cable insulation state is generated; The third unit is used to input the comprehensive feature vector into the timing analysis model based on the long short-term memory network to dynamically evaluate the degree of cable insulation aging; use the adaptive threshold algorithm to detect anomalies in the timing analysis results and identify potential insulation defect locations; use the fuzzy reasoning system to conduct a multi-dimensional evaluation of the detected anomalies, comprehensively consider the severity of the defects, development trends and environmental factors, and generate accurate positioning results and reliability scores for cable insulation aging defects; visualize the positioning results and scoring information, and continuously optimize the parameters of the diagnostic model through a self-learning algorithm.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images