Urban underground comprehensive pipe gallery power cable operation state monitoring method
By combining one-dimensional convolutional neural networks and generative adversarial networks, the problem of difficult monitoring of cable operation status in underground utility tunnels has been solved, achieving efficient and accurate cable status detection and reducing accident risks.
Patent Information
- Application Number
- CN202511787775.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-01
- Publication Date
- 2026-03-27
AI Technical Summary
The operation status of cables in underground utility tunnels is difficult to monitor in real time, resulting in a high risk of accidents. Existing technologies have low monitoring efficiency and insufficient safety control capabilities.
A one-dimensional convolutional neural network (1D-CNN) combined with a generative adversarial network (GAN) is used to monitor the partial discharge signal of the cable. High-frequency pulse current signals are collected by sensors to generate partial discharge map (TRPD) data. Data augmentation and normalization are then performed to construct a 1D-CNN model for monitoring the cable's operating status.
It enables rapid and accurate monitoring of cable operating status, reduces the risk of model overfitting, improves recognition accuracy and efficiency, and ensures the safe operation of cables.
Smart Images
Figure CN121746769A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for monitoring the operational status of power cables in urban underground utility tunnels, belonging to the field of power technology. Background Technology
[0002] Urban integrated utility tunnels, also known as integrated utility trenches or underground common trenches, are a type of modern, scientific, and intensive urban infrastructure formed by centrally installing two or more pipelines, such as electricity, communications, water supply, hot water, cooling, greywater, and gas, in the same underground space below the road. Integrated utility tunnels solve the problems of monitoring, maintenance, and expansion of various pipelines during urban development, and are a hot topic in urban modernization and one of the indicators of a city's level of modernization.
[0003] The biggest advantage of integrated utility tunnels is that they allow for the construction and maintenance of underground pipelines without disturbing roads. This means that most of the maintenance work is done underground. The main and branch lines within underground integrated utility tunnels are distributed crisscrossed over long distances, and personnel are sparsely available.
[0004] Most of the work is carried out underground. The underground utility tunnels have a long network of main and branch lines, are sparsely populated, poorly lit, and poorly ventilated. If an accident occurs in the underground pipelines, the safety of personnel and equipment will be at great risk.
[0005] Therefore, it is necessary to conduct real-time monitoring and safety protection of the professional pipelines in underground utility tunnels, improve the rapid response and safety control capabilities of urban underground utility tunnel operation and management, and ensure the safety of underground utility tunnels. Summary of the Invention
[0006] This invention designs and develops a method for monitoring the operating status of power cables in urban underground utility tunnels. It uses a one-dimensional convolutional neural network to monitor the cable operating status, thereby improving monitoring efficiency.
[0007] The technical solution provided by this invention is as follows:
[0008] A method for monitoring the operational status of power cables in urban underground utility tunnels, comprising:
[0009] Step 1: In the selected integrated utility tunnel, use sensors to collect high-frequency pulse current signals of partial discharge during cable operation, and combine them with the corresponding time-series information to generate partial discharge map TRPD data.
[0010] Step 2: Perform one-dimensional and normalization processing on TRPD data based on 1D-CNN convolutional network, divide the data into training and test sets, perform data augmentation on the training set sample data, and expand the initial training set.
[0011] Step 3: Build a 1D-CNN model and train the 1D-CNN network based on the data-augmented training set;
[0012] Step 4: Input the normalized test data into the 1D-CNN network model to monitor the cable operation status and output the monitoring results;
[0013] The formula for calculating the 1D-CNN convolution is as follows:
[0014]
[0015] In the formula, f is the input sequence, g is the convolution kernel, * is the convolution operation, t is the position in the output sequence, k is the size of the convolution kernel, and n-1 is the upper limit of the convolution kernel.
[0016] Preferably,
[0017] The structure of the 1D-CNN model includes: an input layer, four convolutional layers, four pooling layers, a fully connected layer, and an output layer. The input layer contains 1×2116 partial discharge TRPD waveform data. The convolutional kernel size is 1×9 with a compensation of 1. The pooling kernel size is 1×4 with a stride of 1. The last pooling layer is connected to the weighted connection layer. Data filtering and feature extraction are performed between layers through the ReLU nonlinear activation function.
[0018] Preferably, in step two, 75% of the data is used as the training set and 25% of the data is used as the test set.
[0019] Preferably, in step two, GAN-based data augmentation technology is used to expand the training samples for partial discharge identification. A noise signal z ~ N(0,1) is used as the input to the generator G, and a pseudo-partial discharge signal is synthesized to deceive the discriminator D. The discriminator D provides a prediction about this partial discharge signal. After multiple rounds of sufficient training, a Nash equilibrium state is reached, including:
[0020] Step a: The generator network G maps a one-dimensional noise vector to a feature vector of shape with an input length of 1, and mixes the artificially generated spectrum with the real TRPD spectrum of partial discharge;
[0021] Step b: The discriminator network D maps a TRPD spectrum of shape with an input length of 1 to a binary fraction, which is used to evaluate the probability that the production data is true;
[0022] Step c: The GAN network connects network G and network D together. The generator decodes the latent space vector into feature vectors. The discriminator evaluates the authenticity of these feature parameters. The latent vector is mapped to the discriminator's evaluation result and output through the GAN network.
[0023] Step d: Train the discriminator using real and fake TRPD spectral samples labeled "real" or "fake";
[0024] Step e: Train the generator based on the gradient of the loss of the GAN model relative to the generator weights. The direction of weight update is to maximize the probability that the discriminator will classify the decoded TRPD spectrogram as "true".
[0025] Preferably, in step three, after determining the network architecture, the CNN needs to be trained, including two processes: forward propagation and backward propagation. The forward propagation process is as follows:
[0026] O p =F n (...F2(F1(X P W (1) W (2) W (N) );
[0027] In the formula, W represents the network weight, and X represents the weight of the network. P Let F be the input feature and F be the activation function.
[0028] Backpropagation requires continuous optimization of the weights W and biases b. The update formulas for weights W and biases b are as follows:
[0029]
[0030] In the formula, η is the loss function, l is the number of layers, and J is the loss function.
[0031] Preferably, step four includes:
[0032] Step 1: Construct a one-dimensional convolutional neural network model, adjust the network depth and convolutional kernel size, and determine the final network structure;
[0033] Step 2: Compile network parameters, use classification cross-entropy as the loss function, and use gradient descent algorithm as the optimizer;
[0034] Step 3: Train the 1D-CNN network using the data-augmented training samples, use the classification cross-entropy loss function to measure the distance between the predicted value and the true value, calculate the gradient of the loss with respect to the network parameters, and optimize the network parameters based on the SGD algorithm.
[0035] Preferably, the one-dimensional time-domain waveform feature vector of TRPD is normalized, and the normalization function is:
[0036]
[0037] In the formula, x min x is the minimum value of the waveform eigenvector. maxThis represents the maximum value of the waveform eigenvector.
[0038] The beneficial effects of this invention are as follows: The method for monitoring the operating status of power cables in urban underground utility tunnels provided by this invention can collect and identify the discharge signals of cables in underground utility tunnels, detect cable abnormalities in a timely manner, and realize the detection of cable operating status.
[0039] Based on GAN data augmentation technology, the risk of model overfitting under small sample sizes can be effectively reduced, and the accuracy of discharge signal identification can be improved. Based on one-dimensional convolutional neural networks, the cumbersome steps of manually extracting statistical feature parameters are eliminated, making it more suitable for engineering practice. 1D-CNN can quickly extract one-dimensional feature parameters of TRPD, effectively improving the online identification of partial discharge in cables and enhancing the accuracy and efficiency of cable operation status monitoring. Attached Figure Description
[0040] Figure 1 This is a flowchart illustrating the method for monitoring the operational status of power cables in urban underground utility tunnels as described in this invention.
[0041] Figure 2 This is the partial discharge pattern recognition process based on 1D-CNN described in this invention.
[0042] Figure 3 This is a comparison chart showing the impact of sample expansion on recognition performance as described in this invention.
[0043] Figure 4 This is a comparative diagram of the identification results of various defect types using 1D-CNN as described in this invention. Detailed Implementation
[0044] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.
[0045] like Figure 1-4 As shown, this invention provides a method for monitoring the operational status of power cables in urban underground utility tunnels. 1D-CNN can quickly extract one-dimensional feature parameters of TRPD (Partial Discharge), effectively improving the online identification of partial discharge. The method includes:
[0046] Step 1: In the selected integrated utility tunnel, use sensors to collect high-frequency pulse current signals of partial discharge during cable operation, and combine them with the corresponding time-series information to generate partial discharge map TRPD data.
[0047] In this invention, as a preferred embodiment, a high-frequency current sensor is used to detect the high-frequency pulse current signal of partial discharge;
[0048] Step 2: Perform one-dimensional and normalization processing on TRPD data based on 1D-CNN convolutional network, divide the data into training and test sets, perform data augmentation on the training set sample data, and expand the initial training set.
[0049] Using a noise signal z ~ N(0,1) as input to the generator G, a pseudo partial discharge signal is synthesized to deceive the discriminator D. The discriminator D provides a prediction about this partial discharge signal. After multiple rounds of sufficient training, a Nash equilibrium state is reached, including:
[0050] Step a: The generator network G maps a one-dimensional noise vector to a feature vector of shape with an input length of 1, and mixes the artificially generated spectrum with the real TRPD spectrum of partial discharge;
[0051] Step b: The discriminator network D maps a TRPD spectrum of shape with an input length of 1 to a binary fraction, which is used to evaluate the probability that the production data is true;
[0052] Step c: The GAN network connects network G and network D together. The generator decodes the latent space vector into feature vectors. The discriminator evaluates the authenticity of these feature parameters. The latent vector is mapped to the discriminator's evaluation result and output through the GAN network.
[0053] Step d: Train the discriminator using real and fake TRPD spectral samples labeled "real" or "fake";
[0054] Step e: Train the generator based on the gradient of the loss of the GAN model relative to the generator weights. The direction of weight update is to maximize the probability that the discriminator will classify the decoded TRPD spectrogram as "true".
[0055] The one-dimensional time-domain waveform eigenvector of TRPD is normalized using the following normalization function:
[0056]
[0057] In the formula, x min x is the minimum value of the waveform eigenvector. max This represents the maximum value of the waveform eigenvector.
[0058] Use 75% of the data as the training set and 25% as the test set.
[0059] Step 3: Build a 1D-CNN model and train the 1D-CNN network based on the data-augmented training set;
[0060] In this invention, as a preferred embodiment, the structure of the 1D-CNN model includes: an input layer, four convolutional layers, four pooling layers, a fully connected layer, and an output layer. The input layer receives partial discharge TRPD waveform data with a size of 1×2^116. The convolutional kernel size is 1×9 with a compensation of 1, the pooling kernel size is 1×4 with a stride of 1, and the last pooling layer is connected to the weighted connected layer. Data filtering and feature extraction are performed between layers using the ReLU nonlinear activation function. The output layer obtains the final classification result through the Softmax function.
[0061] Table 1 1D-CNN Model Structure Settings
[0062]
[0063] After determining the network architecture, the CNN needs to be trained, which includes two processes: forward propagation and back propagation. The forward propagation process is as follows:
[0064] O p =F n (...F2(F1(X P W (1) W (2) W (N) );
[0065] In the formula, W is the network weight, X is the input feature, and F is the activation function;
[0066] Backpropagation requires continuous optimization of the weights W and biases b. The update formulas for weights W and biases b are as follows:
[0067]
[0068] In the formula, η is the loss function and J is the loss function.
[0069] Step 4: Input the normalized test data into the 1D-CNN network model to monitor the cable's operating status and output the monitoring results, including:
[0070] Step 1: Construct a one-dimensional convolutional neural network model, adjust the network depth and convolutional kernel size, and determine the final network structure;
[0071] Step 2: Compile network parameters, use classification cross-entropy as the loss function, and use gradient descent algorithm as the optimizer;
[0072] Step 3: Train the 1D-CNN network using the data-augmented training samples, use the classification cross-entropy loss function to measure the distance between the predicted value and the true value, calculate the gradient of the loss with respect to the network parameters, and optimize the network parameters based on the SGD algorithm.
[0073] Step 4: Use the trained 1D-CNN model for pattern recognition on the test set samples.
[0074] The formula for calculating 1D-CNN convolution is:
[0075]
[0076] In the formula, f is the input sequence, g is the convolution kernel, * is the convolution operation, t is the position in the output sequence, k is the size of the convolution kernel, and n-1 is the upper limit of the convolution kernel.
[0077] Example 1
[0078] 1D-CNN Pattern Recognition Method Based on Data Augmentation Techniques
[0079] Python was chosen as the computer programming language, Tensorflow as the programming learning framework, and the computer processor was an i5.
[0080] The partial discharge TRPD spectrum was obtained based on HFCT online detection technology, and the sample was expanded by data augmentation technology. Finally, a 1D-CNN model was used to perform pattern recognition on four typical discharge types: corona discharge, internal discharge, suspension discharge, and surface discharge.
[0081] Through a partial discharge simulation experiment on a cable, 4000 original partial discharge samples were obtained and divided into training and test sets. The training set had 3000 samples and the test set had 1000 samples. The training sample set was expanded based on GAN data augmentation technology, and the hyperparameters of the 1D-CNN model were set to the optimal parameters with a learning rate of 0.0001.
[0082] like Figure 3 As shown, from Figure 3 As can be seen, GAN-based data augmentation techniques can effectively improve the recognition accuracy of 1D-CNN models. When the training set is the original sample set, the pattern recognition rate is 97.5%; when the training sample set is expanded by 3000 samples, the recognition rate reaches 97.9%; when the training sample set is expanded by 6000 samples, the recognition rate further improves to 98.8%; however, when the training sample set is expanded by 9000 samples, the recognition rate drops to 98.6%. Simultaneously, sample expansion exponentially increases training time: with 3000 samples, the training time is 662.2 seconds, while with 9000 samples, the training time reaches 1179.9 seconds.
[0083] Therefore, data augmentation techniques can overcome the overfitting problem caused by small sample sizes and improve recognition accuracy.
[0084] In this invention, as a preferred embodiment, 6000 artificially generated samples are added to the original training samples as the training set for the 1D-CNN model. The confusion matrix for partial discharge pattern recognition based on data augmentation technology is shown in Table 2.
[0085] Table 2. Partial Discharge Pattern Recognition Confusion Rate Based on Data Augmentation Technology (%)
[0086]
[0087] Table 2 shows that the average accuracy rate for identifying partial discharge patterns under four typical defects was 98.8%. Among the four types of partial discharge, corona discharge and surface discharge had a 100% recognition rate, while internal discharge had a 98.4% recognition rate. 1.6% of the internal discharge test samples were incorrectly identified as floating discharge by the 1D-CNN. The floating discharge had the lowest recognition rate, but still reached 96.8%, with 2.2% of the floating samples being identified as internal discharge and 1% as surface discharge. The main reason for this discrepancy is that the cable defect model was not properly constructed during the partial discharge experiment, resulting in a high waveform similarity between internal and floating discharges. This caused a deviation in the optimization of model weights during the TRPD spectral feature parameter learning process of the 1D-CNN model.
[0088] Example 2
[0089] Comparison with existing classifiers
[0090] The performance of the 1D-CNN-based pattern recognition method is compared with existing classification models. Existing classification model one uses a backpropagation neural network (BPNN) with 2000 input layer nodes, 100 hidden layer nodes, and a learning rate of 0.0005; existing model two uses a support vector machine (SVM) with a radial basis function kernel and a penalty factor of 1000. Statistical features of the partial discharge signal's PRPD and TRPD spectra are extracted as input vectors for both classification models. These statistical features consist of 19 parameters, including discharge quantity factor, skewness, positive and negative half-cycle correlation coefficients, and kurtosis. Training samples are partial discharge training sample sets augmented with data augmentation techniques. The pattern recognition confusion matrices for BPNN and SVM are shown in Tables 3 and 4.
[0091] Table 3. BPNN Pattern Recognition Confusion Matrix (%)
[0092]
[0093]
[0094] Table 4. SVM Pattern Recognition Confusion Matrix (%)
[0095]
[0096] As shown in Tables 3 and 4, the average accuracy of pattern recognition based on BPNN is 82.55%, and the average accuracy of pattern recognition based on SVM is 85.95%, both significantly lower than the average recognition rate of 98.8% of this invention. BPNN has a high recognition rate for corona discharge and surface discharge, but for internal discharge and levitation discharge—two types of partial discharge with similar waveforms—the BPNN model cannot effectively improve the recognition accuracy. The SVM model has slightly better recognition accuracy than BPNN; however, its recognition rate for internal discharge and levitation discharge is still not high.
[0097] like Figure 4 As shown, from Figure 4 As can be seen, the 1D-CNN model's recognition accuracy for various defects is significantly higher than that of traditional BPNN and SVM models. For corona discharge and surface discharge, two relatively obvious partial discharge phenomena, the 1D-CNN model exhibits extremely high recognition accuracy. Internal discharge and levitation discharge show high waveform similarity. For internal discharge, the 1D-CNN recognition accuracy is 98.4%, which is 18.8% and 14.8% higher than BPNN and SVM, respectively; while for levitation discharge, the 1D-CNN recognition accuracy is 96.8%, which is 21% and 17.6% higher than BPNN and SVM, respectively. 1D-CNN, through multi-level deep convolution operations, can accurately extract features from the original data, while existing classifiers, which use statistical feature parameters, tend to lose detailed features of the data, reducing pattern recognition accuracy.
[0098] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.
Claims
1. A method for monitoring the operational status of power cables in urban underground utility tunnels, characterized in that, include: Step 1: In the selected integrated utility tunnel, use sensors to collect high-frequency pulse current signals of partial discharge during cable operation, and combine them with the corresponding time-series information to generate partial discharge map TRPD data. Step 2: Perform one-dimensional and normalization processing on TRPD data based on 1D-CNN convolutional network, divide the data into training and test sets, perform data augmentation on the training set sample data, and expand the initial training set. Step 3: Build a 1D-CNN model and train the 1D-CNN network based on the data-augmented training set; Step 4: Input the normalized test data into the 1D-CNN network model to monitor the cable operation status and output the monitoring results; The formula for calculating the 1D-CNN convolution is as follows: In the formula, f is the input sequence, g is the convolution kernel, * is the convolution operation, t is the position in the output sequence, k is the size of the convolution kernel, and n-1 is the upper limit of the convolution kernel.
2. The method for monitoring the operational status of power cables in urban underground utility tunnels according to claim 1, characterized in that, The structure of the 1D-CNN model includes: an input layer, four convolutional layers, four pooling layers, a fully connected layer, and an output layer. The input layer contains 1×2116 partial discharge TRPD waveform data. The convolutional kernel size is 1×9 with a compensation of 1. The pooling kernel size is 1×4 with a stride of 1. The last pooling layer is connected to the weighted connection layer. Data filtering and feature extraction are performed between layers through the ReLU nonlinear activation function.
3. The method for monitoring the operational status of power cables in urban underground utility tunnels according to claim 2, characterized in that, In step two, 75% of the data is used as the training set and 25% of the data is used as the test set.
4. The method for monitoring the operational status of power cables in urban underground utility tunnels according to claim 3, characterized in that, In step two, GAN-based data augmentation technology is used to expand the training samples for partial discharge identification. A noise signal z ~ N(0,1) is used as the input to the generator G, and a pseudo-partial discharge signal is synthesized to deceive the discriminator D. The discriminator D provides a prediction about this partial discharge signal. After multiple rounds of sufficient training, a Nash equilibrium state is reached, including: Step a: The generator network G maps a one-dimensional noise vector to a feature vector of shape with an input length of 1, and mixes the artificially generated spectrum with the real TRPD spectrum of partial discharge; Step b: The discriminator network D maps a TRPD spectrum of shape with an input length of 1 to a binary fraction, which is used to evaluate the probability that the production data is true; Step c: The GAN network connects network G and network D together. The generator decodes the latent space vector into feature vectors. The discriminator evaluates the authenticity of these feature parameters. The latent vector is mapped to the discriminator's evaluation result and output through the GAN network. Step d: Train the discriminator using real and fake TRPD spectral samples labeled "real" or "fake"; Step e: Train the generator based on the gradient of the loss of the GAN model relative to the generator weights. The direction of weight update is to maximize the probability that the discriminator will classify the decoded TRPD spectrogram as "true".
5. The method for monitoring the operational status of power cables in urban underground utility tunnels according to claim 4, characterized in that, In step three, after determining the network architecture, the CNN needs to be trained, which includes two processes: forward propagation and back propagation. The forward propagation process is as follows: O p =F n (...F2(F1(X P W (1) )W (2) )W (N) ); In the formula, W represents the network weight, and X represents the weight of the network. P Let F be the input feature and F be the activation function. Backpropagation requires continuous optimization of the weights W and biases b. The update formulas for weights W and biases b are as follows: In the formula, η is the loss function, l is the number of layers, and J is the loss function.
6. The method for monitoring the operational status of power cables in urban underground utility tunnels according to claim 5, characterized in that, Step four includes: Step 1: Construct a one-dimensional convolutional neural network model, adjust the network depth and convolutional kernel size, and determine the final network structure; Step 2: Compile network parameters, use classification cross-entropy as the loss function, and use gradient descent algorithm as the optimizer; Step 3: Train the 1D-CNN network using the data-augmented training samples, use the classification cross-entropy loss function to measure the distance between the predicted value and the true value, calculate the gradient of the loss with respect to the network parameters, and optimize the network parameters based on the SGD algorithm.
7. The method for monitoring the operational status of power cables in urban underground utility tunnels according to claim 5, characterized in that, The one-dimensional time-domain waveform eigenvector of TRPD is normalized using the following normalization function: In the formula, x min x is the minimum value of the waveform eigenvector. max This represents the maximum value of the waveform eigenvector.