A power distribution network transmission line composite fault intelligent diagnosis method
The DCAdResNet-CBAM model, with its adaptive stripping factor and dual-channel architecture, addresses the issues of low efficiency and poor accuracy in fault diagnosis of power distribution transmission lines, achieving efficient and accurate intelligent diagnosis of complex faults and improving the model's generalization performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-04-28
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies suffer from low efficiency and poor accuracy in fault diagnosis of power distribution network transmission lines. They are particularly difficult to meet the needs of rapid diagnosis in complex fault scenarios. Furthermore, commonly used AI models are highly dependent on the selection of fault features, which leads to a decrease in diagnostic accuracy.
The DCAdResNet-CBAM model, which employs an adaptive stripping factor and a dual-channel architecture, decomposes electrical signals into IMF components using SVMD, generates two-dimensional feature images by combining Markov transfer fields, and constructs the DCAdResNet-CBAM model for intelligent diagnosis, thereby reducing model training time and improving generalization performance.
It achieves efficient and accurate fault diagnosis in complex fault scenarios, reduces model training time and improves diagnostic accuracy, avoids the subjectivity problem of parameter selection in traditional methods, and improves the generalization ability of the model.
Smart Images
Figure CN122432633A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of data analysis and intelligent fault diagnosis technology for composite faults in power transmission lines, specifically relating to an intelligent diagnosis method for composite faults in power distribution network transmission lines. Background Technology
[0002] As a core infrastructure ensuring the orderly operation of social production and daily life, the stability of the power system is crucial to public safety; among them, the reliability of the distribution network transmission lines, as the key carriers of power transmission, is particularly important. Distribution network transmission lines are exposed to harsh natural environments and frequent operating conditions for extended periods, making them prone to various faults. Failure to identify and address these faults promptly and accurately can lead to serious safety accidents such as large-scale power outages and electrical fires.
[0003] Currently, fault identification in power distribution transmission lines mainly relies on manual analysis and judgment of fault waveform data and protection device operation information. However, with the continuous expansion of the coverage of modern power distribution networks and the increasing complexity of network topology, the frequency and types of faults have increased significantly. Moreover, there are often correlations and couplings between various faults. Manual analysis is not only inefficient and unable to meet the needs of rapid fault diagnosis, but it is also prone to subjective judgment bias, leading to diagnostic errors and affecting the timeliness and accuracy of fault handling.
[0004] In recent years, artificial intelligence (AI) has demonstrated enormous application potential in the field of power distribution network fault diagnosis technology due to its powerful data processing capabilities, especially in the scenario of complex fault identification in substations, where its advantages are even more prominent. Currently, researchers mainly apply AI technology to fault diagnosis through two technical approaches: one is to directly input fault data into the AI model, allowing the model to autonomously complete feature learning and fault identification; the other is to combine signal processing technology with AI algorithms to extract features and train the model.
[0005] Currently, commonly used AI models for fault diagnosis in distribution networks include decision trees, support vector machines, and neural networks. However, these models still have significant shortcomings in practical engineering applications: On the one hand, the effective training of such models is highly dependent on massive amounts of high-quality fault data. In scenarios where there are insufficient fault data samples or the data quality is poor, the model struggles to fully learn the fault characteristics, leading to a significant decrease in diagnostic accuracy. On the other hand, existing AI models generally suffer from numerous parameters and a high dependence on the selection of fault features. Inappropriate feature selection not only reduces the computational efficiency of the model but also affects the accuracy of fault diagnosis, making it difficult to meet the needs of accurate and efficient diagnosis of complex faults in large-scale distribution networks. Summary of the Invention
[0006] To address the aforementioned technical issues, this invention proposes an intelligent diagnostic method for complex faults in power distribution network transmission lines. This method effectively avoids the subjectivity and inefficiency of manual analysis while improving diagnostic accuracy and enabling intelligent diagnosis of complex faults. Furthermore, by proposing an adaptive stripping factor, the training time of the model is reduced while maintaining accuracy, and a dual-channel architecture is constructed to enhance the model's generalization performance.
[0007] The present invention provides an intelligent diagnosis method for complex faults in power distribution network transmission lines, comprising the following steps: Step 1: Extract the fault portion from the original three-phase voltage signal and splice it together in the order of phases A, B, and C to obtain the fault characteristic signal; Step 2: Use SVMD to decompose the original fault characteristic signal to obtain multiple IMF components; Step 3: Calculate the Pearson correlation coefficient between each IMF component and the original signal, set the correlation threshold, remove noisy IMF components, and reconstruct the signal. Step 4: Use MTF (Markov Transfer Field) to convert the reconstructed signal into two dimensions to obtain a two-dimensional feature image; Step 5: Construct the DCAdResNet-CBAM model; Step 6: Input the obtained two-dimensional feature image dataset into the DCAdResNet-CBAM model for training and testing to achieve intelligent fault diagnosis.
[0008] Furthermore, step 1 specifically includes: The three-phase signals are scanned sequentially, and the areas where voltage changes abruptly are marked as fault periods. The three-phase fault periods are extracted and spliced together in the order of phases A, B, and C to form the fault characteristic signal.
[0009] Furthermore, step 2 specifically includes: For the input signal Assume it can be decomposed into: , The following conditions must be met: , By solving the minimization problem under the above constraints, the modal components can be extracted. , in, , and As constraints, for Constrained filters for Constrained filter, where t is time. Let i be the time partial derivative operator, and i be the i-th IMF component in the first L-1 orders. It is an L-order mode; The residual signal includes the first L-1 modes obtained. and unprocessed parts ; For convolution operators, The center frequency of the Lth mode. To balance , and The parameter is j; j is the symbol for the imaginary part.
[0010] Furthermore, step 3 specifically includes: Calculate the Pearson correlation coefficient between each IMF component and the original signal: , Set a correlation threshold: , Remove IMF components with correlation less than the correlation threshold, retain IMF components with correlation greater than the correlation threshold, and reconstruct the signal; in, It is the Pearson correlation coefficient; and These are the observed values of the two variables. and These are the sample means of the two variables.
[0011] Furthermore, step 4 specifically includes: Step 401, for a time series ,Will Divided into There are quantile units, and the quantile corresponding to each data value is . The transitions between quantiles are constructed by calculating them along the time axis in the form of a first-order Markov chain. Markov transition matrix : , Step 402: Construct the matrix To obtain the dependency between location and time step: , in quantiles Located at point quantile The probability after, quantiles Transition to quantiles The transition probability; Step 403: Preferably, different color depths are mapped according to the numerical values of the state matrix elements, thereby visualizing the state matrix M and obtaining a two-dimensional feature image.
[0012] Further, in step 5, the DCAdResNet-CBAM model is constructed, including: The model structure is based on the ResNet-18 model, and CBAM is integrated into each residual block of the ResNet-18 model. Introducing an adaptive stripping factor In ResNet, Remove from each residual block CBAM, It can be 0 or 1; Adaptive updates are performed during training to avoid overfitting due to excessive model complexity; Introducing adaptive fusion parameters A dual-channel mechanism is used to improve the model's generalization ability; the output of channel 1 is... The output of channel 2 is The model output is , It updates adaptively in real time as training progresses.
[0013] Furthermore, step 6 specifically includes: Step 601: Divide the obtained fault feature image data into training and test sets for each class in an 8:2 ratio to form a fault feature image dataset.
[0014] Step 602: Input the fault feature image dataset into the DCAdResNet-CBAM model for training and testing to achieve intelligent fault diagnosis.
[0015] The beneficial effects of this invention are as follows: 1) Traditional line fault diagnosis methods require extracting electrical signal characteristic parameters at the time of the fault. The correlation between these parameters and the fault type relies on expert experience, and inappropriate parameters may fail to reflect the fault characteristics. The method presented in this invention maps electrical signals into two-dimensional image information using MTF (Mean Transformer Type Information), avoiding the subjectivity problem of parameter selection in traditional fault diagnosis methods. 2) The DCAdResNet-CBAM model constructed in this invention proposes an adaptive stripping factor to reduce the model training parameters while ensuring accuracy, thereby improving training effect and reducing training time; and constructs a dual-channel structure to improve the generalization performance of the model, and proposes an adaptive fusion parameter to dynamically adjust the output weights of different channels, thereby improving the diagnostic effect and generalization performance of the model in handling complex tasks. Attached Figure Description
[0016] Figure 1 Flowchart of the method described in this invention; Figure 2 This is a substation topology diagram; Figure 3 It is a characteristic signal diagram; Figure 4 This is an IMF decomposition time-domain diagram; Figure 5 This is an IMF component spectrum diagram; Figure 6 This is a threshold calculation graph; Figure 7 This is a comparison chart of the original signal and the reconstructed signal; Figure 8 It is a two-dimensional feature map; Figure 9 Here is a diagram of the ResNet-CBAM architecture; Figure 10 The results of the evaluation indicators are shown in the figure. Detailed Implementation
[0017] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.
[0018] like Figure 1 As shown, the intelligent diagnosis method for complex faults in power distribution network transmission lines according to the present invention includes the following steps: Step 1: Extract the fault portion from the original three-phase voltage signal and splice it together in the order of phases A, B, and C to obtain the fault characteristic signal; Step 2: Use SVMD to decompose the original fault characteristic signal to obtain multiple IMF components; Step 3: Calculate the Pearson correlation coefficient between each IMF component and the original signal, set the correlation threshold, remove noisy IMF components, and reconstruct the signal. Step 4: Use MTF (Markov Transfer Field) to convert the reconstructed signal into two dimensions to obtain a two-dimensional feature image; Step 5: Construct the DCAdResNet-CBAM model; Step 6: Input the obtained two-dimensional feature image dataset into the DCAdResNet-CBAM model for training and testing to achieve intelligent fault diagnosis.
[0019] Taking a distribution network in a certain city and a Simulink simulation model as examples, two topologies are used. The real fault dataset A and the simulated fault dataset B are constructed, as shown in Tables 1 and 2, respectively. Several faults are used as examples: single-phase grounding, two-phase-to-phase, two-phase-to-ground, and three-phase-to-ground. The invention will be further explained in detail below. The distribution network topology diagram is shown below. Figure 2 As shown.
[0020] Table 1. Sample grouping of real fault dataset
[0021] Table 2. Sample grouping of simulation fault dataset
[0022] In step 1, the three-phase signals are scanned sequentially, and regions where voltage abrupt changes are marked as fault periods. The three-phase fault periods are then extracted and spliced together in the order of phases A, B, and C to form fault characteristic signals. Taking several typical faults as examples, the spliced C-phase ground fault characteristic signals, CA phase-to-phase characteristic signals, CA-phase ground fault characteristic signals, and ABC-phase ground fault characteristic signals are shown below. Figure 3 As shown in (a), (b), (c), and (d).
[0023] Step 2 specifically includes: Step 201, for the input signal Assume it can be decomposed into: ; Step 202: For this assumption to hold, the following conditions must be met: ; Step 203: Solve the minimization problem of the above constraints to extract the modal components, i.e.:
[0024] in, It is an L-order mode. The residual signal includes the first L-1 modes obtained. and unprocessed parts . For convolution operators, The center frequency of the Lth mode. To balance , and The parameters. The decomposed IMF components are as follows: Figure 4 As shown. The spectrum of each IMF component is as follows. Figure 5 As shown, the decomposition is sufficient and no spectral aliasing occurs.
[0025] Step 3 specifically includes: Step 301: Calculate the Pearson correlation coefficient between each IMF component and the original signal: ; Step 302, set the correlation threshold: ; Step 303: Remove IMF components with correlation below the correlation threshold, retain IMF components with correlation above the correlation threshold, and reconstruct the signal. The Pearson correlation coefficients of the IMF components for each signal are shown in Table 3, and the reconstruction thresholds are as follows: Figure 6 As shown, the reconstructed signal after decomposition is as follows: Figure 7 As shown.
[0026] Table 3. Pearson correlation coefficients of IMF components
[0027] in, It is the Pearson correlation coefficient. and These are the observed values of two variables. and These are the sample means of the two variables.
[0028] Step 4 specifically includes: Step 401, for a time series ,Will Divided into There are quantile units, and the quantile corresponding to each data value is . The transitions between quantiles are constructed by calculating them along the time axis in the form of a first-order Markov chain. Markov transition matrix : , Step 402: Construct the matrix To obtain the dependency between location and time step: , in quantiles Located at point quantile The probability after, quantiles Transition to quantiles The transition probability, ; Step 403: Visualize the state matrix M to obtain a two-dimensional feature image. The feature images for C-phase grounding, CA phase-to-phase, CA grounding, and ABC grounding are shown below. Figure 8 As shown in (a), (b), (c), and (d).
[0029] Step 5 specifically includes: Step 501: Construct the ResNet-18 model.
[0030] Step 502: Add CBAM to each residual block of the ResNet-18 model to construct the ResNet-CBAM model. Its structure is as follows: Figure 9 As shown, channel attention and spatial attention are embedded in the residual block. The input is processed through multiple convolutions and the attention mechanism. The processed features are added to the original features to obtain the output.
[0031] Step 503, introduce an adaptive stripping factor Construct the AdResNet-CBAM model to represent the ResNet 1st generation. Remove from each residual block CBAM, It can be 0 or 1. The model is adaptively updated during training to avoid overfitting due to excessive model complexity.
[0032] Step 504, introduce adaptive fusion parameters A dual-channel mechanism is employed to construct DCAdResNet-CBAM, improving the model's generalization ability. The output of channel 1 is... The output of channel 2 is The model output is , It updates adaptively in real time as training progresses.
[0033] Step 6 specifically includes: Step 601: Divide the obtained fault feature image data into training and test sets for each class in an 8:2 ratio to form a fault feature image dataset.
[0034] Step 602: Input the fault feature image dataset into the DCAdResNet-CBAM model for training and testing to achieve intelligent fault diagnosis.
[0035] DCAdResNet-CBAM was used to train and test on datasets A and B. The accuracy and loss curves on dataset A are shown below. Figure 10 As shown in (a). The test set confusion matrix is as follows. Figure 10 As shown in (b), the accuracy and loss curves of the model on dataset B are as follows. Figure 10 As shown in (c). The test set confusion matrix is as follows. Figure 10 As shown in (d). DCAdResNet-CBAM is compared with several other traditional methods, and the comparison results are shown in Table 4.
[0036] Table 4 Comparison of Test Indicators
[0037] As shown in Table 4, the model proposed in this invention achieved good diagnostic results on both datasets, with its accuracy significantly better than other models. It was only slightly inferior to the DCResNet-CBAM model on dataset B. However, the model proposed in this invention has an adaptive stripping factor, so its training time on both datasets is less than that of the DCResNet-CBAM model, proving that the model has achieved a good balance between diagnostic accuracy and computational cost.
[0038] The above description is merely a preferred embodiment of the present invention and is not intended to further limit the present invention. All equivalent changes made based on the description and drawings of the present invention are within the protection scope of the present invention.
Claims
1. A method for intelligent diagnosis of complex faults in power distribution network transmission lines, characterized in that, The method includes: Step 1: Extract the fault portion from the original three-phase voltage signal and splice it together in the order of phases A, B, and C to obtain the fault characteristic signal; Step 2: Use SVMD to decompose the original fault characteristic signal to obtain multiple IMF components; Step 3: Calculate the Pearson correlation coefficient between each IMF component and the original fault feature signal. Based on the preset correlation threshold, remove the noisy IMF components and reconstruct the signal. Step 4: Use MTF to convert the reconstructed signal into two dimensions to obtain a two-dimensional feature image; Step 5: Construct the DCAdResNet-CBAM model; Step 6: Input the obtained two-dimensional feature image dataset into the DCAdResNet-CBAM model for training and testing to achieve intelligent fault diagnosis.
2. The intelligent diagnosis method for composite faults in power distribution network transmission lines according to claim 1, characterized in that, In step 1, the fault portion of the original three-phase voltage signal is extracted and spliced together in the order of phases A, B, and C to obtain the fault characteristic signal. The specific steps are as follows: The three-phase signals are scanned sequentially, and the areas where voltage changes abruptly are marked as fault periods. The three-phase fault periods are extracted and spliced together in the order of phases A, B, and C to form the fault characteristic signal.
3. The intelligent diagnosis method for composite faults in power distribution network transmission lines according to claim 1, characterized in that, In step 2, the original fault characteristic signal is decomposed using SVMD to obtain multiple IMF components; specifically: For the input signal Assume it can be decomposed into: ， The following conditions must be met: ， By solving the minimization problem under the above constraints, the modal components can be extracted. ， in, , and As constraints, for Constrained filters for Constrained filter, where t is time. Let i be the time partial derivative operator, and i be the i-th IMF component in the first L-1 orders. It is an L-order mode; The residual signal includes the first L-1 modes obtained. and unprocessed parts ; For convolution operators, The center frequency of the Lth mode. To balance , and The parameter is j; j is the symbol for the imaginary part.
4. The intelligent diagnosis method for composite faults in power distribution network transmission lines according to claim 1, characterized in that: Step 3 specifically involves: Calculate the Pearson correlation coefficient between each IMF component and the original signal: ， Set a correlation threshold: ， Remove IMF components with correlation less than the correlation threshold, retain IMF components with correlation greater than the correlation threshold, and reconstruct the signal; in, It is the Pearson correlation coefficient of the i-th IMF component; and These are the observed values of the two variables; and These are the sample means of the two variables.
5. The intelligent diagnosis method for composite faults in power distribution network transmission lines according to claim 1, characterized in that: Step 4 specifically involves: Step 401: Convert the time series Divided into There are quantile units, and the quantile corresponding to each data value is . The transitions between quantiles are constructed by calculating them along the time axis in the form of a first-order Markov chain. Markov transition matrix : ， Step 402: Construct the matrix To obtain the dependency between location and time step: ， in quantiles Located at point quantile The probability after, quantiles Transition to quantiles The transition probability; Step 403: Visualize the state matrix M to obtain a two-dimensional feature image.
6. The intelligent diagnosis method for composite faults in power distribution network transmission lines according to claim 1, characterized in that: Step 5 involves constructing the DCAdResNet-CBAM model, including: The model structure is based on the ResNet-18 model, and CBAM is integrated into each residual block of the ResNet-18 model. Introducing an adaptive stripping factor In ResNet, Remove from each residual block CBAM, It can be 0 or 1; Adaptive updates are performed during training; Introducing adaptive fusion parameters It adopts a dual-channel mechanism, and the output of channel 1 is The output of channel 2 is The model output is , It updates adaptively in real time as training progresses.
7. The intelligent diagnosis method for composite faults in power distribution network transmission lines according to claim 1, characterized in that: Step 6 specifically involves: Step 601: Divide the obtained fault feature image data into training and test sets for each class in an 8:2 ratio to form a fault feature image dataset; Step 602: Input the fault feature image dataset into the DCAdResNet-CBAM model for training and testing to achieve intelligent fault diagnosis.