Power distribution network fault diagnosis method and system based on improved empirical mode decomposition and hybrid neural network

Abstract

The invention discloses a power distribution network fault diagnosis method and system based on improved empirical mode decomposition and a hybrid neural network, and the method comprises the following steps: obtaining high-frequency fault disturbance waveform data, and dividing a fault disturbance waveform into a plurality of data segments through a moving time window; obtaining an intrinsic mode component set for each data segment by using an improved empirical mode decomposition method; a random component in the intrinsic mode component is screened and removed by using an improved run detection method containing an adaptive factor; mapping the processed three-phase voltage transient waveform to a complex field plane by using phase domain transformation; and constructing a power distribution network fault diagnosis model by using the multi-branch hybrid neural network, the input signal being the fault data converted to the complex field plane, and the output data being the fault type.

Description

Technical Field

[0001] This invention belongs to the field of power distribution network fault diagnosis technology, and particularly relates to an improved empirical mode decomposition and hybrid neural network method and system for power distribution network fault diagnosis. Background Technology

[0002] With the large-scale integration of distributed energy resources and new loads, the power grid structure and operation modes are becoming increasingly complex and diverse, and power flow is showing a bidirectional trend. How to address power system fault diagnosis in complex operating environments is a pressing issue. With the rapid development of artificial intelligence technology and the reduction in communication and data storage costs, adopting data-driven methods for distribution network fault diagnosis will become a major trend and direction for future development.

[0003] Traditional empirical mode decomposition (EMD) suffers from mode aliasing and decomposition stability issues. Specifically, during waveform decomposition, spurious components exist in the upper and lower envelopes of the waveform to be decomposed, causing frequency abrupt changes at the endpoints. In fault classification, on the one hand, neural network-based methods suffer from low detection accuracy and overfitting. On the other hand, fault classification models based on attention-based networks have problems such as large model parameter capacity and long training computation time. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides an improved empirical mode decomposition and hybrid neural network method and system for distribution network fault diagnosis. The improved empirical mode decomposition method and multi-branch hybrid neural network are used for fault diagnosis, thereby improving the accuracy, real-time performance and reliability of distribution network fault diagnosis.

[0005] The technical solution of this invention is implemented as follows: An improved method for fault diagnosis in power distribution networks using empirical mode decomposition and hybrid neural networks includes the following steps: S1. Obtain high-frequency fault disturbance waveform data and divide the fault disturbance waveform into multiple data segments using a moving time window; S2. Apply the improved empirical mode decomposition method to each data segment to obtain the set of intrinsic mode components; S3. Use an improved run detection method with adaptive factors to filter and remove random components in intrinsic mode components; S4. Use phase domain transformation to map the processed three-phase voltage transient waveform to the complex domain plane; S5. A distribution network fault diagnosis model is constructed using a multi-branch hybrid neural network. The input signal is fault data transformed to the complex number plane, and the output data is the fault type.

[0006] Furthermore, in S1, the method for dividing the fault disturbance waveform is as follows: The fault disturbance waveform data is divided into multiple non-overlapping data segments using a moving time window, as expressed by the following formula: If the original data is x 0=[ x (1), x (2), x (3),…, x ( n )]; No. k The data segments obtained in each time window are represented as follows: x k =[ x ( ka – k +1), x ( ka – k +2),…, x ( ka )]; The length of each time window is a The original data is divided into: n / a Each segment.

[0007] Furthermore, in S2, the improved empirical mode decomposition method specifically includes: S2.1, let it include a The sampling point of the first sampling point k The time window data segment is x k The local maxima and local minima of the curve are determined, and the points where the local maxima and local minima are found are used as known points. New points are then added within the known points by constructing a polynomial function to obtain the upper envelope curve. and lower envelope curve ; S2.2 Calculate the average values ​​of the upper and lower envelope curves. The specific calculation formula is as follows: ; The waveform similarity calculation method is used to find matching waveforms. The waveform similarity index is represented by the sum of the squared amplitude distances between the current curve and the curve to be matched at the first endpoint, local maxima, and local minima. The optimal matching curve is determined, and the average value of the envelope curve is calculated. The spurious components at the endpoints are replaced with the true components to maintain consistent frequency characteristics of the screening waveform; the average value of the replaced upper and lower envelopes is defined as... ; S2.3 Calculate the fault disturbance waveform data segment and The difference and judge Whether it belongs to an intrinsic mode component, the fault disturbance waveform is screened using the following formula: ; in, This is the cumulative deviation indicator. and The first j Second and third j -1st iteration k Segment Fault Disturbance Waveform Difference exist t The value at time; S2.4 Preserve intrinsic modal components Update the fault disturbance waveform data segment ; S2.5. Apply the improved empirical mode decomposition method to each segment of fault disturbance waveform data in sequence to obtain the set of intrinsic mode components.

[0008] Furthermore, in S2.3, if the following conditions are met... Then determine the difference in the current fault disturbance waveform. These are intrinsic modal components; If not satisfied Then As the fault disturbance waveform to be decomposed, return to execute S2.1.

[0009] Furthermore, in S2.4, the fault disturbance waveform data segment is updated. The specific formula is as follows: ; Determine whether the updated fault disturbance waveform data segment meets the termination condition. The specific determination condition is as follows: It basically shows a monotonous trend or If the amplitude is very small, it can be considered as an error. If the condition is met, the empirical mode decomposition is considered complete; otherwise, it will be updated. As the curve to be decomposed, return to execute S2.1.

[0010] Furthermore, in step S3, the specific steps of the improved run-length detection method with adaptive factors are as follows: S3.1 Convert the data of each intrinsic mode component into a time series of 0 and 1. Data in the intrinsic mode component that is less than the median is represented by 0, and data that is greater than the median is represented by 1. S3.2, Statistics l Total number of runs in the intrinsic mode component transition sequence yl A sequence of identical numbers without interruption in the transformation sequence represents a run. The number of runs is calculated. l The number of 0s and 1s in the transformation sequence of each intrinsic mode component is respectively represented by... n 1 and n 2 represents; in accordance with n 1 and n 2. Calculate the critical value for run detection y l,max and y l,min Using the mathematical expectation of the number of runs E yl and variance D yl Calculate the critical value for the number of runs: ; ; ; in, Z P Values ​​are normally distributed. P The significance level; S3.3 Determine whether each intrinsic modal component is a random sequence.

[0011] Furthermore, in S3.3, it is determined whether each intrinsic mode component is a random sequence, the main changing trend of each power quality disturbance waveform curve is extracted, and an adaptive factor is introduced. ε If the following conditions are met: ; Then determine the first l The class's intrinsic modal components are random sequences.

[0012] Furthermore, in S4, the specific method of using phase domain transformation is as follows: S4.1 Remove the identified random sequence from the intrinsic mode components and reconstruct the fault disturbance transient signal; S4.2, convert the three-phase voltage transient signal v 1( t ), v 2( t ), v 3( t Transformation to complex plane signal v α ( t )and v β ( t The method is as follows: ; S4.3. The size of the image matrix is ​​set proportionally according to the sampling frequency of the transient waveform, and the sample size is determined by the sampling frequency of the transient waveform. α , β The boundary of the complex number field plane is determined by 1.2 times the maximum value on the axis; S4.4. Convert the complex domain data into a 0-1 square matrix according to the set image matrix size. If there are transformation points in the image matrix, set the matrix element to 1.

[0013] Furthermore, in S5, a multi-branch hybrid neural network is used to diagnose the type of fault in the power distribution network; The power distribution network fault diagnosis model is a hybrid neural network with multiple branches. The input of the diagnosis model is a 0-1 square matrix. The hybrid neural network includes convolutional layers, pooling layers and fully connected layers. The two-dimensional convolutional neural network model branch is used to extract the spatial features of the image, and the one-dimensional convolutional neural network model branch is used to extract the multi-scale features of the feature image.

[0014] An improved empirical mode decomposition and hybrid neural network fault diagnosis system for distribution networks applies the improved empirical mode decomposition and hybrid neural network fault diagnosis method for distribution networks described in any one of the above-mentioned methods.

[0015] Compared with the prior art, the present invention achieves the following beneficial effects: This invention provides an improved method and system for fault diagnosis in distribution networks based on empirical mode decomposition (EMD) and hybrid neural networks. It includes a fault waveform preprocessing method based on moving time windows and jointly improved EMD, and a fault diagnosis method based on multi-branch hybrid neural networks. This method can efficiently process long-term fault signals and high-frequency sampled fault signals. The universality and stability of the method provided by this invention are better than other signal analysis methods. The proposed method can accurately classify fault types in distribution networks and reliably extract fault features. Compared to other deep learning-based fault diagnosis methods, the method of this invention improves the classification accuracy of the model through hybrid neural networks. Compared to attention mechanism networks, the method of this invention has the advantages of being lightweight and efficient. Attached Figure Description

[0016] Picture 1 This is a flowchart of an improved empirical mode decomposition and hybrid neural network method for fault diagnosis in power distribution networks, provided in an embodiment of the present invention. Picture 2 This is the run detection preprocessing result of an improved empirical mode decomposition and hybrid neural network-based power distribution network fault diagnosis method provided in this embodiment of the invention; Picture 3 This invention provides a fault diagnosis model based on a hybrid neural network, which is an improved empirical mode decomposition and hybrid neural network-based method for fault diagnosis in power distribution networks. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0018] Example like Picture 1 to Picture 3 An improved method for fault diagnosis in distribution networks using empirical mode decomposition and hybrid neural networks includes the following steps: S1. Obtain high-frequency fault disturbance waveform data and divide the fault disturbance waveform into multiple data segments using a moving time window; In S1, the method for dividing the fault disturbance waveform is as follows: The fault disturbance waveform data is divided into multiple non-overlapping data segments using a moving time window, as expressed by the following formula: If the original data is x 0=[ x (1), x (2), x (3),…, x ( n )]; No. k The data segments obtained in each time window are represented as follows: x k =[ x ( ka – k +1), x ( ka – k +2),…, x ( ka )]; The length of each time window is a The original data is divided into: n / a Each segment.

[0019] S2. Apply the improved empirical mode decomposition method to each data segment to obtain the set of intrinsic mode components; In S2, the improved empirical mode decomposition method specifically includes: S2.1, let it include a The sampling point of the first sampling point k The time window data segment is x kThe local maxima and local minima of the curve are determined, and the points where the local maxima and local minima are found are used as known points. New points are then added within the known points by constructing a polynomial function to obtain the upper envelope curve. and lower envelope curve ; S2.2 Calculate the average values ​​of the upper and lower envelope curves. The specific calculation formula is as follows: ; The waveform similarity calculation method is used to find matching waveforms. The waveform similarity index is represented by the sum of the squared amplitude distances between the current curve and the curve to be matched at the first endpoint, local maxima, and local minima. The optimal matching curve is determined, and the average value of the envelope curve is calculated. The spurious components at the endpoints are replaced with the true components to maintain consistent frequency characteristics of the screening waveform; the average value of the replaced upper and lower envelopes is defined as... ; S2.3 Calculate the fault disturbance waveform data segment and The difference and judge Whether a component belongs to an Inherent Mode Form (IMF) component is determined to screen the fault disturbance waveform. The specific formula is as follows: ; in, This is the cumulative deviation indicator. and The first j Second and third j -1st iteration k Segment Fault Disturbance Waveform Difference exist t The value at time; If satisfied Then determine the difference in the current fault disturbance waveform. These are intrinsic modal components; If not satisfied Then As the fault disturbance waveform to be decomposed, return to execute S2.1.

[0020] S2.4 Preserve intrinsic modal components Update the fault disturbance waveform data segment The specific formula is: ; Determine whether the updated fault disturbance waveform data segment meets the termination condition. The specific determination condition is as follows: It basically shows a monotonous trend or If the amplitude is very small, it can be considered as an error. If the condition is met, the empirical mode decomposition is considered complete; otherwise, it will be updated. As the curve to be decomposed, return to execute S2.1.

[0021] S2.5. Apply the improved empirical mode decomposition method to each segment of fault disturbance waveform data in sequence to obtain the set of inherent mode components, ensuring the stability and timeliness of the method.

[0022] S3. Use an improved run detection method with adaptive factors to filter and remove random components in intrinsic mode components; In step S3, the specific steps of the improved run-length detection method with adaptive factors are as follows: S3.1 Convert the data of each intrinsic mode component into a time series of 0 and 1. Data in the intrinsic mode component that is less than the median is represented by 0, and data that is greater than the median is represented by 1. S3.2, Statistics l Total number of runs in the intrinsic mode component transition sequence y l A sequence of identical numbers without interruption in the transformation sequence represents a run. The number of runs is calculated. l The number of 0s and 1s in the transformation sequence of each intrinsic mode component is respectively represented by... n 1 and n 2 represents; in accordance with n 1 and n 2. Calculate the critical value for run detection y l,max and y l,min Using the mathematical expectation of the number of runs E yl and variance D yl Calculate the critical value for the number of runs: ; ; ; in, Z P Values ​​are normally distributed. P The significance level; S3.3 Determine whether each intrinsic mode component is a random sequence, extract the main changing trend of each power quality disturbance waveform curve, and introduce an adaptive factor. ε If the following conditions are met: ; Then determine the first l The class's intrinsic modal components are random sequences.

[0023] By changing the adaptive factor ε It can accurately extract fault disturbance waveforms. Picture 2 The results of preprocessing the perturbation waveforms under different adaptive factors are shown. a is the original waveform; b is the preprocessed waveform with ε=5; c is the preprocessed waveform with ε=20.

[0024] S4. Use phase domain transformation to map the processed three-phase voltage transient waveform to the complex domain plane; In S4, the specific method of using phase domain transformation is as follows: S4.1 Remove the identified random sequence from the intrinsic mode components and reconstruct the fault disturbance transient signal; S4.2, convert the three-phase voltage transient signal v 1( t ), v 2( t ), v 3( t Transformation to complex plane signal v α ( t )and v β ( t The method is as follows: ; S4.3. The size of the image matrix is ​​set proportionally according to the sampling frequency of the transient waveform, and the sample size is determined by the sampling frequency of the transient waveform. α , β The boundary of the complex number field plane is determined by 1.2 times the maximum value on the axis; S4.4. Convert the complex domain data into a 0-1 square matrix according to the set image matrix size. If there are transformation points in the image matrix, set the matrix element to 1.

[0025] S5. A distribution network fault diagnosis model is constructed using a multi-branch hybrid neural network. The input signal is fault data transformed to the complex number plane, and the output data is the fault type.

[0026] In S5, a multi-branch hybrid neural network is used to diagnose the type of fault in the power distribution network. The power distribution network fault diagnosis model is a hybrid neural network with multiple branches. The two-dimensional convolutional neural network (CNN) branch is used to extract spatial features of the image, while the one-dimensional CNN branch is used to extract multi-scale features of the feature image. The input to the diagnostic model is a 0-1 square matrix; the input data size and number of channels for the two-dimensional CNN are (16,16) and 3, respectively, while the input data size and number of channels for the one-dimensional CNN are 256 and 3, respectively. The hybrid neural network includes convolutional layers, pooling layers, and fully connected layers.

[0027] Utilizing multi-branch hybrid neural networks can improve the classification accuracy of the model. This invention mainly considers single-phase grounding faults, two-phase grounding faults, and three-phase grounding faults. Picture 3 The diagram shows a hybrid neural network structure for a fault diagnosis model. Branch 1 includes convolutional and pooling layers to compress two-dimensional image channel data and extract key image features, while branch 2 includes image shape transformation and a one-dimensional convolutional neural network.

[0028] An improved empirical mode decomposition and hybrid neural network fault diagnosis system for power distribution networks applies the improved empirical mode decomposition and hybrid neural network fault diagnosis method described above.

[0029] This invention provides an improved method and system for fault diagnosis in power distribution networks based on empirical mode decomposition (EMD) and hybrid neural networks. Traditional EMD suffers from stability and reliability issues when dealing with long-sequence fault signals. Furthermore, traditional EMD methods have endpoint problems, easily leading to inconsistencies in the frequencies of the decomposed intrinsic mode components. This invention proposes a fault waveform preprocessing method based on a moving-time window-based joint improvement of EMD. The moving-time window segments the high-sampling-frequency fault waveform, improving computational efficiency and waveform decomposition stability. A waveform sieving component extension method replaces spurious components in the sieving waveform with real components, resolving the mode aliasing problem in traditional EMD, while simultaneously extracting noise random components from the fault disturbance waveform. On one hand, compared to traditional EMD methods, this invention is suitable for processing long-sequence and high-sampling-frequency waveform data and features high computational efficiency. On the other hand, most improved EMD methods typically employ mirror symmetry and original waveform extension. This invention's method, by extending the waveform at the endpoints of the sieving components, helps to ensure consistent frequency characteristics of the sieving waveform, thereby solving the mode aliasing problem in traditional EMD.

[0030] This invention utilizes a moving time window to divide the fault time series into multiple non-overlapping fault periods. An improved empirical mode decomposition method is then applied to each fault period to accurately and stably extract fault feature components. Attention mechanism networks suffer from problems such as large model parameter count and long training computation time. This invention proposes a fault diagnosis method based on a multi-branch hybrid neural network, which accurately determines the fault type of the distribution network by extracting spatial and multi-scale features of the fault. First, the fault waveform is converted into a two-dimensional image using waveform transformation. Based on this, a two-dimensional convolutional neural network branch extracts spatial location feature information from the image, while a one-dimensional convolutional neural network branch extracts multi-scale feature information, solving the gradient vanishing problem and giving the model a certain degree of robustness and stability, thus improving its fault diagnosis accuracy. Compared to traditional fault diagnosis models based on deep neural networks, this invention's method uses a hybrid neural network to improve the model's ability to extract multi-dimensional feature information and its detection accuracy. The multi-branch hybrid neural network effectively solves the gradient vanishing and overfitting problems during model training. Compared to attention mechanism networks, this invention's method uses a smaller total number of model parameters and fewer model layers, offering advantages in terms of lightweight design and efficiency.

[0031] Based on the disclosure and teachings of the foregoing specification, those skilled in the art can make changes and modifications to the above embodiments. Therefore, the present invention is not limited to the specific embodiments disclosed and described above, and some modifications and changes to the present invention should also fall within the protection scope of the claims of the present invention. Furthermore, although some specific terms are used in this specification, these terms are only for convenience of explanation and do not constitute any limitation on the present invention.

Claims

1. An improved method for fault diagnosis in power distribution networks using empirical mode decomposition and hybrid neural networks, characterized in that, Includes the following steps: S1. Obtain high-frequency fault disturbance waveform data and divide the fault disturbance waveform into multiple data segments using a moving time window; S2. Apply the improved empirical mode decomposition method to each data segment to obtain the set of intrinsic mode components; S3. Use an improved run detection method with adaptive factors to filter and remove random components in intrinsic mode components; S4. Use phase domain transformation to map the processed three-phase voltage transient waveform to the complex domain plane; S5. A distribution network fault diagnosis model is constructed using a multi-branch hybrid neural network. The input signal is fault data transformed to the complex number plane, and the output data is the fault type.

2. The method for fault diagnosis of distribution networks based on improved empirical mode decomposition and hybrid neural networks according to claim 1, characterized in that, In S1, the method for dividing the fault disturbance waveform is as follows: The fault disturbance waveform data is divided into multiple non-overlapping data segments using a moving time window, as expressed by the following formula: If the original data is x 0=[ x (1), x (2), x (3),…, x ( n )]; No. k The data segments obtained in each time window are represented as follows: x k =[ x ( ka – k +1), x ( ka – k +2),…, x ( ka )]; The length of each time window is a The original data is divided into: n / a Each segment.

3. The method for fault diagnosis of distribution networks based on improved empirical mode decomposition and hybrid neural networks according to claim 2, characterized in that, In S2, the improved empirical mode decomposition method specifically includes: S2.1, let it include a The sampling point of the first sampling point k The time window data segment is x k The local maxima and local minima of the curve are determined, and the points where the local maxima and local minima are found are used as known points. New points are then added within the known points by constructing a polynomial function to obtain the upper envelope curve. and lower envelope curve ; S2.2 Calculate the average values ​​of the upper and lower envelope curves. The specific calculation formula is as follows: ； The waveform similarity calculation method is used to find matching waveforms. The waveform similarity index is represented by the sum of the squared amplitude distances between the current curve and the curve to be matched at the first endpoint, local maxima, and local minima. The optimal matching curve is determined, and the average value of the envelope curve is calculated. The spurious components at the endpoints are replaced with the true components to maintain consistent frequency characteristics of the screening waveform; the average value of the replaced upper and lower envelopes is defined as... ; S2.3 Calculate the fault disturbance waveform data segment and The difference and judge Whether it belongs to an intrinsic mode component, the fault disturbance waveform is screened using the following formula: ； in, This is the cumulative deviation indicator. and The first j Second and third j -1st iteration k Segment Fault Disturbance Waveform Difference exist t The value at time; S2.4 Preserve intrinsic modal components Update the fault disturbance waveform data segment ; S2.

5. Apply the improved empirical mode decomposition method to each segment of fault disturbance waveform data in sequence to obtain the set of intrinsic mode components.

4. The improved empirical mode decomposition and hybrid neural network method for distribution network fault diagnosis according to claim 3, characterized in that, In S2.3, if the following conditions are met... Then determine the difference in the current fault disturbance waveform. These are intrinsic modal components; If not satisfied Then As the fault disturbance waveform to be decomposed, return to execute S2.

1.

5. The improved empirical mode decomposition and hybrid neural network method for distribution network fault diagnosis according to claim 3, characterized in that, In step S2.4, the fault disturbance waveform data segment is updated. The specific formula is as follows: ； Determine whether the updated fault disturbance waveform data segment meets the termination condition. The specific determination condition is as follows: It basically shows a monotonous trend or If the amplitude is very small, it can be considered as an error. If the condition is met, the empirical mode decomposition is considered complete; otherwise, it will be updated. As the curve to be decomposed, return to execute S2.

1.

6. The method for fault diagnosis of distribution networks based on improved empirical mode decomposition and hybrid neural networks according to claim 1, characterized in that, In step S3, the specific steps of the improved run-length detection method with adaptive factors are as follows: S3.1 Convert the data of each intrinsic mode component into a time series of 0 and 1. Data in the intrinsic mode component that is less than the median is represented by 0, and data that is greater than the median is represented by 1. S3.2, Statistics l Total number of runs in the intrinsic mode component transition sequence y l A sequence of identical numbers without interruption in the transformation sequence represents a run. The number of runs is calculated. l The number of 0s and 1s in the transformation sequence of each intrinsic mode component is respectively represented by... n 1 and n 2 represents; in accordance with n 1 and n 2. Calculate the critical value for run detection y l,max and y l,min Using the mathematical expectation of the number of runs E yl and variance D yl Calculate the critical value for the number of runs: ； ； ； in, Z P Values ​​are normally distributed. P The significance level; S3.3 Determine whether each intrinsic modal component is a random sequence.

7. The improved empirical mode decomposition and hybrid neural network method for distribution network fault diagnosis according to claim 6, characterized in that, In step S3.3, it is determined whether each intrinsic mode component is a random sequence, the main changing trend of each power quality disturbance waveform curve is extracted, and an adaptive factor is introduced. ε If the following conditions are met: ； Then determine the first l The class's intrinsic modal components are random sequences.

8. The improved empirical mode decomposition and hybrid neural network method for distribution network fault diagnosis according to claim 7, characterized in that, In S4, the specific method of using phase domain transformation is as follows: S4.1 Remove the identified random sequence from the intrinsic mode components and reconstruct the fault disturbance transient signal; S4.2, convert the three-phase voltage transient signal v 1( t ), v 2( t ), v 3( t Transformation to complex plane signal v α ( t )and v β ( t The method is as follows: ； S4.

3. The size of the image matrix is ​​set proportionally according to the sampling frequency of the transient waveform, and the sample size is determined by the sampling frequency of the transient waveform. α , β The boundary of the complex number field plane is determined by 1.2 times the maximum value on the axis; S4.

4. Convert the complex domain data into a 0-1 square matrix according to the set image matrix size. If there are transformation points in the image matrix, set the matrix element to 1.

9. The method for fault diagnosis of distribution networks based on improved empirical mode decomposition and hybrid neural networks according to claim 8, characterized in that, In S5, a multi-branch hybrid neural network is used to diagnose the type of fault in the power distribution network. The power distribution network fault diagnosis model is a hybrid neural network with multiple branches. The input of the diagnosis model is a 0-1 square matrix. The hybrid neural network includes convolutional layers, pooling layers and fully connected layers. The two-dimensional convolutional neural network model branch is used to extract the spatial features of the image, and the one-dimensional convolutional neural network model branch is used to extract the multi-scale features of the feature image.

10. A power distribution network fault diagnosis system with improved empirical mode decomposition and hybrid neural networks, characterized in that, A distribution network fault diagnosis method based on an improved empirical mode decomposition and hybrid neural network as described in any one of claims 1 to 9 is applied.

Images