Multi-modal Transform distribution network power frequency signal fault identification method based on Markov transfer field
Through the Markov transfer field and multimodal Transformer network architecture, the power frequency signal is converted into image features and fused with time series features, which solves the problems of insufficient accuracy and adaptability of existing power system fault detection methods and realizes efficient fault identification and real-time monitoring.
Patent Information
- Application Number
- CN202510975074.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-15
- Publication Date
- 2025-10-28
AI Technical Summary
Existing power system fault detection methods lack accuracy, have poor real-time performance, and poor adaptability when faced with complexity and dynamic changes. They find it difficult to fully explore the complex relationships and deep-level characteristics of power frequency signals, especially the coupling relationship between three-phase electrical quantities.
A multimodal Transformer network architecture based on Markov transfer fields is adopted. By converting power frequency signals into Markov transfer field image features and fusing them with time-series features, the Transformer model is used for fault identification, achieving adaptive learning of long-term dependencies.
It improves the accuracy and response speed of fault identification, enhances the robustness and adaptability of the system, enables a more comprehensive analysis of the health status of the power network, provides rich fault analysis information, and assists power engineers in making fault handling and maintenance decisions.
Smart Images
Figure CN120850119A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power fault detection technology, specifically to a multimodal Transformer method for identifying power frequency signal faults in distribution networks based on Markov transfer fields. Background Art
[0002] Power frequency signal fault identification in distribution networks is a crucial task in power systems, especially in the construction and operation of smart grids. Traditional power system fault detection methods largely rely on time-series analysis, rule-based fault diagnosis, and traditional machine learning algorithms. While these methods offer some fault detection capabilities, they fall short when faced with the complexity and dynamic changes of power systems. As power systems continue to expand, a large number of suspected faulty power frequency signals need to be detected, but traditional methods still suffer from insufficient accuracy, poor real-time performance, and poor adaptability. Specifically: 1) Traditional methods, such as threshold-based detection and cluster analysis, have limited ability to extract features from power frequency voltage and current data, making it difficult to fully explore the complex relationships and deep features in the data. For example, clustering methods can usually only capture simple distribution features of the data, and are weak in extracting complex features such as time-series dependencies and phase correlations.
[0003] 2) Traditional fault detection models lack in-depth analysis of the coupling relationships between three-phase electrical quantities when processing power frequency voltage and current data. In actual power systems, there are complex coupling and interdependence relationships between three-phase voltage and current signals, such as phase difference, amplitude relationship, and frequency variation. These factors together determine the operating state and potential faults of the system.
[0004] 3) Markov transfer fields can transform one-dimensional power frequency signals into two-dimensional images, providing a new perspective for the extraction of fault features.
[0005] 4) In terms of feature extraction of fault signals, traditional methods often rely on single-modality or simple feature extraction methods, making it difficult to fully capture the hidden features of complex fault signals. This patent introduces a multimodal transformer network architecture, thereby improving the accuracy of fault classification and diagnosis.
[0006] In summary, existing machine learning and time series analysis methods cannot fully capture the nonlinear dependencies and dynamic changes between power signals, resulting in inaccurate detection of complex faults. Rule-based fault detection methods are unable to adapt to new fault types in power systems, and are prone to false alarms or missed alarms. Therefore, there is an urgent need for a new fault identification method that can fully utilize the complex spatiotemporal dependencies in power frequency signals to improve the accuracy, real-time performance, and adaptability of fault identification. Summary of the Invention
[0007] The technical problem to be solved by the present invention is to provide a multimodal Transformer method for power frequency signal fault identification in distribution networks based on Markov transfer fields, so as to overcome the shortcomings of the prior art.
[0008] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A multimodal Transformer method for power frequency signal fault identification in distribution networks based on Markov transfer fields, comprising the following steps: Step S01: Obtain three-phase power frequency data from the data acquisition device to obtain the input dataset. ; Step S02: Process the input dataset Preprocessing is performed to obtain a continuous power frequency signal. ; Step S03: Convert the continuous power frequency signal Perform Markov transfer field transformation to obtain Markov transfer field image features. ; Step S04: For continuous power frequency signals Feature extraction is performed to obtain time-series features. ; Step S05: Transfer Markov transferred field image features and time series characteristics Perform feature fusion to obtain fused features ; Step S06: Merge features The input is fed into the Transformer model for training, resulting in a power frequency signal fault identification model; Step S07: Use the power frequency signal fault identification model to identify faults.
[0009] The beneficial effects of this invention are as follows: This invention can transform a one-dimensional power frequency signal into a two-dimensional image through a Markov transition field, providing a new perspective for fault feature extraction. This invention further extracts features from the power frequency signal, constructing efficient relevant information points that can represent waveform features. A Transformer model is used for time-series modeling, utilizing a self-attention mechanism to automatically capture long-term dependencies in the signal. The Transformer can adaptively learn the relationships between different time periods, thereby improving the sensitivity and detection accuracy of the fault evolution process.
[0010] This invention enhances the robustness and adaptability of fault identification through multimodal data fusion. The fusion of multimodal signals enables the system to more comprehensively analyze the health status of the power network, improving the accuracy of fault identification, especially under complex fault modes. The method of fusing image features generated from power frequency signals through Markov transfer field transformation with time-series features obtained based on piecewise mean extraction achieves effective complementarity and information gain between different data modes.
[0011] This invention can be integrated into monitoring and protection devices for power systems to achieve real-time monitoring and fault diagnosis of power lines and equipment. This system not only significantly improves the accuracy and response speed of fault detection but also provides rich fault analysis information to assist power engineers in fault handling and maintenance decisions.
[0012] The entire system employs the backpropagation algorithm to jointly train all modules, and uses an optimizer to update model parameters, achieving gradual convergence of the overall model. It can capture long-distance dependencies and complex nonlinear characteristics, improving the accuracy of identifying abnormal patterns and potential hazards in power line monitoring.
[0013] Based on the above technical solution, the present invention can be further improved as follows.
[0014] Furthermore, the preprocessing in step S02 specifically includes the following steps: Step S21: For the input dataset Noise filtering is performed to remove unwanted high-frequency noise and interference at specific frequencies; Step S22: The input dataset after noise filtering. Perform data standardization and calculate the input dataset. The eigenmean of each feature And calculate the standard deviation of each feature. The calculation formula is as follows: ; In the formula, It is the sample size. It is the first Feature values of each sample; Step S23: Standardize the feature values of each sample using the following formula: ; In the formula, These are the standardized eigenvalues; It is the first Feature values of each sample; This represents the standard deviation of the feature.
[0015] Furthermore, step S21 also includes: Before the noise filtering process, examine the input dataset. The timestamps or sampling points of the three-phase power frequency data within the time frame ensure that the signal data are sampled under the same time reference. If inconsistent sampling points are found, alignment processing is required. If there are signals with time misalignment or different sampling frequencies, interpolation is used for compensation, and feature points are aligned to ensure synchronization.
[0016] Furthermore, step S03 specifically includes the following steps: Step S31: Based on the inherent period of the power frequency signal Determine each cycle Number of internal sampling points , will cycle The preprocessed continuous power frequency signal described above Divide into several periodic subsequences ;in, For periodicity; For frequency; The total number of cycles, ; Step S32: Based on each of the said periodic subsequences The amplitude range determines the period. Minimum value of inner amplitude and maximum value Each of the said periodic subsequences Evenly divided into The amplitude range is The interval is used to obtain the discrete state sequence. : ; Step S33: Based on the discrete state sequence A Markov transition matrix is constructed, and the probability values in the Markov transition matrix are normalized and mapped to the gray values of the corresponding image pixels to obtain the image features of the Markov transition field. ; The Markov transfer field image features It includes the probability value of state transition, the initial state, and the state after the transition; wherein, the probability value of the state transition is the gray value of the corresponding image pixel; the initial state is the horizontal coordinate of the Markov transition matrix; and the state after the transition is the vertical coordinate of the Markov transition matrix.
[0017] Furthermore, step S04 specifically includes the following steps: Step S41: At the sampling rate Under these conditions, a fixed time window length of 20 milliseconds is selected, and the number of sampling points within each time window is [number missing]. ; Step S42: Convert the continuous power frequency signal Divided into The non-overlapping segments, the first The discrete signal within each segment is denoted as ,in, ;No. Arithmetic mean of each segment The formula is:
[0018] The arithmetic mean of all segments By combining the data, the temporal features are obtained. .
[0019] Furthermore, step S05 specifically includes the following steps: Step S51: Using feature projection technology, the features of the horse image are... With the aforementioned time-series features The transformation process, which maps each feature to the same feature space, is represented as follows: ; in , They are respectively and Projection matrix, , They are respectively and Bias term; Step S52: In the cross-attention module, image features are... as query vector Temporal characteristics as a key vector and value vector Calculate the cross-attention weights and use the following formula to achieve feature fusion: ; The fusion features are obtained based on the above formula. . Attached Figure Description
[0020] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a flowchart of the power frequency signal preprocessing of the present invention; Figure 3 This is a flowchart illustrating the feature fusion process of the present invention; Figure 4This is a flowchart of the model training process of the present invention. DETAILED DESCRIPTION
[0021] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.
[0022] like Figures 1-4 As shown in Example 1, the multimodal Transformer method for power frequency signal identification in distribution networks based on Markov transfer fields includes the following steps: Step S01: Obtain three-phase power frequency data from the data acquisition device to obtain the input dataset. Data acquisition devices include sensors in the power system; Step S02: Process the input dataset Preprocessing is performed to obtain a continuous power frequency signal. ; Step S03: Convert the continuous power frequency signal Perform Markov transfer field transformation to obtain Markov transfer field image features. ; Step S04: For continuous power frequency signals Feature extraction is performed to obtain time-series features. ; Step S05: Transfer Markov transferred field image features and time series characteristics Perform feature fusion to obtain fused features ; Step S06: Merge features The input is fed into the Transformer model for training, resulting in a power frequency signal fault identification model; Step S07: Fault identification using a power frequency signal fault identification model. This study proposes a system optimization framework based on the integration of power frequency data acquisition and machine learning models. The overall process can be divided into the following core steps: First, the original data is divided into three subsets: a training set, a test set, and a detection set, and a format conversion operation is performed to meet the input requirements of subsequent models. In the model building stage, an adaptive prediction model is trained using the training set data. For the converted data, Markov transition matrix features are extracted from the time-series data to model the signal's state transitions; simultaneously, power frequency information is extracted from the original signal for basic modeling of time-series behavior. Multimodal information from the Markov model and power frequency features is fused, and the fused multimodal features are received. Complex time-series and feature dependencies are modeled using a Transformer. The generalization performance is verified using a test set, and finally, model evaluation is completed on the detection set, with quantitative analysis of key indicators. Notably, the system introduces a negative feedback controller as a dynamic adjustment module, receiving real-time prediction error or confidence information from the model output. By adjusting data acquisition parameters or model hyperparameters, a closed-loop optimization mechanism is formed.
[0023] This invention transforms one-dimensional power frequency signals into two-dimensional images using Markov transfer fields, providing a new perspective for fault feature extraction. The invention further extracts features from the power frequency signal, constructing efficient relevant information points that can represent waveform characteristics. A Transformer model is used for time-series modeling, utilizing a self-attention mechanism to automatically capture long-term dependencies in the signal. The Transformer can adaptively learn the relationships between different time periods, thereby improving the sensitivity to fault evolution and detection accuracy.
[0024] This invention enhances the robustness and adaptability of fault identification through multimodal data fusion. The fusion of multimodal signals enables the system to more comprehensively analyze the health status of the power network, improving the accuracy of fault identification, especially under complex fault modes. The method of fusing image features generated from power frequency signals through Markov transfer field transformation with time-series features obtained based on piecewise mean extraction achieves effective complementarity and information gain between different data modes.
[0025] This invention can be integrated into monitoring and protection devices for power systems to achieve real-time monitoring and fault diagnosis of power lines and equipment. This system not only significantly improves the accuracy and response speed of fault detection but also provides rich fault analysis information to assist power engineers in fault handling and maintenance decisions.
[0026] The entire system employs the backpropagation algorithm to jointly train all modules, and uses an optimizer to update model parameters, achieving gradual convergence of the overall model. It can capture long-distance dependencies and complex nonlinear characteristics, improving the accuracy of identifying abnormal patterns and potential hazards in power line monitoring.
[0027] Example 2 is a further improvement based on Example 1, and its details are as follows: The preprocessing in step S02 specifically includes the following steps: Step S21: Process the input dataset Noise filtering is performed to remove unwanted high-frequency noise and interference at specific frequencies; Step S22: Process the input dataset after noise filtering. Data standardization is performed to eliminate amplitude differences between different signals, ensuring that all signals are analyzed on the same scale, and the input dataset is then calculated. The eigenmean of each feature And calculate the standard deviation of each feature. The calculation formula is as follows: ; In the formula, It is the sample size. It is the first The feature values of each sample. Calculate the standard deviation of each feature. , is used to measure the degree of dispersion of eigenvalues; Step S23: To avoid the impact of order-of-magnitude differences between features on the model's training performance, standardize the feature values of each sample using the following formula: ; In the formula, These are the standardized eigenvalues; It is the first Feature values of each sample; This represents the standard deviation of the feature.
[0028] Data standardization can scale different features to the same scale, thereby eliminating the differences in the units of measurement between features and ensuring that each feature is treated fairly during model training.
[0029] Example 3 is a further improvement based on Example 2, and its details are as follows: Step S21 also includes: Before noise filtering, examine the input dataset. The timestamps or sampling points of the three-phase power frequency data within the system are used to ensure that the signal data are sampled under the same time reference. If inconsistent sampling points are found, alignment processing is required. If there are signals with time misalignment or different sampling frequencies, interpolation is used for compensation to align feature points and ensure synchronization. The three-phase signals (phase A, phase B, and phase C) are processed separately to ensure the independence of each phase signal. During subsequent analysis, the time synchronization and integrity of the three-phase signals are maintained to avoid time misalignment problems caused by sampling differences.
[0030] Example 4 is a further improvement based on Example 1, and its details are as follows: Step S03 specifically includes the following steps: Markov Transition Field (MTF) is an effective method for converting time-series data into images, commonly used in image recognition and signal processing. MTF conversion not only preserves the original information of the time series but also visually represents the characteristics of the data in graphical form. Since power frequency signals under normal conditions have a fixed periodicity, they can be divided into multiple periods. For a 50Hz power frequency signal, one period is 20 milliseconds long.
[0031] Step S31: Based on the inherent period of the power frequency signal (For example, for a 50Hz power frequency signal, T =1 / 50 of a second, or 20 milliseconds), to determine each cycle. Number of internal sampling points , will cycle Continuous power frequency signal after internal preprocessing Divide into several periodic subsequences ;in, For periodicity; For frequency; The total number of cycles, ; Step S32: Based on each periodic subsequence The amplitude range determines the period. Minimum value of inner amplitude and maximum value , each periodic subsequence Evenly divided into The amplitude range is The interval is used to obtain the discrete state sequence. : ; This expression is a set of state sequences generated by discretizing each periodic subsequence. A symbolic representation corresponding to a periodic subsequence, used for subsequent analysis; Step S33: Based on discrete state sequence Construct a Markov transition matrix, normalize the probability values in the Markov transition matrix, and map them to the gray values of the corresponding image pixels to obtain the Markov transition field image features. ; Markov transfer field image features It includes the probability value of state transition, the initial state, and the state after the transition; where the probability value of state transition is the gray value of the corresponding image pixel; the initial state is the horizontal coordinate of the Markov transition matrix; and the state after the transition is the vertical coordinate of the Markov transition matrix.
[0032] Based on the characteristics of current and voltage signals, this invention sets... =2; After performing Markov transition field transformation on a one-dimensional signal, the state transition matrix is directly regarded as a grayscale image.
[0033] Example 5 is a further improvement based on Example 1, and its details are as follows: Step S04 specifically includes the following steps: This invention further extracts features from power frequency signals to construct efficient information points that can represent waveform characteristics. The signal is segmented using a fixed time window, and then the arithmetic mean of all sampling points within each segment is calculated, thereby obtaining mean features that reflect the overall characteristics and energy level of the signal.
[0034] Step S41: At the sampling rate Under these conditions, a fixed time window length of 20 milliseconds is selected, and the number of sampling points within each time window is [number missing]. ; Step S42: Convert the continuous power frequency signal Divided into The non-overlapping segments, the first The discrete signal within each segment is denoted as ,in, ;No. Arithmetic mean of each segment The formula is:
[0035] arithmetic mean of all segments By combining the features, we can obtain the temporal characteristics. , It is a mean sequence, reflecting the overall trend of the signal. The segments need to be of equal length to ensure the temporal characteristics. The physical meaning of .
[0036] Arithmetic mean That is to reflect the first The overall statistical characteristics and energy level of each segmented signal.
[0037] Example 6 is a further improvement based on Example 1, and its details are as follows: Step S05 specifically includes the following steps: Step S51: Use feature projection technology to project the features of the Marco image. With time series characteristics The transformation process, which maps each feature to the same feature space, is represented as follows: ; in , They are respectively and Projection matrix, , They are respectively and Bias term; Step S52: In the cross-attention module, image features are... as query vector Temporal characteristics as a key vector and value vector Calculate the cross-attention weights and use the following formula to achieve feature fusion: ; The fusion features are obtained based on the above formula. .
[0038] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1. A multimodal Transformer-based fault identification method for power frequency signals in distribution networks, characterized in that, Includes the following steps: Step S01: Obtain three-phase power frequency data from the data acquisition device to obtain the input dataset. ; Step S02: For the input dataset Preprocessing is performed to obtain a continuous power frequency signal. ; Step S03: Convert the continuous power frequency signal Perform Markov transfer field transformation to obtain Markov transfer field image features. ; Step S04: For the continuous power frequency signal Feature extraction is performed to obtain time-series features. ; Step S05: Transfer the Markov transfer field image features and the aforementioned time-series features Perform feature fusion to obtain fused features ; Step S06: The fusion features The input is fed into the Transformer model for training, resulting in a power frequency signal fault identification model; Step S07: Use the power frequency signal fault identification model to identify faults.
2. The multimodal Transformer distribution network power frequency signal fault identification method based on Markov transfer field according to claim 1, characterized in that, The preprocessing in step S02 specifically includes the following steps: Step S21: For the input dataset Noise filtering is performed to remove unwanted high-frequency noise and interference at specific frequencies; Step S22: The input dataset after noise filtering. Perform data standardization and calculate the input dataset. The eigenmean of each feature And calculate the standard deviation of each feature. The calculation formula is as follows: ； In the formula, It is the sample size. It is the first Feature values of each sample; Step S23: Standardize the feature values of each sample using the following formula: ； In the formula, These are the standardized eigenvalues; It is the first Feature values of each sample; This represents the standard deviation of the feature.
3. The multimodal Transformer distribution network power frequency signal fault identification method based on Markov transfer field according to claim 2, characterized in that, Step S21 further includes: Before the noise filtering process, examine the input dataset. The timestamps or sampling points of the three-phase power frequency data within the time frame ensure that the signal data are sampled under the same time reference. If inconsistent sampling points are found, alignment processing is required. If there are signals with time misalignment or different sampling frequencies, interpolation is used for compensation, and feature points are aligned to ensure synchronization.
4. The multimodal Transformer distribution network power frequency signal fault identification method based on Markov transfer field according to claim 1, characterized in that, Step S03 specifically includes the following steps: Step S31: Based on the inherent period of the power frequency signal Determine each cycle Number of internal sampling points , will cycle The preprocessed continuous power frequency signal described above Divide into several periodic subsequences ;in, For periodicity; For frequency; The total number of cycles, ; Step S32: Based on each of the said periodic subsequences The amplitude range determines the period. Minimum value of inner amplitude and maximum value Each of the said periodic subsequences Evenly divided into The amplitude range is The interval is used to obtain the discrete state sequence. : ； Step S33: Based on the discrete state sequence A Markov transition matrix is constructed, and the probability values in the Markov transition matrix are normalized and mapped to the gray values of the corresponding image pixels to obtain the image features of the Markov transition field. ; The Markov transfer field image features It includes the probability value of state transition, the initial state, and the state after the transition; wherein, the probability value of the state transition is the gray value of the corresponding image pixel; the initial state is the horizontal coordinate of the Markov transition matrix; and the state after the transition is the vertical coordinate of the Markov transition matrix.
5. The multimodal Transformer distribution network power frequency signal fault identification method based on Markov transfer field according to claim 1, characterized in that, Step S04 specifically includes the following steps: Step S41: At the sampling rate Under these conditions, a fixed time window length of 20 milliseconds is selected, and the number of sampling points within each time window is [number missing]. ; Step S42: Convert the continuous power frequency signal Divided into The non-overlapping segments, the first The discrete signal within each segment is denoted as ,in, ;No. Arithmetic mean of each segment The formula is: ； The arithmetic mean of all segments By combining the data, the temporal features are obtained. .
6. The multimodal Transformer distribution network power frequency signal fault identification method based on Markov transfer field according to claim 1, characterized in that, Step S05 specifically includes the following steps: Step S51: Using feature projection technology, the features of the horse image are... With the aforementioned time-series features The transformation process, which maps each feature to the same feature space, is represented as follows: ； in , They are respectively and Projection matrix, , They are respectively and Bias term; Step S52: In the cross-attention module, image features are... as query vector Temporal characteristics as a key vector and value vector Calculate the cross-attention weights and use the following formula to achieve feature fusion: ； The fusion features are obtained based on the above formula. .