Multi-mode-based transformer data fusion and governance method

Abstract

The invention relates to a multi-modal-based transformer data fusion and treatment method. The method comprises the following steps: performing standardized access and data dynamic mapping on multi-modal data of a transformer; constructing a unique identifier of the mapped transformer data; constructing a transformer state data set according to the unique identifier, and performing multi-strategy data management to complete cleaning management of the state data set; based on the cleaned data set, performing multi-modal alignment fusion to obtain a transformer multi-modal fusion data set; and based on the multi-modal fusion data set of the transformer, evaluating the data management method. According to the method, standardized access, management and alignment fusion of multi-modal data can be realized under a unified framework, and the method is of great significance in improving the accuracy, reliability and timeliness of state sensing of the power grid transformer.

Description

Technical Field

[0001] The present invention belongs to the technical field of transformer diagnosis, and in particular to a transformer data fusion and management method based on multimodality. Background Art

[0002] As power systems continue to expand and become increasingly complex, transformers, as crucial equipment within them, are directly impacted by their operational status, impacting the safety and stability of the power grid. By introducing effective transformer data governance methods, in-depth analysis of operational data can be conducted to identify potential risks and optimize maintenance strategies, thereby enhancing the intelligence of equipment operation and maintenance. Against this backdrop, the collection, fusion, and governance of multi-source, multi-modal data have gradually become a research hotspot. However, given the diverse data sources and complex data types, achieving standardized access, alignment, and fusion of transformer multimodal data, as well as effective governance of data quality issues, remains a challenge.

[0003] Manual rule-based approaches based on specific scenarios are the foundation of transformer multi-source data governance, but rule development relies on domain knowledge and lacks automation and adaptability. Models such as random forests, support vector machines, and convolutional neural networks extract transformer features and implement multi-source data fusion and governance at the feature level. While these methods can achieve data governance, they lack data standardization when dealing with multimodal and highly spatiotemporally correlated transformer state perception data, making it difficult to achieve unified data description and processing. Transformer state perception data includes multimodal information such as voltage, current, temperature, and oil chromatography gas. These data come from different sensors and may have inconsistent sampling frequencies and timestamps, resulting in data misalignment in time and space. Cross-modal correlation anomalies exist between different modal data. Traditional anomaly detection methods, such as statistical analysis, are effective in single-variable scenarios but struggle to handle multimodal and highly spatiotemporally correlated anomaly patterns. Existing methods struggle to address these issues in multimodal data governance, resulting in loose data integration and an inability to fully explore the inherent connections between multimodal data.

[0004] Therefore, studying a multimodal transformer data fusion and governance technology can achieve standardized access, governance and alignment fusion of multimodal data under a unified framework, which is of great significance to improving the accuracy, reliability and timeliness of power grid transformer status perception. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings of the existing technology and propose a transformer data fusion and governance method based on multimodality, which can realize the standardized access, governance and alignment fusion of multimodal data under a unified framework, which is of great significance to improving the accuracy, reliability and timeliness of power grid transformer status perception.

[0006] The present invention solves the technical problem by adopting the following technical solutions:

[0007] A transformer data fusion and management method based on multimodality includes the following steps:

[0008] Step 1: Standardize access and dynamically map transformer multimodal data;

[0009] Step 2: Construct a unique identifier for the mapped transformer data;

[0010] Step 3: Construct a transformer status dataset based on the unique identifier, perform multi-strategy data management, and complete the cleaning and management of the status dataset;

[0011] Step 4: Based on the cleaned dataset, perform multimodal alignment and fusion to obtain the transformer multimodal fusion dataset;

[0012] Step 5: Evaluate the data governance method based on the transformer multimodal fusion dataset.

[0013] Furthermore, the step 1 includes the following steps:

[0014] Step 1.1: Access the transformer's real-time data stream and historical batch data through stream-batch integration standardization and save them to the storage system;

[0015] Step 1.2: Dynamically map the stored data to the schema and save the result.

[0016] Moreover, the specific implementation method of step 1.1 is as follows: when accessing real-time data streams, access the real-time data streams of timestamps, voltage, current, and oil temperature through the Kafka message queue; when accessing historical batch data, upload the original data files to the HDFS storage system, access the batch data files of oil chromatograms, device numbers, and device status through the HDFS storage system, use Flink's batch processing engine to merge the historical files, and finally write the stream batch data into a unified HDFS storage system. The stream batch data includes: timestamps, device numbers, voltage, current, oil temperature, oil chromatogram data, and device status.

[0017] Moreover, the specific implementation method of the dynamic schema mapping in step 1.2 is as follows: the transformer data is of object type, containing various specific data attributes; the device code is of string type, the oil temperature, voltage, current, and gas content in the oil chromatogram are of numeric type, the timestamp is in ISO 8601 format, and the device status is a Boolean value. The data is verified when it enters the stream processing framework, and the real-time data stream and batch data are dynamically mapped through templates to ensure unified data format and smooth interaction under a unified data architecture.

[0018] Moreover, the identifier in step 2 is based on a composite primary key, where the primary key = device ID + timestamp, the ID identifies the transformer, and the timestamp records the data time point. For each piece of data, a unique primary key is generated.

[0019] Furthermore, the step 3 includes the following steps:

[0020] Step 3.1: Perform data preprocessing, identify matching transformers based on timestamps and device IDs, and normalize numerical data to scale it to the range of [0, 1]:

[0021]

[0022] in, is the original data, X max 、X min are the maximum and minimum values of the data respectively, X is the normalized data, ranging from [0,1];

[0023] For missing values in the data, the normalized mean data of the sliding window is used to fill them:

[0024]

[0025] Among them, X t is the data at the current moment, W is the window size, t is the current moment, X i For non-missing values in the window, only non-missing values in the window are used to calculate the mean;

[0026] Step 3.2: Divide the transformer data by time window to calculate the dynamic field radius;

[0027] The transformer data is divided into 15-minute time windows, and the data within the window is grouped by device ID. The distance from each data point in the window to the kth nearest neighbor is then calculated to generate a k-distance graph. The inflection point of the k-distance curve is selected as the candidate area radius value Eps. The inflection point corresponds to the area with significant density changes. At the same time, for windows with uniform data distribution, the mean k-distance is used as Eps. For windows with high density, the local quartile is used to avoid overfitting:

[0028]

[0029] Among them, d k is the k-distance of the data points, which is used to determine the local density of the data points. is the mean of the k-distances of all points in the sliding window, σ(d k ) is the standard deviation, is the adjustment coefficient, which is used when the data is evenly distributed. ;Q3(d k) is the third quartile of k-distance, Q1(d k ) is the first quartile of the k-distance, IQR= Q3- Q1 is the interquartile range, when the window is a high-density area, use ;

[0030] Step 3.3, improve DBSCAN clustering based on dynamic domain radius; calculate the difference value of the transformer data window by improving the DBSCAN clustering method;

[0031]

[0032] Among them, X t is the current moment data, X t-1 is the data of the previous moment, For the first-order difference, when the absolute value of the difference exceeds the threshold of 3 times the standard deviation, the data point at the current moment is marked as a candidate mutation point:

[0033]

[0034] Where θ is the threshold, To take the mean, To obtain the standard deviation; the difference result is then used as a new dimension and input into DBSCAN together with the original data. Under dynamic EPS, the clustering results are divided into three categories: core points, boundary points, and noise points; when a candidate mutation point is also marked as a noise point, it is confirmed as a mutation anomaly;

[0035] Step 3.4: Perform multimodal anomaly labeling based on dynamic EPS and improved DBSCAN clustering;

[0036] The multimodal anomaly marking method uses the Person correlation coefficient to calculate the correlation between different modes of the transformer to identify cross-modal correlation anomalies. The modal calculation is:

[0037]

[0038] Among them: O and Z are the data of two modal datasets, O i and Z i is the data point, and is the mean, r is the correlation coefficient, and its value range is [-1, 1]. When r is approximately equal to 0, it means there is no correlation. When r < 0.2 and the historical r > 0.8, it is marked as a cross-modal anomaly. Based on the data quality, modality importance and anomaly confidence, the weight is dynamically assigned to improve the robustness of anomaly detection;

[0039] Step 3.5, outlier processing and cleaning; after completing the anomaly marking, the outlier processing and cleaning are performed to improve the data quality; after deleting the outliers, the corresponding missing values are filled with the mean, and the mean filling method is still the sliding window mean filling, thus completing the transformer dataset management.

[0040] Furthermore, the step 4 includes the following steps:

[0041] Step 4.1: Time alignment: The input data is a cleaned transformer dataset, including voltage, current, oil temperature, and oil chromatogram gas content. Dynamic time warping (DTW) is used to eliminate inconsistencies in the data time steps:

[0042] Calculate the optimal alignment path to keep the time series consistent and achieve time alignment:

[0043]

[0044] Among them, D(i,j) is the cumulative distance between two time steps, |X i -Y j | is the Euclidean distance between two points, and the current cumulative distance is equal to the sum of the Euclidean distance between the two points and the cumulative distance of the smallest neighboring element that can reach the point. min() is the minimum function, which updates the alignment path by selecting the historical minimum cumulative distance to ensure reasonable matching of time steps.

[0045] Step 4.2: Build a temporal attention mechanism to allow the model to focus on important time steps;

[0046] The temporal attention mechanism calculates the attention weight at time step t:

[0047]

[0048] Then calculate the time series weighted features:

[0049]

[0050] Among them, h t is the feature vector of time step t, h i is the eigenvector of time step i, is the attention weight, W t is the weight matrix, T is the time dimension length of the input sequence, H temp It is a time series weighted feature, and exp() is an exponential function;

[0051] Step 4.3: Perform spatial alignment: The voltage, current, and temperature data of different transformers are distributed differently in physical locations. Use graph convolution GCN to model the spatial relationship, use the adjacency matrix A to represent the connection weights between the data, and calculate the spatial features:

[0052]

[0053] Among them, H (l+1) is the l+1 layer feature, H l is the l-th layer feature, W l is the learnable weight matrix of layer l, is a nonlinear activation function;

[0054] There is potential correlation between different modal data, so a cross-modal attention mechanism is designed to calculate the correlation between modal features and interactively fuse multimodal information;

[0055] The inter-modal feature correlation is calculated as:

[0056]

[0057] Compute cross-modal fusion:

[0058]

[0059] Among them, S is the inter-modal feature correlation, H cross For cross-modal fusion feature representation, Q, K, and V are the query, key, and value vectors in the attention calculation, respectively. d is the feature dimension, and Softmax() is a normalization function that makes the sum of attention weights equal to 1.

[0060] Finally, use multi-head attention for feature fusion:

[0061]

[0062] Among them, H final For the final fusion feature representation, each attention head (head1,…head n ) performs independent cross-modal interactions to fully learn different information subspaces, Concat() is the concatenation function, W O is the multi-head attention weight matrix, and n is the number of attention heads.

[0063] Moreover, the specific implementation method of step 5 is: divide the transformer data set into training set, validation set and test set in a ratio of 70%:15%:15%, and divide the multimodal data x at each moment into t The input is fed into a bidirectional LSTM network. The forward LSTM extracts past information, and the backward LSTM extracts future information. The output is then mapped to the label space through a fully connected layer. After the LSTM output, the CRF layer uses the contextual relationship to calculate the CRF score, improving the global consistency of label predictions.

[0064]

[0065] Where P is the entire label sequence (y1,y2,…,y step ), is the category score of the current time step predicted by LSTM, is the state transition fraction of adjacent time steps calculated by CRF; B is the normalization factor to ensure that the sum of probabilities is 1, and step is the total number of time steps;

[0066] The loss function uses the cross-entropy loss function (Cross-Entropy Loss) and CRF loss:

[0067]

[0068]

[0069] Among them, L CE is the cross entropy loss function, is the true label value, is the probability predicted by the model, L CRF is the CRF loss, log() is the logarithmic function;

[0070] The Adam optimizer was used with a learning rate of 0.001. The bidirectional LSTM-CRF model was trained by minimizing the combined loss of the cross entropy loss and the CRF loss function. The bidirectional LSTM-CRF model was evaluated using metrics such as F1_score, accuracy, precision, and recall, calculated using the following formula:

[0071]

[0072] Among them, F1_score is the F1 score, which measures the performance of the classification model; Precision is the precision rate, that is, the proportion of labels predicted as positive examples that are actually positive examples; Recall is the recall rate, which is also the proportion of labels that are actually positive examples that are predicted as positive examples.

[0073] The advantages and positive effects of the present invention are:

[0074] This invention standardizes access and dynamic data mapping for transformer multimodal data; constructs a unique identifier for the mapped transformer data; constructs a transformer status dataset based on the unique identifier and implements multi-strategy data governance; performs multimodal alignment and fusion on the cleaned dataset to obtain a transformer multimodal fusion dataset; and evaluates data governance methods based on the transformer multimodal fusion dataset. This invention enables standardized access, governance, and alignment and fusion of multimodal data within a unified framework, which is of great significance for improving the accuracy, reliability, and timeliness of power grid transformer status perception. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 is a flow chart of the present invention;

[0076] Figure 2 A flowchart of the present invention for constructing a transformer status data set based on a unique identifier to perform multi-strategy data governance;

[0077] Figure 3 This is a flowchart of the alignment and fusion of multimodal data of the present invention. DETAILED DESCRIPTION

[0078] The present invention is further described below in conjunction with the accompanying drawings.

[0079] A transformer data fusion and governance method based on multimodality, such as Figure 1 As shown, the following steps are included:

[0080] Step 1: Standardize access and dynamically map transformer multimodal data.

[0081] Step 1.1: Access the transformer's real-time data stream and historical batch data through stream-batch integration standardization and save them to the storage system.

[0082] The specific implementation method of step 1.1 is as follows: when accessing real-time data streams, access the real-time data streams of timestamps, voltage, current, and oil temperature through the Kafka message queue; when accessing historical batch data, upload the original data files to the HDFS storage system, access the batch data files of oil chromatograms, device numbers, and device status through the HDFS storage system, use Flink's batch processing engine to merge the historical files, and finally write the stream batch data (including timestamps, device numbers, voltage, current, oil temperature, oil chromatogram data, and device status) to a unified HDFS storage system.

[0083] Step 1.2: Dynamically map the stored data to the schema and save the result.

[0084] The specific implementation method for dynamic schema mapping in step 1.2 is as follows: transformer data is of object type, containing various specific data attributes; the device code is of string type, oil temperature, voltage, current, and gas content in the oil chromatogram are of numeric type, the timestamp is in ISO 8601 format, and the device status is a Boolean value. Data is verified when it enters the stream processing framework, and real-time data streams and batch data are dynamically mapped through templates to ensure unified data formats and smooth interaction under a unified data architecture.

[0085] Step 2: Construct a unique identifier for the mapped transformer data.

[0086] The identifier in step 2 is based on a composite primary key, where the primary key = device ID + timestamp. The ID identifies the transformer, and the timestamp records the data time point. For each piece of data, a unique primary key is generated.

[0087] Step 3: Build a transformer status dataset based on the unique identifier, perform multi-strategy data governance, and complete the cleaning and governance of the status dataset.

[0088] like Figure 2 As shown, this step includes the following steps:

[0089] Step 3.1: Perform data preprocessing, identify matching transformers based on timestamps and device IDs, and normalize numerical data to scale it to the range of [0, 1]:

[0090]

[0091] in, is the original data, X max 、X min are the maximum and minimum values of the data respectively, X is the normalized data, ranging from [0,1];

[0092] For missing values in the data, the normalized mean data of the sliding window is used to fill them:

[0093] Among them, X t is the data at the current moment, W is the window size, t is the current moment, X i For non-missing values in the window, only non-missing values in the window are used to calculate the mean;

[0094] Step 3.2: Divide the transformer data by time window to calculate the dynamic field radius;

[0095] The transformer data is divided into 15-minute time windows, and the data within the window is grouped by device ID. The distance from each data point in the window to the kth nearest neighbor is then calculated to generate a k-distance graph. The inflection point of the k-distance curve is selected as the candidate area radius value Eps. The inflection point corresponds to the area with significant density changes. At the same time, for windows with uniform data distribution, the mean k-distance is used as Eps. For windows with high density, the local quartile is used to avoid overfitting:

[0096]

[0097] Among them, d k is the k-distance of the data points, which is used to determine the local density of the data points. is the mean of the k-distances of all points in the sliding window, is the standard deviation, is the adjustment coefficient, which is used when the data is evenly distributed. ;Q3(d k ) is the third quartile of k-distance, Q1(d k ) is the first quartile of the k-distance, IQR= Q3- Q1 is the interquartile range, when the window is a high-density area, use ;

[0098] Step 3.3: Improve DBSCAN clustering based on dynamic domain radius; calculate the differential value of the transformer data window by improving the DBSCAN clustering method.

[0099]

[0100] Among them, X t is the current moment data, X t-1 is the data of the previous moment, For the first-order difference, when the absolute value of the difference exceeds the threshold of 3 times the standard deviation, the data point at the current moment is marked as a candidate mutation point:

[0101]

[0102] Where θ is the threshold, To take the mean, To obtain the standard deviation; the difference result is then used as a new dimension and input into DBSCAN together with the original data. Under dynamic EPS, the clustering results are divided into three categories: core points, boundary points, and noise points; when a candidate mutation point is also marked as a noise point, it is confirmed as a mutation anomaly;

[0103] Step 3.4: Perform multimodal anomaly labeling based on dynamic EPS and improved DBSCAN clustering;

[0104] The multimodal anomaly marking method uses the Person correlation coefficient to calculate the correlation between different modes of the transformer to identify cross-modal correlation anomalies. The modal calculation is:

[0105]

[0106] Among them: O and Z are the data of two modal datasets, O i and Z i is the data point, and is the mean, r is the correlation coefficient, and its value range is [-1, 1]. When r is approximately equal to 0, it means there is no correlation. When r < 0.2 and the historical r > 0.8, it is marked as a cross-modal anomaly. Based on the data quality, modality importance and anomaly confidence, the weight is dynamically assigned to improve the robustness of anomaly detection;

[0107] Step 3.5: Outlier processing and cleaning.

[0108] After outlier marking is completed, outlier processing and cleaning are performed to improve data quality. After deleting outliers, the corresponding missing values are filled using the mean, and the mean filling method is still the sliding window mean filling, thus completing the transformer dataset governance.

[0109] Step 4: Based on the cleaned dataset, perform multimodal alignment and fusion to obtain the transformer multimodal fusion dataset.

[0110] like Figure 3 As shown, this step includes the following steps

[0111] Step 4.1: Time alignment: The input data is a cleaned transformer dataset, including voltage, current, oil temperature, and oil chromatogram gas content. Dynamic time warping (DTW) is used to eliminate inconsistencies in the data time steps:

[0112] Calculate the optimal alignment path to keep the time series consistent and achieve time alignment:

[0113]

[0114] Among them, D(i,j) is the cumulative distance between two time steps, |X i -Y j | is the Euclidean distance between two points, and the current cumulative distance is equal to the sum of the Euclidean distance between the two points and the cumulative distance of the smallest neighboring element that can reach the point. min() is the minimum function, which updates the alignment path by selecting the historical minimum cumulative distance to ensure reasonable matching of time steps.

[0115] Step 4.2: Build a temporal attention mechanism to allow the model to focus on important time steps;

[0116] The temporal attention mechanism calculates the attention weight at time step t:

[0117]

[0118] Then calculate the time series weighted features:

[0119]

[0120] Among them, h t is the feature vector of time step t, h i is the eigenvector of time step i, is the attention weight, W t is the weight matrix, T is the time dimension length of the input sequence, H temp It is a time series weighted feature, and exp() is an exponential function;

[0121] Step 4.3: Perform spatial alignment: The voltage, current, and temperature data of different transformers are distributed differently in physical locations. Use graph convolution GCN to model the spatial relationship, use the adjacency matrix A to represent the connection weights between the data, and calculate the spatial features:

[0122]

[0123] Among them, H (l+1) is the l+1 layer feature, H l is the l-th layer feature, W l is the learnable weight matrix of layer l, is a nonlinear activation function;

[0124] There is potential correlation between different modal data, so a cross-modal attention mechanism is designed to calculate the correlation between modal features and interactively fuse multimodal information;

[0125] The inter-modal feature correlation is calculated as:

[0126]

[0127] Compute cross-modal fusion:

[0128]

[0129] Among them, S is the inter-modal feature correlation, H cross For cross-modal fusion feature representation, Q, K, and V are the query, key, and value vectors in the attention calculation, respectively. d is the feature dimension, and Softmax() is a normalization function that makes the sum of attention weights equal to 1.

[0130] Finally, use multi-head attention for feature fusion:

[0131]

[0132] Among them, H final For the final fusion feature representation, each attention head (head1,…head n ) performs independent cross-modal interactions to fully learn different information subspaces, Concat() is the concatenation function, W O is the multi-head attention weight matrix, and n is the number of attention heads.

[0133] Step 5: Evaluate the data governance method based on the transformer multimodal fusion dataset.

[0134] The transformer dataset is divided into training set, validation set and test set in the ratio of 70%:15%:15%, and the multimodal data x at each moment is tThe input is fed into a bidirectional LSTM network. The forward LSTM extracts past information, and the backward LSTM extracts future information. The output is then mapped to the label space through a fully connected layer. After the LSTM output, the CRF layer uses the contextual relationship to calculate the CRF score, improving the global consistency of label predictions.

[0135]

[0136] Where P is the entire label sequence (y1,y2,…,y step ), is the category score of the current time step predicted by LSTM, is the state transition fraction of adjacent time steps calculated by CRF; B is the normalization factor to ensure that the sum of probabilities is 1, and step is the total number of time steps;

[0137] The loss function uses the cross-entropy loss function (Cross-Entropy Loss) and CRF loss:

[0138]

[0139]

[0140] Among them, L CE is the cross entropy loss function, is the true label value, is the probability predicted by the model, L CRF is the CRF loss, log() is the logarithmic function;

[0141] The Adam optimizer was used with a learning rate of 0.001. The bidirectional LSTM-CRF model was trained by minimizing the combined loss of the cross entropy loss and the CRF loss function. The bidirectional LSTM-CRF model was evaluated using metrics such as F1_score, accuracy, precision, and recall, calculated using the following formula:

[0142]

[0143] Among them, F1_score is the F1 score, which measures the performance of the classification model; Precision is the precision rate, that is, the proportion of labels predicted as positive examples that are actually positive examples; Recall is the recall rate, which is also the proportion of labels that are actually positive examples that are predicted as positive examples.

[0144] It should be emphasized that the embodiments described in the present invention are illustrative rather than restrictive. Therefore, the present invention includes but is not limited to the embodiments described in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solutions of the present invention also fall within the scope of protection of the present invention.

Claims

1. A transformer data fusion and management method based on multimodality, characterized by: The following steps are involved: Step 1: Standardize access and dynamically map transformer multimodal data; Step 2: Construct a unique identifier for the mapped transformer data; Step 3: Construct a transformer status dataset based on the unique identifier, perform multi-strategy data management, and complete the cleaning and management of the status dataset; Step 4: Based on the cleaned dataset, perform multimodal alignment and fusion to obtain the transformer multimodal fusion dataset; Step 5: Evaluate the data governance method based on the transformer multimodal fusion dataset.

2. The method for transformer data fusion and management based on multimodality according to claim 1, characterized in that: The step 1 comprises the following steps: Step 1.1: Access the transformer's real-time data stream and historical batch data through stream-batch integration standardization and save them to the storage system; Step 1.2: Dynamically map the stored data to the schema and save the result.

3. The method for transformer data fusion and management based on multimodality according to claim 2, characterized in that: The specific implementation method of step 1.1 is as follows: when accessing real-time data streams, access the real-time data streams of timestamps, voltage, current, and oil temperature through the Kafka message queue; when accessing historical batch data, upload the original data files to the HDFS storage system, access the batch data files of oil chromatograms, device numbers, and device status through the HDFS storage system, use Flink's batch processing engine to merge the historical files, and finally write the stream batch data into a unified HDFS storage system. The stream batch data includes: timestamps, device numbers, voltage, current, oil temperature, oil chromatogram data, and device status.

4. The method for transformer data fusion and management based on multimodality according to claim 2, characterized in that: The specific implementation method of the dynamic schema mapping in step 1.2 is as follows: the transformer data is of object type, containing various specific data attributes; the device code is of string type, the oil temperature, voltage, current, and gas content in the oil chromatogram are of numeric type, the timestamp is in ISO 8601 format, and the device status is a Boolean value. The data is verified when it enters the stream processing framework, and the real-time data stream and batch data are dynamically mapped through templates to ensure unified data format and smooth interaction under a unified data architecture.

5. The method for transformer data fusion and management based on multimodality according to claim 1, characterized in that: The identifier in step 2 is based on a composite primary key, where the primary key = device ID + timestamp, the ID identifies the transformer, and the timestamp records the data time point. For each piece of data, a unique primary key is generated.

6. The method for transformer data fusion and management based on multimodality according to claim 1, characterized in that: The step 3 comprises the following steps: Step 3.1: Perform data preprocessing, identify matching transformers based on timestamps and device IDs, and normalize numerical data to scale it to the range of [0, 1]: ; in, is the original data, X max 、X min are the maximum and minimum values of the data respectively, X is the normalized data, ranging from [0,1]; For missing values in the data, the normalized mean data of the sliding window is used to fill them: ; Among them, X t is the data at the current moment, W is the window size, t is the current moment, X i For non-missing values in the window, only non-missing values in the window are used to calculate the mean; Step 3.2: Divide the transformer data by time window to calculate the dynamic field radius; The transformer data is divided into 15-minute time windows, and the data within the window is grouped by device ID. The distance from each data point in the window to the kth nearest neighbor is then calculated to generate a k-distance graph. The inflection point of the k-distance curve is selected as the candidate area radius value Eps. The inflection point corresponds to the area with significant density changes. At the same time, for windows with uniform data distribution, the mean k-distance is used as Eps. For windows with high density, the local quartile is used to avoid overfitting: ; Among them, d k is the k-distance of the data points, which is used to determine the local density of the data points. is the mean of the k-distances of all points in the sliding window, is the standard deviation, is the adjustment coefficient, which is used when the data is evenly distributed. ;Q3(d k ) is the third quartile of k-distance, Q1(d k ) is the first quartile of the k-distance, IQR= Q3- Q1 is the interquartile range, when the window is a high-density area, use ; Step 3.3, improve DBSCAN clustering based on dynamic domain radius; calculate the difference value of the transformer data window by improving the DBSCAN clustering method; ; Among them, X t is the current moment data, X t-1 is the data of the previous moment, For the first-order difference, when the absolute value of the difference exceeds the threshold of 3 times the standard deviation, the data point at the current moment is marked as a candidate mutation point: ； in, is the threshold, To take the mean, To obtain the standard deviation; the difference result is then used as a new dimension and input into DBSCAN together with the original data. Under dynamic EPS, the clustering results are divided into three categories: core points, boundary points, and noise points; when a candidate mutation point is also marked as a noise point, it is confirmed as a mutation anomaly; Step 3.4: Perform multimodal anomaly labeling based on dynamic EPS and improved DBSCAN clustering; The multimodal anomaly marking method uses the Person correlation coefficient to calculate the correlation between different modes of the transformer to identify cross-modal correlation anomalies. The modal calculation is: ; Among them: O and Z are the data of two modal datasets, O i and Z i is the data point, and is the mean, r is the correlation coefficient, and its value range is [-1, 1]. When r is approximately equal to 0, it means there is no correlation. When r < 0.2 and the historical r > 0.8, it is marked as a cross-modal anomaly. Based on the data quality, modality importance and anomaly confidence, the weight is dynamically assigned to improve the robustness of anomaly detection; Step 3.5, outlier processing and cleaning; after completing the anomaly marking, the outlier processing and cleaning are performed to improve the data quality; after deleting the outliers, the corresponding missing values are filled with the mean, and the mean filling method is still the sliding window mean filling, thus completing the transformer dataset management.

7. The method for transformer data fusion and management based on multimodality according to claim 1, characterized in that: The step 4 comprises the following steps: Step 4.1: Time alignment: The input data is a cleaned transformer dataset, including voltage, current, oil temperature, and oil chromatogram gas content. Dynamic time warping (DTW) is used to eliminate inconsistencies in the data time steps: Calculate the optimal alignment path to keep the time series consistent and achieve time alignment: ; Among them, D(i,j) is the cumulative distance between two time steps, |X i -Y j | is the Euclidean distance between two points, and the current cumulative distance is equal to the sum of the Euclidean distance between the two points and the cumulative distance of the smallest neighboring element that can reach the point. min() is the minimum function, which updates the alignment path by selecting the historical minimum cumulative distance to ensure reasonable matching of time steps. Step 4.2: Build a temporal attention mechanism to allow the model to focus on important time steps; The temporal attention mechanism calculates the attention weight at time step t: ; Then calculate the time series weighted features: ; Among them, h t is the feature vector of time step t, h i is the eigenvector of time step i, is the attention weight, W t is the weight matrix, T is the time dimension length of the input sequence, H temp It is a time series weighted feature, and exp() is an exponential function; Step 4.3: Perform spatial alignment: The voltage, current, and temperature data of different transformers are distributed differently in physical locations. Use graph convolution GCN to model the spatial relationship, use the adjacency matrix A to represent the connection weights between the data, and calculate the spatial features: ; Among them, H (l+1) is the l+1 layer feature, H l is the l-th layer feature, W l is the learnable weight matrix of layer l, is a nonlinear activation function; There is potential correlation between different modal data, so a cross-modal attention mechanism is designed to calculate the correlation between modal features and interactively fuse multimodal information; The inter-modal feature correlation is calculated as: ; Compute cross-modal fusion: ; Among them, S is the inter-modal feature correlation, H cross For cross-modal fusion feature representation, Q, K, and V are the query, key, and value vectors in the attention calculation, respectively. d is the feature dimension, and Softmax() is a normalization function that makes the sum of attention weights equal to 1. Finally, use multi-head attention for feature fusion: ; Among them, H final For the final fusion feature representation, each attention head (head1,…head n ) performs independent cross-modal interactions to fully learn different information subspaces, Concat() is the concatenation function, W O is the multi-head attention weight matrix, and n is the number of attention heads.

8. The method for transformer data fusion and management based on multimodality according to claim 1, characterized in that: The specific implementation method of step 5 is: divide the transformer dataset into training set, validation set and test set according to the ratio of 70%:15%:15%, and divide the multimodal data x at each moment into t The input is fed into a bidirectional LSTM network. The forward LSTM extracts past information, and the backward LSTM extracts future information. The output is then mapped to the label space through a fully connected layer. After the LSTM output, the CRF layer uses the contextual relationship to calculate the CRF score, improving the global consistency of label predictions. ; Where P is the entire label sequence (y1,y2,…,y step ), is the category score of the current time step predicted by LSTM, is the state transition fraction of adjacent time steps calculated by CRF; B is the normalization factor to ensure that the sum of probabilities is 1, and step is the total number of time steps; The loss function uses the cross-entropy loss function (Cross-Entropy Loss) and CRF loss: ; ; Among them, L CE is the cross entropy loss function, is the true label value, is the probability predicted by the model, L CRF is the CRF loss, log() is the logarithmic function; The Adam optimizer was used with a learning rate of 0.

001. The bidirectional LSTM-CRF model was trained by minimizing the combined loss of the cross entropy loss and the CRF loss function. The bidirectional LSTM-CRF model was evaluated using metrics such as F1_score, accuracy, precision, and recall, calculated using the following formula: ; Among them, F1_score is the F1 score, which measures the performance of the classification model; Precision is the precision rate, that is, the proportion of labels predicted as positive examples that are actually positive examples; Recall is the recall rate, which is also the proportion of labels that are actually positive examples that are predicted as positive examples.

Images