A multi-modal based transformer data fusion and management method

Abstract

The present application relates to a kind of transformer data fusion and management method based on multi-modal, by standardization access and data dynamic mapping to transformer multi-modal data;Unique identifier of mapping transformer data is constructed;According to unique identifier, transformer state data set is constructed, multi-strategy data management is carried out, and the cleaning management of state data set is completed;Based on the data set after cleaning, multi-modal alignment fusion is carried out, and transformer multi-modal fusion data set is obtained;Based on transformer multi-modal fusion data set, data management method is evaluated.The present application can realize the standardization access, management and alignment fusion of multi-modal data under unified framework, and it has important significance to improve the accuracy, reliability and timeliness of power grid transformer state perception.

Description

Technical Field

[0001] This invention belongs to the field of transformer diagnostic technology, and in particular to a method for transformer data fusion and management based on multimodal methods. Background Technology

[0002] With the continuous expansion and increasing complexity of power systems, transformers, as crucial equipment, directly impact the safety and stability of the power grid. Introducing effective transformer data governance methods allows for in-depth analysis of operational data, identifying potential risks, optimizing maintenance strategies, and thus improving the intelligence level of equipment operation and maintenance. Against this backdrop, the acquisition, fusion, and governance of multi-source, multi-modal data has gradually become a research hotspot. However, the diverse data sources and complex data types present challenges in achieving standardized access, alignment, and fusion of multi-modal transformer data, as well as effective governance of data quality issues.

[0003] The use of scenario-specific manual rule-based methods is fundamental to the governance of multi-source transformer data. However, rule formulation 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 achieve multi-source data fusion and governance at the feature layer. While these methods can achieve data governance, they suffer from insufficient data standardization when dealing with multimodal and spatiotemporally correlated transformer state-aware data, making it difficult to achieve unified data description and processing. Transformer state-aware data includes multimodal information such as voltage, current, temperature, and oil chromatographic gases. This data comes from different sensors, and the sampling frequency and timestamps may be inconsistent, leading to spatiotemporal misalignment. Cross-modal correlation anomalies exist between different modal data. Traditional anomaly detection methods, such as statistical analysis, are effective in univariate scenarios but struggle to handle multimodal and spatiotemporally correlated anomaly patterns. Existing methods are unable to address these problems in multimodal data governance, resulting in loose data integration and an inability to fully explore the intrinsic connections between multimodal data.

[0004] Therefore, researching 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 for improving the accuracy, reliability, and timeliness of power grid transformer state perception. Summary of the Invention

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

[0006] The technical problem solved by this invention is achieved through the following technical solution:

[0007] A multimodal transformer data fusion and governance method includes the following steps:

[0008] Step 1: Standardize and dynamically map the transformer multi-mode data;

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

[0010] Step 3: Construct a transformer state dataset based on the unique identifier, perform multi-strategy data governance, and complete the cleaning and governance of the state 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, step 1 includes the following steps:

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

[0015] Step 1.2: Perform dynamic schema mapping on the stored data and save the results.

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

[0017] Furthermore, the specific implementation method of dynamic schema mapping in step 1.2 is as follows: transformer data is an object type, containing various specific data attributes; equipment code is a string type; oil temperature, voltage, current, and gas content in oil chromatography are numerical types; timestamp is in ISO 8601 format; equipment status is a Boolean value; data is validated when entering the stream processing framework; real-time data stream and batch data are dynamically mapped through templates to ensure data format uniformity and smooth interaction under a unified data architecture.

[0018] Furthermore, 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, step 3 includes the following steps:

[0020] Step 3.1: Perform data preprocessing. Identify and match transformers based on timestamps and device IDs, and normalize numerical data, scaling it to the range [0,1].

[0021]

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

[0023] For missing values ​​in the data, normalized mean data from a sliding window is used for imputation:

[0024]

[0025] Among them, X t The data is at the current time, W is the window size, t is the current time, and X is the current time. i For non-missing values ​​within the window, only non-missing values ​​within the window are used to calculate the mean;

[0026] Step 3.2: Calculate the dynamic range radius by dividing the transformer data into time windows;

[0027] Transformer data is divided into 15-minute time windows, with data grouped by device ID within each window. The distance from each data point within the window to its k-nearest neighbor is then calculated, generating a k-distance plot. The inflection point of the k-distance curve is selected as the candidate neighborhood radius value, Eps, corresponding to areas of significant density change. For windows with uniform data distribution, the mean k-distance is used as Eps; for windows in high-density regions, local quartiles are used to avoid overfitting.

[0028]

[0029] Where, d k The k-distance is used to determine the local density of data points. It is the mean k-distance of all points within the sliding window, σ(d k ) is the standard deviation. This is an adjustment coefficient used when the data is uniformly distributed. ;Q3(d kQ1(d) is the third quartile of the k-distance. k Q1 is the first quartile of the k-distance, and IQR = Q3 - Q1 is the interquartile range. When the window is a high-density region, the interquartile range is used. ;

[0030] Step 3.3: Improve DBSCAN clustering based on dynamic neighborhood radius; calculate the difference value of the transformer data window by improving the DBSCAN clustering method;

[0031]

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

[0033]

[0034] Where θ is the threshold. To take the average, To obtain the standard deviation, the difference results are then used as a new dimension and input into DBSCAN along 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 labeling method uses the Person correlation coefficient to calculate the correlation between different modes of the transformer to identify cross-modal anomalies. The mode calculation is as follows:

[0037]

[0038] Where: O and Z are data from two modality datasets, respectively, O i and Z i For data points, and The mean is denoted as r, which is the correlation coefficient, ranging from -1 to 1. When r is approximately equal to 0, it indicates no correlation. When r < 0.2 and the historical r > 0.8, it is marked as a cross-modal anomaly. Based on data quality, modal importance, and anomaly confidence, weights are dynamically allocated to improve the robustness of anomaly detection.

[0039] Step 3.5, Outlier Handling and Cleaning: After outlier marking, outlier handling and cleaning are performed to improve data quality. After deleting outliers, the corresponding missing values ​​are filled with the mean. The mean filling method is still sliding window mean filling, thus completing the transformer dataset governance.

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

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

[0042] Calculate the optimal alignment path to keep the time series consistent, thus achieving time alignment:

[0043]

[0044] Where 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 value function, which updates the alignment path by selecting the historical minimum cumulative distance to ensure reasonable matching of time steps;

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

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

[0047]

[0048] Then, the time-weighted features are calculated:

[0049]

[0050] Among them, h t h is the eigenvector at time step t. i Let i be the feature vector at time step i. For attention weights, W t Here, H is the weight matrix, T is the time dimension length of the input sequence, and H is the weight matrix. temp It is a time-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 location. Graph Convolutional CN is used to model spatial relationships, and an adjacency matrix A is used to represent the connection weights between data points. Spatial features are then calculated.

[0052]

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

[0054] There are potential correlations between data from different modalities. Therefore, a cross-modal attention mechanism is designed to calculate the correlation between features of different modalities and interactively fuse multimodal information.

[0055] The correlation between modal features is calculated as follows:

[0056]

[0057] Computing cross-modal fusion:

[0058]

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

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

[0061]

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

[0063] Furthermore, the specific implementation method of step 5 is as follows: the transformer dataset is divided into a training set, a validation set, and a test set according to a ratio of 70%:15%:15%, and the multimodal data x at each time step is... t The input is fed into a bidirectional LSTM network. The forward LSTM extracts past information, and the backward LSTM extracts future information. A fully connected layer maps the output to the label space. After the LSTM output, a CRF layer utilizes contextual relationships to calculate the CRF score, improving the global consistency of label prediction.

[0064]

[0065] Where P is the entire label sequence (y1, y2, ..., y step The conditional probability of ). It is the class score predicted by LSTM at the current time step. is the state transition score between 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 functions used are cross-entropy loss and CRF loss:

[0067]

[0068]

[0069] Among them, L CE Let cross-entropy be the loss function. The actual label value. L is the probability predicted by the model. CRF For CRF loss, log() is the logarithmic function;

[0070] Using the Adam optimizer with a learning rate of 0.001, a bidirectional LSTM-CRF model was trained by minimizing the joint 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] Here, F1_score is the F1 score, which measures the performance of the classification model; Precision is the accuracy, which is the proportion of labels predicted as positive that are actually positive; Recall is the recall, which is also the proportion of labels that are actually positive that are predicted as positive.

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

[0074] This invention standardizes and dynamically maps transformer multimodal data; constructs unique identifiers for the mapped transformer data; builds a transformer state dataset based on these unique identifiers and performs multi-strategy data governance; performs multimodal alignment and fusion on the cleaned dataset to obtain a transformer multimodal fusion dataset; and evaluates the data governance methods based on the transformer multimodal fusion dataset. This invention enables standardized access, governance, and alignment fusion of multimodal data within a unified framework, which is of great significance for improving the accuracy, reliability, and timeliness of power grid transformer state perception. Attached Figure Description

[0075] Figure 1 This is a flowchart of the present invention;

[0076] Figure 2 This is a flowchart illustrating the multi-strategy data governance method for constructing a transformer state dataset based on a unique identifier, as described in this invention.

[0077] Figure 3 This is a flowchart illustrating the alignment and fusion of multimodal data according to the present invention. Detailed Implementation

[0078] The present invention will be further described in detail below with reference to the accompanying drawings.

[0079] A multimodal transformer data fusion and governance method, such as Figure 1 As shown, it includes the following steps:

[0080] Step 1: Standardize the access and dynamically map the transformer multi-mode data.

[0081] Step 1.1: Access the real-time data stream and historical batch data of the transformer through the unified batch processing 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 real-time data streams of timestamps, voltage, current, and oil temperature through a Kafka message queue; when accessing historical batch data, upload the original data files to the HDFS storage system, access batch data files of oil chromatography, equipment number, and equipment status through the HDFS storage system, merge the historical files using Flink's batch processing engine, and finally write the stream and batch data into a unified HDFS storage system. The stream and batch data include: timestamps, equipment number, voltage, current, oil temperature, oil chromatography data, and equipment status.

[0083] Step 1.2: Perform dynamic schema mapping on the stored data and save the results.

[0084] The specific implementation method of dynamic schema mapping in step 1.2 is as follows: transformer data is an object type, containing various specific data attributes; equipment code is a string type; oil temperature, voltage, current, and gas content in oil chromatography are numerical types; timestamp is in ISO 8601 format; equipment status is a boolean value; data is validated when entering the stream processing framework; real-time data stream and batch data are dynamically mapped through templates to ensure data format uniformity and smooth interaction under a unified data architecture.

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

[0086] In step 2, the identifier 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: Construct a transformer state dataset based on the unique identifier, perform multi-strategy data governance, and complete the cleaning and governance of the state dataset.

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

[0089] Step 3.1: Perform data preprocessing. Identify and match transformers based on timestamps and device IDs, and normalize numerical data, scaling it to the range [0,1].

[0090]

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

[0092] For missing values ​​in the data, normalized mean data from a sliding window is used for imputation:

[0093] Among them, X t The data is at the current time, W is the window size, t is the current time, and X is the current time. i For non-missing values ​​within the window, only non-missing values ​​within the window are used to calculate the mean;

[0094] Step 3.2: Calculate the dynamic range radius by dividing the transformer data into time windows;

[0095] Transformer data is divided into 15-minute time windows, with data grouped by device ID within each window. The distance from each data point within the window to its k-nearest neighbor is then calculated, generating a k-distance plot. The inflection point of the k-distance curve is selected as the candidate neighborhood radius value, Eps, corresponding to areas of significant density change. For windows with uniform data distribution, the mean k-distance is used as Eps; for windows in high-density regions, local quartiles are used to avoid overfitting.

[0096]

[0097] Where, d k The k-distance is used to determine the local density of data points. It is the mean of the k-distances of all points within the sliding window. It is the standard deviation. This is an adjustment coefficient used when the data is uniformly distributed. ;Q3(d k Q1(d) is the third quartile of the k-distance. k Q1 is the first quartile of the k-distance, and IQR = Q3 - Q1 is the interquartile range. When the window is a high-density region, the interquartile range is used. ;

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

[0099]

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

[0101]

[0102] Where θ is the threshold. To take the average, To obtain the standard deviation, the difference results are then used as a new dimension and input into DBSCAN along 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 labeling method uses the Person correlation coefficient to calculate the correlation between different modes of the transformer to identify cross-modal anomalies. The mode calculation is as follows:

[0105]

[0106] Where: O and Z are data from two modality datasets, respectively, O i and Z i For data points, and The mean is denoted as r, which is the correlation coefficient, ranging from -1 to 1. When r is approximately equal to 0, it indicates no correlation. When r < 0.2 and the historical r > 0.8, it is marked as a cross-modal anomaly. Based on data quality, modal importance, and anomaly confidence, weights are dynamically allocated to improve the robustness of anomaly detection.

[0107] Step 3.5: Outlier handling and cleaning.

[0108] After outlier labeling, outlier processing and cleaning are performed to improve data quality. After deleting outliers, the corresponding missing values ​​are imputed using the mean, employing a sliding window mean imputation method, thus completing the governance of the transformer dataset.

[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: Perform time alignment: The input data is the cleaned transformer dataset, including voltage, current, oil temperature, and oil chromatographic gas content. Dynamic Time Warping (DTW) is used to eliminate inconsistencies in the data's time steps.

[0112] Calculate the optimal alignment path to keep the time series consistent, thus achieving time alignment:

[0113]

[0114] Where 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 value function, which updates the alignment path by selecting the historical minimum cumulative distance to ensure reasonable matching of time steps;

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

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

[0117]

[0118] Then, the time-weighted features are calculated:

[0119]

[0120] Among them, h t h is the eigenvector at time step t. i Let i be the feature vector at time step i. For attention weights, W t Here, H is the weight matrix, T is the time dimension length of the input sequence, and H is the weight matrix. temp It is a time-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 location. Graph Convolutional CN is used to model spatial relationships, and an adjacency matrix A is used to represent the connection weights between data points. Spatial features are then calculated.

[0122]

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

[0124] There are potential correlations between data from different modalities. Therefore, a cross-modal attention mechanism is designed to calculate the correlation between features of different modalities and interactively fuse multimodal information.

[0125] The correlation between modal features is calculated as follows:

[0126]

[0127] Computing cross-modal fusion:

[0128]

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

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

[0131]

[0132] Among them, H final For the final fused feature representation, each attention head (head1,…head) n ) Performs independent cross-modal interactions, enabling different information subspaces to learn fully. Concat() is the concatenation function. O It is a multi-head attention weight matrix, where 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 was divided into training, validation, and test sets in a 70%:15%:15% ratio. The multimodal data at each time step was then processed. tThe input is fed into a bidirectional LSTM network. The forward LSTM extracts past information, and the backward LSTM extracts future information. A fully connected layer maps the output to the label space. After the LSTM output, a CRF layer utilizes contextual relationships to calculate the CRF score, improving the global consistency of label prediction.

[0135]

[0136] Where P is the entire label sequence (y1, y2, ..., y step The conditional probability of ). It is the class score predicted by LSTM at the current time step. is the state transition score between 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 functions used are cross-entropy loss and CRF loss:

[0138]

[0139]

[0140] Among them, L CE Let cross-entropy be the loss function. The actual label value. L is the probability predicted by the model. CRF For CRF loss, log() is the logarithmic function;

[0141] Using the Adam optimizer with a learning rate of 0.001, a bidirectional LSTM-CRF model was trained by minimizing the joint 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] Here, F1_score is the F1 score, which measures the performance of the classification model; Precision is the accuracy, which is the proportion of labels predicted as positive that are actually positive; Recall is the recall, which is also the proportion of labels that are actually positive that are predicted as positive.

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

Claims

1. A method for transformer data fusion and governance based on multimodal processing, characterized in that: Includes the following steps: Step 1: Standardize and dynamically map the transformer multi-mode data; Step 2: Construct a unique identifier for the mapped transformer data; Step 3: Construct a transformer state dataset based on the unique identifier, perform multi-strategy data governance, and complete the cleaning and governance of the state dataset; Step 3.1: Perform data preprocessing. Identify and match transformers based on timestamps and device IDs, and normalize numerical data, scaling it to the range [0,1]. ； in, For the original data, X max X min These are the maximum and minimum values ​​of the data, respectively, and X is the normalized data, ranging from [0,1]. For missing values ​​in the data, normalized mean data from a sliding window is used for imputation: ； Among them, X t The data is at the current time, W is the window size, t is the current time, and X is the current time. i For non-missing values ​​within the window, only non-missing values ​​within the window are used to calculate the mean; Step 3.2: Calculate the dynamic range radius by dividing the transformer data into time windows; Transformer data is divided into 15-minute time windows, with data grouped by device ID within each window. The distance from each data point within the window to its k-nearest neighbor is then calculated, generating a k-distance plot. The inflection point of the k-distance curve is selected as the candidate neighborhood radius value, Eps, corresponding to areas of significant density change. For windows with uniform data distribution, the mean k-distance is used as Eps; for windows in high-density regions, local quartiles are used to avoid overfitting. ； Where, d k The k-distance is used to determine the local density of data points. It is the mean of the k-distances of all points within the sliding window. It is the standard deviation. This is an adjustment coefficient used when the data is uniformly distributed. ;Q3(d k Q1(d) is the third quartile of the k-distance. k Q1 is the first quartile of the k-distance, and IQR = Q3 - Q1 is the interquartile range. When the window is a high-density region, the interquartile range is used. ; Step 3.3: Improve DBSCAN clustering based on dynamic neighborhood radius; calculate the difference value of the transformer data window by improving the DBSCAN clustering method; ； Among them, X t For the data at the current moment, X t-1 This is the data from the previous moment. For first-order differencing, when the absolute value of the difference exceeds a threshold of 3 times the standard deviation, the data point at the current time step is marked as a candidate mutation point: ； in, For the threshold, To take the average, To obtain the standard deviation, the difference results are then used as a new dimension and input into DBSCAN along 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 labeling method uses the Person correlation coefficient to calculate the correlation between different modes of the transformer to identify cross-modal anomalies. The mode calculation is as follows: ； Where: O and Z are data from two modality datasets, respectively, O i and Z i For data points, and The mean is denoted as r, which is the correlation coefficient, ranging from -1 to 1. When r equals 0, it indicates no correlation. When r < 0.2 and the historical r > 0.8, it is marked as a cross-modal anomaly. Based on data quality, modal importance, and anomaly confidence, weights are dynamically allocated to improve the robustness of anomaly detection. Step 3.5, Outlier Handling and Cleaning: After outlier marking, outlier handling and cleaning are performed to improve data quality. After deleting outliers, the corresponding missing values ​​are filled with the mean. The mean filling method is still sliding window mean filling, thus completing the transformer dataset governance. 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 governance based on multimodal processing according to claim 1, characterized in that: Step 1 includes the following steps: Step 1.1: Access the transformer's real-time data stream and historical batch data through the unified batch processing system and save them to the storage system; Step 1.2: Perform dynamic schema mapping on the stored data and save the results.

3. The method for transformer data fusion and governance based on multimodal processing 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 real-time data streams of timestamps, voltage, current, and oil temperature through a Kafka message queue; when accessing historical batch data, upload the original data files to the HDFS storage system, access batch data files of oil chromatography, equipment number, and equipment status through the HDFS storage system, merge the historical files using Flink's batch processing engine, and finally write the stream and batch data into a unified HDFS storage system. The stream and batch data include: timestamps, equipment number, voltage, current, oil temperature, oil chromatography data, and equipment status.

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

5. The method for transformer data fusion and governance based on multimodal processing according to claim 1, characterized in that: In step 2, the identifier 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 governance based on multimodal processing according to claim 1, characterized in that: Step 4 includes the following steps: Step 4.1: Perform time alignment: The input data is the cleaned transformer dataset, including voltage, current, oil temperature, and oil chromatographic gas content. Dynamic Time Warping (DTW) is used to eliminate inconsistencies in the data's time steps. Calculate the optimal alignment path to keep the time series consistent, thus achieving time alignment: ； Where 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 value function, which updates the alignment path by selecting the historical minimum cumulative distance to ensure reasonable matching of time steps; Step 4.2: Construct a temporal attention mechanism to allow the model to focus on important time steps; The temporal attention mechanism calculates the attention weights at time step t: ； Then, the time-weighted features are calculated: ； Among them, h t h is the eigenvector at time step t. i Let i be the feature vector at time step i. For attention weights, W t Here, H is the weight matrix, T is the time dimension length of the input sequence, and H is the weight matrix. temp It is a time-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 location. Graph Convolutional CN is used to model spatial relationships, and an adjacency matrix A is used to represent the connection weights between data points. Spatial features are then calculated. ； Among them, H (l+1) yes l +1 layer features, H l It is the first l Layer features, W l yes l The learnable weight matrix of the layer, It is a non-linear activation function; There are potential correlations between data from different modalities. Therefore, a cross-modal attention mechanism is designed to calculate the correlation between features of different modalities and interactively fuse multimodal information. The correlation between modal features is calculated as follows: ； Computing cross-modal fusion: ； Where S represents the intermodal feature correlation, and H cross For cross-modal fusion feature representation, Q, K, and V are the query, key, and value vectors in attention calculation, respectively, d is the feature dimension, and Softmax() is the normalization function that makes the sum of attention weights equal to 1. Finally, multi-head attention is used for feature fusion: ； Among them, H final For the final fused feature representation, each attention head head1,…head n Perform independent cross-modal interactions to allow different information subspaces to learn fully. Concat() is the concatenation function. O It is a multi-head attention weight matrix, where n is the number of attention heads.

7. The method for transformer data fusion and governance based on multimodal processing according to claim 1, characterized in that: The specific implementation method of step 5 is as follows: 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 time step into training set, validation set and test set. t The input is fed into a bidirectional LSTM network. The forward LSTM extracts past information, and the backward LSTM extracts future information. A fully connected layer maps the output to the label space. After the LSTM output, a CRF layer utilizes contextual relationships to calculate the CRF score, improving the global consistency of label prediction. ； Where P is the entire label sequence y1, y2, ..., y step The conditional probability, It is the class score predicted by LSTM at the current time step. is the state transition score between 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 functions used are Cross-Entropy Loss and CRF loss: ； ； Among them, L CE Let cross-entropy be the loss function. The actual label value. L is the probability predicted by the model. CRF For CRF loss, log() is the logarithmic function; Using the Adam optimizer with a learning rate of 0.001, a bidirectional LSTM-CRF model was trained by minimizing the joint loss of the cross-entropy loss and the CRF loss function. The bidirectional LSTM-CRF model was evaluated using F1 score, accuracy, precision, and recall, calculated using the following formulas: ； Here, F1_score is the F1 score, which measures the performance of the classification model; Precision is the accuracy, which is the proportion of labels predicted as positive that are actually positive; Recall is the recall, which is also the proportion of labels that are actually positive that are predicted as positive.

Images