Power system sensitive data identification method and device, power information terminal and medium

Abstract

The invention provides an electric power system sensitive data identification method and device, an electric power information terminal and a medium, and relates to the technical field of information security. The method comprises the steps of performing preliminary analysis on to-be-output multi-modal data in a power system to obtain multi-modal preliminary analysis data; inputting the multi-modal preliminary analysis data into a routing layer network, performing modal identification and entity identification on the multi-modal preliminary analysis data, and dynamically routing the multi-modal preliminary analysis data to at least one expert model based on an identification result so as to activate the corresponding expert model; performing sensitive information identification on the received multi-mode preliminary analysis data through the activated expert model, and outputting an identification result; and on the basis of fusion of identification results of the expert models for the same kind of entities and the importance degree of each kind of entities in the identification results in the power system, the sensitivity degree of the to-be-output multi-modal data is determined. The problem that an existing sensitive content detection means is low in flexibility and accuracy is solved.

Description

Technical Field

[0001] This invention relates to the field of information security technology, specifically to a method and device for identifying sensitive data in a power system, a power information terminal, and a medium. Background Technology

[0002] With the accelerated construction of new power systems and the rapid increase in the informatization level of the power grid, the risk of leakage of sensitive power-related information has also increased. Taking large-scale power models as an example, although these models effectively improve individual work efficiency and information acquisition capabilities, their black-box nature may lead to the output of sensitive content.

[0003] Traditional methods for identifying sensitive information using keywords lack the ability to detect long texts and contextual information; methods using named entity recognition models can only identify specific sensitive categories. Faced with the complex and diverse entity types in the power industry, traditional methods for detecting sensitive content suffer from low flexibility and accuracy. Summary of the Invention

[0004] To overcome the problems of low flexibility and accuracy of the above-mentioned sensitive content detection methods, the present invention provides a method and device for identifying sensitive data in power systems, a power information terminal, and a medium.

[0005] On one hand, the present invention provides a method for identifying sensitive data in a power system, comprising: Preliminary analysis of the multimodal data to be output from the power system is performed to obtain preliminary multimodal analysis data; The multimodal preliminary parsing data is input into the routing layer network, where modality recognition and entity recognition are performed on the multimodal preliminary parsing data, and the multimodal preliminary parsing data is dynamically routed to at least one expert model based on the recognition results to activate the corresponding expert model. The activated expert model identifies sensitive information from the received multimodal preliminary analysis data and outputs the identification results. Based on the fusion of the identification results of various expert models for similar entities and the importance of various entities in the power system in the identification results, the sensitivity of the multimodal data to be output is determined in order to complete the identification of sensitive information in the multimodal data to be output. The expert models include a text expert model, an image expert model, and a table expert model; the text expert model is used to process power text data; the image expert model is used to process power image data; and the table expert model is used to process power table data.

[0006] Optionally, the text expert model includes a personal privacy information expert model, a power grid facility information expert model, and a business operation data expert model; The expert model for personal privacy information is used to identify sensitive text information related to personal privacy. The power grid facility information expert model is used to identify sensitive text information related to key equipment within the power system. The enterprise operation data expert model is used to identify sensitive text information related to the operation data of various power companies.

[0007] Optionally, the routing layer network includes a modality recognition module and a gating network; the multimodal preliminary parsing data includes at least one of power text parsing data, power image parsing data, and power table parsing data; the multimodal preliminary parsing data is input into the routing layer network, modality recognition and entity recognition are performed on the multimodal preliminary parsing data, and the multimodal preliminary parsing data is dynamically routed to at least one expert model based on the recognition results, including: The preliminary multimodal analysis data is input into the modality recognition module for modality recognition. The gated network is used to perform entity recognition on the parsed power text data whose modality recognition result is text, so as to dynamically route the power text data to the corresponding text expert model; The power image parsing data, which is the modality recognition result, is dynamically routed to the image expert model; The power table parsing data, which is a table of modality recognition results, is dynamically routed to the table expert model.

[0008] Optionally, the gated network is used to perform entity recognition on the power text parsing data whose modality recognition result is text, so as to dynamically route the power text parsing data to the corresponding text expert model, including: The power text parsing data is segmented and embedded to obtain the corresponding text matrix; The text matrix is ​​subjected to parallel convolution, linear activation and pooling operations using multiple one-dimensional convolution kernels of different sizes to obtain the text feature vectors corresponding to each one-dimensional convolution kernel. The text feature vectors corresponding to each one-dimensional convolution kernel are concatenated and then randomly discarded to obtain a comprehensive text feature vector. The comprehensive text feature vector is processed by full connection and nonlinear mapping to obtain the probability that the power text data belongs to each text expert model; the power text data is then dynamically routed to the text expert model with the highest probability.

[0009] Optionally, if the activated expert model includes an image expert model, the image expert model includes a compressed excitation attention layer; the multimodal preliminary parsing data includes power image parsing data; the step of identifying sensitive information from the received multimodal preliminary parsing data through the activated expert model and outputting the identification result includes: The received power image parsing data is subjected to parallel feature extraction with multi-scale convolution kernels, linear activation, and global average pooling to obtain multi-scale basic feature vectors. The multi-scale basic feature vector is weighted by a compressed excitation attention layer, and channel attention processing is performed on the multi-scale basic feature vector based on the learned weights to obtain a channel attention-adjusted feature vector. The channel attention adjustment feature vector is binarized and segmented to determine the target text region; the target text region is then used to perform text recognition using a convolutional recurrent neural network based on an attention mechanism to obtain structured text. The structured text is semantically encoded to obtain the corresponding semantic encoding vector; Sensitive entity matching is performed based on the similarity between the semantic encoding vector and each entity in the pre-constructed power knowledge graph, and the sensitive text recognition result of the power image parsing data is output. In the channel attention processing, a preset channel weight threshold is used to filter background interference features.

[0010] Optionally, if the activated expert model includes a table expert model; and the multimodal preliminary analysis data includes electricity table analysis data; the step of identifying sensitive information in the received multimodal preliminary analysis data through the activated expert model and outputting the identification result includes: Extract the header text and corresponding cell text from the parsed data of the electricity table; Semantic encoding is performed on the header text and the corresponding cell text respectively to generate a header semantic vector and a cell text feature vector; The semantic vector of the table header is concatenated with the corresponding text feature vector of the cell to form a fused feature vector; the fused feature vector is then subjected to layer normalization. The bidirectional long short-term memory network in the table expert model is used to sequentially extract bidirectional temporal dependencies, perform cross-cell semantic association and linear classification on the normalized fused feature vector to obtain the table entity classification result. Based on the transition scores of each entity category label learned by the conditional random field layer in the table expert model and the table entity classification results, the Viterbi decoding algorithm is used to decode the globally optimal label sequence, thereby obtaining the recognition result of the power table parsing data.

[0011] Optionally, the bidirectional long short-term memory network in the table expert model includes a bidirectional LSTM layer and an in-row attention layer connected in sequence; the cross-cell semantic association includes: The time-series correlation feature vectors output by the bidirectional LSTM layer are concatenated column by column to form a row feature matrix, so that the time-series correlation feature vectors corresponding to each cell in the same row of the power table parsing data are located in the same row of the row feature matrix; Based on the row feature matrix, the intra-row attention weights are determined using a single-head attention mechanism, and the row feature matrix is ​​weighted using the intra-row attention weights to obtain an intra-row association enhancement feature matrix. The feature matrix with enhanced in-row correlation is split by column to restore the matrix dimension.

[0012] Optionally, if the activated expert model includes the personal privacy information expert model and / or the power grid facility information expert model, the step of identifying sensitive information from the received multimodal preliminary parsing data through the activated expert model and outputting the identification result includes: The target rule module of the activated expert model is used to perform rule matching on the text to be recognized, and the rule recognition result is obtained. The text to be recognized is input into the encoding layer of the activated expert model and then deep encoded to obtain a deep encoded vector. The bidirectional long short-term memory network in the activated expert model is used to extract bidirectional temporal dependencies and perform linear classification on the normalized deep encoding vector to obtain the text entity classification result. Based on the transition scores of each entity category label learned by the conditional random field layer in the activated expert model and the text entity classification results, the Viterbi decoding algorithm is used to decode the globally optimal label sequence, thereby obtaining the recognition result output by the activated expert model. Based on the rule recognition results and the model output recognition results, the final recognition result of the text to be recognized is determined; Wherein, if the activated expert model is a personal privacy information expert model, the target rule module is a regular expression for personal privacy information; if the activated expert model is a power grid facility information expert model, the target rule module is constructed based on various power field dictionaries.

[0013] Optionally, if the activated expert model includes the enterprise management data expert model, the step of identifying sensitive information from the received multimodal preliminary analysis data through the activated expert model and outputting the identification result includes: The rule matching module of the enterprise operation data expert model is used to perform rule matching on the received text to be identified to obtain the matching entity; The text to be identified is input into the text classification model of the enterprise management data expert model and subjected to multi-scale one-dimensional convolution processing, fully connected and nonlinear mapping to obtain the recognition result output by the text classification model. The final recognition result is determined based on the matching entity and the recognition result output by the text classification model.

[0014] Optionally, based on the fusion of the identification results of various expert models for the same type of entity and the importance of various entities in the power system in the identification results, the sensitivity of various entities is determined, including: For each type of entity, a weighted fusion is performed based on the sensitivity probability, number of entities, and number of modalities of the entity in the recognition results output by each expert model to obtain the sensitivity score for each type of entity. Based on the importance of various entities in the power system and the sensitivity score of each entity, the sensitivity of the multimodal data to be output is obtained.

[0015] Optionally, after determining the sensitivity of the multimodal data to be output, the method further includes: The multimodal data to be output is controlled to be output in a hierarchical manner based on its sensitivity. The graded output includes at least one of the following: blocking output, desensitized output, post-recording output, and direct output.

[0016] On the other hand, the present invention also provides a power system sensitive data identification device, comprising: The parsing unit is used to perform preliminary parsing of the multimodal data to be output in the power system, and to obtain preliminary multimodal parsing data; The routing unit is used to input the multimodal preliminary parsing data into the routing layer network, perform modality recognition and entity type recognition on the multimodal preliminary parsing data, and dynamically route the multimodal preliminary parsing data to at least one expert model based on the recognition results, so as to activate the corresponding expert model. The identification unit is used to identify sensitive information in the multimodal preliminary analysis data it receives through the activated expert model and output the identification results. The fusion unit is used to determine the sensitivity of the multimodal data to be output based on the fusion of the identification results of various expert models for the same type of entity and the importance of various entities in the power system in the identification results, so as to complete the identification of sensitive information in the multimodal data to be output. The expert models include a text expert model, an image expert model, and a table expert model; the text expert model is used to process power text data; the image expert model is used to process power image data; and the table expert model is used to process power table data.

[0017] On the other hand, the present invention also provides a power information terminal, including: the power system sensitive data identification device of any of the above.

[0018] On the other hand, the present invention also provides an electronic device, comprising: at least one processor and a memory; the memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, the method described in any of the foregoing is implemented.

[0019] On the other hand, the present invention also provides a readable storage medium having an executable program stored thereon, wherein when the executable program is executed, it implements the method described in any one of the above.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a method and apparatus for identifying sensitive data in power systems. It utilizes a routing layer network for modality and entity type identification, and dynamically routes data to at least one expert model based on the identification results. The dynamic routing mechanism activates the corresponding expert model according to data characteristics, ensuring flexibility and relevance in the processing, thereby improving identification efficiency and reducing false positives. Sensitive information is identified through activated expert models (including text, image, and table expert models). Each expert model focuses on a specific modality and data type, enabling it to capture the complex and diverse entity types in the power system and improve identification accuracy. Finally, the identification results from each expert model are fused, and the sensitivity level is determined by combining the entity's importance within the power system. This avoids the bias of a single model's output, making the identification results more aligned with the actual needs of the power industry and improving the comprehensiveness and reliability of sensitive information identification. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an example of a power system sensitive data identification method according to the present invention. Figure 2 This is a schematic diagram of a multimodal data-sensitive content recognition architecture as an example of the present invention; Figure 3 This is a schematic diagram of an example of an MOE-based power text parsing data recognition architecture according to the present invention. Figure 4 This is a schematic diagram of the core data processing flow of a gating network, as an example of the present invention. Figure 5 This is a schematic diagram of the data processing flow of an image expert model as an example of the present invention; Figure 6 A schematic diagram of the data processing flow of a table expert model as an example of the present invention; Figure 7 This is a schematic diagram of the data processing flow of the bidirectional LSTM layer and the inline attention layer in a table expert model of an example of the present invention. Figure 8 This is a schematic diagram illustrating the processing flow of an expert model for personal privacy information, as an example of the present invention. Figure 9 This is a schematic diagram illustrating the text processing flow of a BERT-BiLSTM-CRF model architecture, as an example of the present invention. Figure 10 This is a schematic diagram illustrating the processing flow of an expert model for power grid facility information, as an example of the present invention. Figure 11 This is a schematic diagram illustrating the processing flow of an expert model for enterprise management data, as an example of the present invention. Figure 12 This is a schematic diagram illustrating the deployment location of a sensitive data identification device according to an example of the present invention; Figure 13 This is a structural block diagram of an electronic device according to the present invention. Detailed Implementation

[0022] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0023] Explanation of the proper nouns used in this application: NER (Named Entity Recognition) is a subtask of information extraction that aims to locate and classify named entities in text into predefined categories, such as people, organizations, locations, time expressions, quantities, currency values, percentages, etc. A named entity is a word or phrase that clearly distinguishes an object from objects with similar attributes.

[0024] Regex: A regular expression is a tool for describing string matching patterns. Users define matching rules to find strings that match those rules within a long text. It is commonly used for text matching, search, and replace operations, and can identify complex text patterns with concise expressions.

[0025] TextCNN (Text Convolutional Neural Network) is a deep learning model based on the Convolutional Neural Network (CNN) architecture, specifically designed for text classification tasks. Its core idea is to borrow the ability of CNNs to extract local features in image processing, automatically extracting key semantic features from text through convolutional operations to complete text classification. In text, the order and adjacency relationships between words contain semantic information, and TextCNN uses convolutional operations to capture these local semantic units in the text.

[0026] BERT: Bidirectional Encoder Representations from Transformers, is a bidirectional pre-trained language model based on the Transformer architecture. It achieves contextual understanding through masked language model (MLM) and next sentence prediction (NSP) tasks, significantly improving the performance of natural language processing (NLP) tasks.

[0027] BiLSTM: Bi-directional Long Short-Term Memory, is a combination of forward LSTM and backward LSTM. The LSTM model is based on RNN with the addition of input gates, forget gates, unit states, and output gates. During network training, information can be added or removed through the gate structures. Different neural networks can use the gate structures on the unit states to determine which relevant information to remember or forget.

[0028] CRF (Conditional Random Field) is a discriminative sequence modeling method in natural language processing. It models output variables using conditional probability distributions to solve sequence labeling tasks (such as named entity recognition and part-of-speech tagging). Its core advantage lies in integrating context marker transition probabilities, avoiding the independence assumption of traditional HMM models, and supporting flexible feature addition and deletion mechanisms.

[0029] Example 1 The present invention provides a method for identifying sensitive data in a power system, the schematic diagram of which is shown below. Figure 1 As shown, the method includes: Step S110: Perform preliminary analysis on the multimodal data to be output in the power system to obtain preliminary multimodal analysis data; Step S120: Input the multimodal preliminary parsing data into the routing layer network, perform modality recognition and entity type recognition on the multimodal preliminary parsing data, and dynamically route the multimodal preliminary parsing data to at least one expert model based on the recognition results to activate the corresponding expert model; Step S130: The activated expert model identifies sensitive information in the received multimodal preliminary analysis data and outputs the identification results. Step S140: Based on the fusion of the identification results of various expert models for the same type of entity and the importance of various entities in the power system in the identification results, determine the sensitivity of the multimodal data to be output in order to complete the identification of sensitive information of the multimodal data to be output. The expert models include a text expert model, an image expert model, and a table expert model; the text expert model is used to process power text data; the image expert model is used to process power image data; and the table expert model is used to process power table data.

[0030] In this example implementation, the multimodal data to be output may include multiple modalities, such as text, tables, and images. The multimodal data to be output may include at least one of power-related text data, power-related table data, and power-related image data. For example, when a user accesses a large-scale power model, the large-scale power model obtains relevant content based on the access information and returns response information to the user. The multimodal data to be output may be the access-related content obtained by the large-scale power model. Before returning this content to the user, the sensitive data identification process of this application is performed. This avoids the problem of sensitive data leakage through the large-scale power model. Figure 2As shown, the execution process includes: First, preprocessing the multimodal data to be output from the power system, such as preliminary analysis and cleaning, transforms it into standardized features that subsequent models can understand. For power table data, table structure parsing can be performed, that is, for spreadsheets (such as Excel), the row index, column index, and header cell content of the file are directly read to construct a "header-cell" mapping table, obtaining preliminary multimodal parsing data. Converting the raw data into standardized features lays the foundation for subsequent analysis and ensures comprehensive processing of complex data sources in the power industry. Second, the preliminary multimodal parsing data is input into the routing layer network for modality recognition and entity recognition, and the corresponding expert models are activated. The routing layer network (including the gating network) acts as the intelligent scheduling hub, analyzing data features through lightweight neural networks (such as TextCNN), automatically identifying entity categories (such as personal privacy, power grid facilities, or enterprise operation data) and data types (such as judging image and table data by file format). The dynamic routing mechanism activates the most relevant experts based on data characteristics, avoiding the rigidity problem of traditional fixed models that can only handle specific entity types. Third, the activated expert models (including text, image, and table expert models) identify sensitive information in the received data. The expert model pool is a set of pre-trained, highly specialized deep learning models. Each expert focuses on identifying only one type of sensitive information. Each expert model is deeply optimized for a specific modality: text expert models (such as BERT-BiLSTM-CRF) process power-related text data, using natural language processing techniques to capture long texts and contextual semantics, solving the problem of lack of semantic understanding in traditional sensitive word detection; image expert models are used to identify text entity information contained in image scenes. For example, the image expert model extracts implicit sensitive information from image data such as photos in power-related scenes. The specific implementation process includes: image region segmentation and sensitive region localization: using the YOLOv8 object detection model to automatically locate "text regions" and output the bounding box coordinates of each region; image text extraction (OCR) and semantic encoding: for the located text regions, the PP-OCRv4 model is used to extract the text content and remove blurred and noisy characters. Table expert models (such as BERT-BiLSTM-CRF) process electricity table data to identify various entity information contained in the table and resolve header ambiguity and cross-cell association issues.For example, the table expert model addresses the issues of "header ambiguity" and "splitting of sensitive information across cells" for structured data in Excel / CSV format in the power industry (such as user electricity consumption details and equipment procurement lists). The specific implementation process is as follows: Header semantic understanding and alignment: The BERT model is used to semantically encode the header text (such as "account number," "equipment model," and "contract amount"), generating a 768-dimensional header semantic vector; Cross-cell association feature extraction: For sensitive information split across columns within a row (such as "Li" in column A, "four" in column B, and "139xxxx" in column C, which are actually complete personal information), the BiLSTM model is used to model the semantic association of cells within the row. The input is a "cell content sequence," and the output is the "cross-cell semantic association degree" (e.g., if the association degree of columns A+B+C is ≥0.8, it is determined to be a combined sensitive entity). Each model is optimized based on domain knowledge, enabling it to accurately capture the complex and diverse entity types in the power industry (such as equipment models and financial data). Finally, based on the fusion of the identification results of various expert models for similar entities, and combined with the importance of various entities in the power system in the identification results, multimodal result fusion (such as weighted comprehensive risk score calculation) avoids the one-sidedness of single model output and improves reliability through cross-validation; at the same time, considering the importance of entities (such as core substation information being more sensitive than ordinary data) makes the sensitivity assessment more in line with the actual needs of the power industry. This application is approved as follows. Figure 2 The multimodal hybrid expert processing architecture shown automatically routes different types of data to the most specialized expert models (personal privacy information expert model, power grid facility information expert model, enterprise operation data expert model, image expert model, and table expert model) for processing through a routing layer network (intelligent scheduling hub). It then integrates the recognition results from different expert models to jointly determine the comprehensive sensitivity level of the data to be output. Ultimately, this achieves accurate, interpretable, and adaptive sensitive data discovery and security control. A pipelined processing architecture of "perception-routing-decision" is adopted to automatically identify and control sensitive information in the input multimodal power data.

[0031] In some implementations, the text expert model includes a personal privacy information expert model, a power grid facility information expert model, and a business operation data expert model; The expert model for personal privacy information is used to identify sensitive text information related to personal privacy. The power grid facility information expert model is used to identify sensitive text information related to key equipment within the power system. The enterprise operation data expert model is used to identify sensitive text information related to the operation data of various power companies.

[0032] In this example implementation, the text expert model is specifically divided into a personal privacy information expert model, a power grid facility information expert model, and a business operation data expert model, each optimized for a specific type of sensitive text information. For example... Figure 3 As shown, the personal privacy information expert model focuses on identifying entities such as names, ID numbers, mobile phone numbers, and addresses, and is implemented using a hybrid approach of rule matching and NER models: the rule module uses regular expressions to match fixed pattern information (such as the 18 digits of an ID number). For example, the deep learning model uses the BERT-BiLSTM-CRF architecture. The power grid facility information expert model processes key equipment information in the power system, such as substation latitude and longitude, equipment information, and geographic information system data; it is implemented using a hybrid approach of power industry dictionary matching and NER models. The NER model can use the BERT-BiLSTM-CRF architecture, but the parameters are adjusted to accommodate longer texts (maximum sequence length 256), and an integrated power industry dictionary (such as a substation name dictionary and an equipment model dictionary) is used to enhance recognition capabilities. The enterprise operation data expert model identifies commercially sensitive information such as undisclosed electricity consumption, electricity prices, load curves, financial data, and development plans; it combines rule matching (regular expressions for amounts and account numbers) and the TextCNN text classification model. The outputs of all three expert models are entity category labels and corresponding confidence scores. This specialization allows each model to be deeply optimized in a specific domain. The personal privacy model efficiently processes patterned data, the power grid facility model adapts to industry terminology, and the enterprise management model quickly classifies text types, thereby improving the overall accuracy and adaptability of the system. At the same time, the modular design supports independent updates, such as when adding a new expert, the system does not need to be rebuilt.

[0033] In some implementations, the routing layer network includes a modality recognition module and a gating network; the multimodal preliminary parsing data includes at least one of power text parsing data, power image parsing data, and power table parsing data; the multimodal preliminary parsing data is input into the routing layer network, modality recognition and entity recognition are performed on the multimodal preliminary parsing data, and the multimodal preliminary parsing data is dynamically routed to at least one expert model based on the recognition results, including: The preliminary multimodal analysis data is input into the modality recognition module for modality recognition. The gated network is used to perform entity recognition on the power text parsing data whose modality recognition result is text, so as to dynamically route the power text parsing data to the corresponding text expert model; The power image parsing data, which is the modality recognition result, is dynamically routed to the image expert model; The power table parsing data, which is a table of modality recognition results, is dynamically routed to the table expert model.

[0034] In this example implementation, the multimodal preliminary parsing data includes at least one of power text parsing data, power image parsing data, and power table parsing data. The routing layer network includes a modality recognition module and a gating network, responsible for subsequent modality recognition and routing decisions on the aforementioned multimodal preliminary parsing data. The modality recognition module performs rapid classification based on data format and preprocessing features: it identifies the modality of image data and table data according to file format, and routes power image parsing data identified as image modality to an image expert model, and power table parsing data identified as table modality to a table expert model. For power text parsing data whose modality is identified as text, it is sent to a dedicated gating network for deep entity recognition. The gating network can be a lightweight TextCNN model, which obtains the probability that the text belongs to a personal privacy information expert model, a power grid facility information expert model, and a business operation data expert model. The routing decision dynamically routes the power text parsing data to the text expert model with the highest probability based on this probability distribution. The entire routing process constructs an efficient hierarchical decision-making system: the modality recognition module acts as the first-level filter, quickly diverting data of different formats; the gating network acts as the second-level intelligent router, specifically for analyzing text data and assigning experts.

[0035] In some implementations, the gating network is used to perform entity recognition on the parsed power text data whose modality recognition results are text, so as to dynamically route the parsed power text data to the corresponding text expert model, including: The power text parsing data is segmented and embedded to obtain the corresponding text matrix; The text matrix is ​​subjected to parallel convolution, linear activation and pooling operations using multiple one-dimensional convolution kernels of different sizes to obtain the text feature vectors corresponding to each one-dimensional convolution kernel. The text feature vectors corresponding to each one-dimensional convolution kernel are concatenated and then randomly discarded to obtain a comprehensive text feature vector. The comprehensive text feature vector is processed by full connection and nonlinear mapping to obtain the probability that the power text parsing data belongs to each text expert model; the power text parsing data is then dynamically routed to the text expert model with the highest probability.

[0036] In this example implementation, the gating network is a lightweight neural network that receives a general feature vector, analyzes the overall features of the input data, and outputs a probability distribution indicating which expert models should process the data to achieve fast and accurate routing. The gating network dynamically activates the most relevant text expert model based on the input data; unactivated experts do not participate in the calculation. The final output is a weighted combination of the results from the activated experts. The specific process of entity recognition using the gating network on power text parsing data includes: the input to the gating network is power text parsing data (a piece of text); the model selected is the lightweight TextCNN; the output is a 3-dimensional probability vector, representing the probabilities of selecting three text experts. The expert with the highest probability is activated. For example, as shown... Figure 4 The diagram illustrates the core data processing flow of the gating network. The input to the gating network is parsed power text data. Preprocessing and embedding are performed on the input: the text is first segmented and converted into a sequence of word indices. An embedding layer maps each index to a dense 100-dimensional vector, transforming the text into a two-dimensional matrix (sequence length × 100). Next, multi-scale convolution is performed: four different (2, 3, 4, 5) one-dimensional convolution kernels are used to scan the embedding matrix in parallel. Each kernel size has 128 kernels to extract local n-gram features of different lengths. Each convolution operation is followed by a ReLU activation function. Global max pooling is performed on the feature maps generated by each convolution kernel, resulting in a 128-dimensional feature vector. Finally, feature concatenation is performed: the four pooled feature vectors (each 128-dimensional) are concatenated into a 512-dimensional composite feature vector to fuse text features of different granularities. Next, the classification output is performed: the concatenated feature vector is first passed through a Dropout layer to prevent overfitting, then through a fully connected layer to further combine the features, and finally through a Softmax output layer to obtain the final expert probability distribution, which represents the probability that the text belongs to one of the three expert models: personal privacy information, power grid facility information, or enterprise operation information. Routing decision: Based on the output probability distribution, the system determines which text expert model to route the text to for subsequent fine-grained recognition. The actual model parameters used in this example, as well as the model training parameters and parameters used during the model training process, are shown in Tables 1-3. Table 1 shows the model parameters.

[0037] Table 2 Training Parameters

[0038] Table 3 Training process parameters

[0039] The above-described gating network routing process ensures the extraction of features at different granularities through multi-scale convolutions, ranging from phrases to long-distance dependencies; Dropout and fully connected layers enhance the model's generalization ability; and the Softmax output provides an interpretable probability distribution. The lightweight design of the gating network makes it efficient and low-latency, achieving high routing accuracy on power text parsing data.

[0040] In some implementations, if the activated expert model includes an image expert model, the image expert model includes a compressed excitation attention layer; the multimodal preliminary parsing data includes power image parsing data; the step of identifying sensitive information from the received multimodal preliminary parsing data through the activated expert model and outputting the identification result includes: The received power image parsing data is subjected to parallel feature extraction with multi-scale convolution kernels, linear activation, and global average pooling to obtain multi-scale basic feature vectors. The multi-scale basic feature vector is weighted by a compressed excitation attention layer, and channel attention processing is performed on the multi-scale basic feature vector based on the learned weights to obtain a channel attention-adjusted feature vector. The channel attention adjustment feature vector is binarized and segmented to determine the target text region; the target text region is then used to perform text recognition using a convolutional recurrent neural network based on an attention mechanism to obtain structured text. The structured text is semantically encoded to obtain the corresponding semantic encoding vector; Sensitive entity matching is performed based on the similarity between the semantic encoding vector and each entity in the pre-constructed power knowledge graph, and the sensitive text recognition result of the power image parsing data is output. In the channel attention processing, a preset channel weight threshold is used to filter background interference features.

[0041] In this example implementation, when the image expert model is activated, it specifically processes sensitive information in the power image parsing data, achieving detection through "sensitive area localization - OCR text recognition - semantic matching". This may involve an input preprocessing layer for extracting basic features from the input image and performing feature enhancement to capture text details. For example, such as... Figure 5As shown, the EfficientNet-B1 architecture is used to extract basic features from power image analysis data. Specifically, three convolutional kernels of different sizes (3×3, 5×5, and 7×7) are used in parallel, with 128 kernels of each size, capturing features at different scales, especially small text details. ReLU activation is then used to increase non-linearity, and global average pooling is used to preserve local parametric features, ultimately outputting a 512-dimensional multi-scale feature vector. The output of the EfficientNet-B1 architecture enters the Compressed Excitation Attention Layer (SE Attention Layer). The SE Attention Layer uses a compression-excitation mechanism to enhance useful text region features and suppress background interference such as reflections. For example, the channel dimension is compressed by 16 times to improve computational efficiency, while a channel weight threshold of 0.3 is set; background channel features below this threshold are filtered out. Finally, a 512-dimensional feature vector with adjusted weights (channel attention-adjusted feature vector) is output. The output of the SE Attention Layer enters the PP-OCRv4 text extraction layer. This layer performs text detection and recognition in an end-to-end manner. First, DBNet uses binarization to segment and locate text regions in the image, then CRNN combined with an attention mechanism is used to recognize the text, outputting structured text. Next, the structured text enters the BERT semantic encoding layer for semantic understanding. For example, the parameters of the BERT semantic encoding layer are: hidden layer dimension 768, maximum processing length of 128 text sequences, Dropout 0.2 to prevent overfitting, and finally outputting a 768-dimensional semantic vector. This semantic vector enters the semantic verification layer, which matches the encoded semantic vector with entities in the knowledge graph, calculates the matching degree using the cosine similarity algorithm, and verifies whether the text entities in the power image parsing data are power-sensitive entities. The final output is [entity text, entity type, coordinates in the image, confidence score]. The entity types identified in this part can include various types, such as personal privacy, power grid equipment, business data entities, etc., and can also be other types. This example improves text detection accuracy through multi-scale convolution and attention mechanisms. The SE attention layer effectively enhances useful features and improves the accuracy of sensitive information recognition in complex images.

[0042] In some implementations, if the activated expert model includes a table expert model; and the multimodal preliminary parsing data includes electricity table parsing data; the step of identifying sensitive information in the received multimodal preliminary parsing data through the activated expert model and outputting the identification result includes: Extract the header text and corresponding cell text from the parsed data of the electricity table; Semantic encoding is performed on the header text and the corresponding cell text respectively to generate a header semantic vector and a cell text feature vector; The semantic vector of the table header is concatenated with the corresponding text feature vector of the cell to form a fused feature vector; the fused feature vector is then subjected to layer normalization. The bidirectional long short-term memory network in the table expert model is used to sequentially extract bidirectional temporal dependencies, perform cross-cell semantic association and linear classification on the normalized fused feature vector to obtain the table entity classification result. Based on the transition scores of each entity category label learned by the conditional random field layer in the table expert model and the table entity classification results, the Viterbi decoding algorithm is used to decode the globally optimal label sequence, thereby obtaining the recognition result of the power table parsing data.

[0043] In this example implementation, the table expert model identifies sensitive information in structured data in Excel / CSV format used in the power industry. The data parsed from the power table can be structured data consisting of "header text + cell text" as model input. For example, as shown... Figure 6As shown, the basic model of the table expert model adopts the "BERT-BiLSTM-CRF" architecture. The model input is "table header text + cell text" (e.g., table header "Name" → cell "Zhang San"). First, table header semantic embedding is performed to encode the table header semantics into a vector, which serves as "contextual guidance" for the cell text. A lightweight BERT is used, with an input sequence length of 32 (table header text is usually ≤10 characters), an output dimension of 256 (dimensionality reduction through fully connected layers to adapt to subsequent concatenation), Dropout: 0.2, and an output of a 256-dimensional table header semantic vector. Next, cell text encoding is performed to encode the cell text into a semantic vector. BERT-base-Chinese is used for fine-tuning, with an input sequence length of 64, a hidden layer dimension of 768, Dropout: 0.2, and an output of a 768-dimensional cell text token-level feature vector. Next, feature concatenation is performed, which involves concatenating the semantic vector of the table header with the text feature vector of each cell, injecting contextual information from the table header, resulting in a 1024-dimensional concatenation. This concatenated vector is then normalized using LayerNorm, outputting a 1024-dimensional token-level fusion feature vector. This vector is then passed through a BiLSTM layer (512×2 hidden layers, bidirectional Dropout 0.2, tanh activation function) for bidirectional temporal dependency extraction and feature concatenation at each time step, yielding a temporal association feature vector. This temporal association feature vector is then fed into an intra-row attention layer for cross-cell association. The intra-row attention layer calculates the attention weights for different cells within the same row (e.g., the weight of the "Account Number" cell for the "Name" cell is 0.7), strengthening the semantic association across cells. The output dimension is a 1024-dimensional token-level feature vector with enhanced intra-row association, consistent with the input dimension, only updating the feature weights. The linear classification output CRF layer functions similarly to the above, learning the constraints between labels. The labels are the 23 labels mentioned above (including two newly added monetary labels). Decoding algorithm: Viterbi decoding (searching for the label sequence with the highest global probability). Output: Sensitive label sequence. The final model output is: [Sensitive entity, type, cell position, confidence].

[0044] For example, the bidirectional long short-term memory network in the table expert model includes a bidirectional LSTM layer and an in-row attention layer connected in sequence; the cross-cell semantic association includes: The time-series correlation feature vectors output by the bidirectional LSTM layer are concatenated column by column to form a row feature matrix, so that the time-series correlation feature vectors corresponding to each cell in the same row of the power table parsing data are located in the same row of the row feature matrix; Based on the row feature matrix, the intra-row attention weights are determined using a single-head attention mechanism, and the row feature matrix is ​​weighted using the intra-row attention weights to obtain an intra-row association enhancement feature matrix. Split the feature matrix with enhanced in-line association by columns to restore the matrix dimension.

[0045] In the present exemplary embodiment, the bidirectional long short-term memory network in the table expert model includes a bidirectional LSTM layer, an in-line attention layer, and a linear classification layer connected in sequence. The bidirectional LSTM layer is used to capture temporal features from left to right and from right to left to enhance the temporal logic. As Figure 7 shown, the input of the bidirectional LSTM layer is the 1024-dimensional fused feature after feature splicing, and the output of the bidirectional LSTM layer is the 1024-dimensional temporally enhanced feature. By obtaining the temporal dependence between Tokens through the bidirectional LSTM layer, sequence associations are generated for isolated Token features, preventing the model from treating "Zhang" and "San" as two independent characters. The in-line attention layer is used to strengthen cross-cell associations. Its input is the 1024-dimensional temporally enhanced feature output by the bidirectional LSTM layer, and its output is the 1024-dimensional in-line association enhanced feature (retaining header context + temporal logic + newly added cross-cell associations), thus making the same-row cells mutually associated and solving the problem of cross-cell split entity recognition (e.g., "Li" in column A + "Si" in column B → "Li Si"). The output of the in-line attention layer is linearly classified by the linear classification layer, and 23-dimensional entity categories are output. Exemplarily, the data processing process of the in-line attention layer is as follows: First, data reconstruction is performed, that is, the temporal feature vectors of all cells in the same row are concatenated in column order to obtain a row feature matrix (whose dimension is the number of cells × text length × 1024). The single-head attention mechanism is: Let the query vector Q, the key vector K, and the value vector V all be equal to the row feature matrix, then the in-line attention weight is calculated as follows: ; In the formula, n is the row vector dimension of the row feature matrix, such as 1024.

[0046] The in-line association enhanced feature matrix H is calculated as follows: ; By weighted summation of the attention weight and the value vector V, the in-line association enhanced feature matrix H is obtained. Output reconstruction: The in-line association enhanced feature matrix H is split into feature vectors of individual cells, keeping the output shape of the in-line attention layer the same as the input shape. The output of the in-line attention layer is a 1024-dimensional in-line association enhanced Token-level feature vector. The training parameters and model parameters of the table expert model are shown in Table 4 below.

[0047] Table 4

[0048] This example employs a single-head self-attention architecture. Single-head attention can focus on global relationships within a row, avoiding excessive complexity. An in-row self-attention mask is used to calculate weights only between cells within the same row, masking cross-row interference. The table expert model processing described above resolves ambiguity through header-cell semantic fusion, captures cross-column relationships through in-row attention, and ensures sequence labeling accuracy using BiLSTM and CRF.

[0049] In some implementations, if the activated expert model includes the personal privacy information expert model and / or the power grid facility information expert model, the step of identifying sensitive information from the received multimodal preliminary parsing data through the activated expert model and outputting the identification result includes: The target rule module of the activated expert model is used to perform rule matching on the text to be recognized, and the rule recognition result is obtained. The text to be recognized is input into the encoding layer of the activated expert model and then deep encoded to obtain a deep encoded vector. The bidirectional long short-term memory network in the activated expert model is used to extract bidirectional temporal dependencies and perform linear classification on the normalized deep encoding vector to obtain the text entity classification result. Based on the transition scores of each entity category label learned by the conditional random field layer in the activated expert model and the text entity classification results, the Viterbi decoding algorithm is used to decode the globally optimal label sequence, thereby obtaining the recognition result output by the activated expert model. Based on the rule recognition results and the model output recognition results, the final recognition result of the text to be recognized is determined.

[0050] In this example implementation, both the personal privacy information expert model and the power grid facility information expert model employ a combination of rule modules and deep learning models (such as BERT-BiLSTM-CRF). If the activated expert model is the personal privacy information expert model, the target rule module is a regular expression for personal privacy information; if the activated expert model is the power grid facility information expert model, the target rule module is constructed based on various power domain dictionaries. The processing flow for the personal privacy information expert model is as follows: Figure 8As shown, personal privacy information from the input power text parsing data is input in parallel into the target rule module and the deep learning model. The target rule module uses regular expressions to match information with fixed patterns, such as ID card numbers, mobile phone numbers, and bank card numbers; this can be achieved using preset regular expressions. Since personal privacy data is mostly this type of regularized data, this module is used extensively in this step. After the deep learning model receives the input sequence, the BERT layer first performs layer normalization before inputting it into a 12-layer transformer encoder. The BERT layer outputs a context vector for each character, 768 dimensions per character, forming a new vector sequence, which is then input into the BiLSTM layer. In this layer, bidirectional encoding is used, and the bidirectional encoded vectors are concatenated to output bidirectional features for each character, 768 dimensions per character. After passing through a fully connected layer, the label score for each character is calculated, outputting 9 dimensions per character, corresponding to 9 labels. This is then input into the CRF layer to learn label transition rules, forming a 9×9 transition matrix. Finally, Viterbi decoding is used to find the globally optimal label path, forming the final prediction sequence. The final predicted sequence and the rule recognition results of the rule module are fused (e.g., vector concatenation) to obtain the final recognition result of the model.

[0051] For example, such as Figure 9 The following shows the complete process of the BERT-BiLSTM-CRF model processing text and identifying personal privacy information: Input and encoding: The original text is converted into a token ID sequence and an attention mask by the BERT tokenizer, and special symbols such as [CLS] and [SEP] are added.

[0052] BERT encoding layer: The ID of each token is converted into three embedding vectors: word embedding, position embedding, and paragraph embedding. The three embedding representations are added together to form the initial representation. The initial representation is then deep encoded by 12 Transformer modules, and its output is a deep feature vector (768 dimensions) for each token that incorporates full-text context information.

[0053] BiLSTM layer: The feature vector output by BERT is fed into a bidirectional LSTM (BiLSTM) to further capture sequence features. The forward LSTM reads the sequence from left to right, and the backward LSTM reads the sequence from right to left. The outputs of the two are concatenated at each time step. A Dropout layer is then added to prevent overfitting.

[0054] Linear classification layer: The features of each token output by BiLSTM are mapped to a score vector with a length equal to the number of labels, indicating that the token is predicted as the original score for each label.

[0055] CRF layer: The CRF layer comes with a state transition matrix (transition score), which is the constraint relationship between pre-learned labels, and can avoid illegal label sequences that may be generated by the linear layer.

[0056] Viterbi Decoding: This algorithm searches for the globally optimal path, that is, it comprehensively considers the score matrix output by the linear layer and the transition score of the CRF to find the legal label sequence with the highest probability, rather than simply selecting the highest-scoring label for each token.

[0057] Output: The decoded tag sequence is merged into the final entity fragment.

[0058] The model ultimately outputs a list of identified personal privacy information entities, including the entity text, type, start and end positions in the original text, and confidence level (the confidence level of the rule matching result can be set to a preset value such as 90%-100%). The model parameters involved in processing text using the BERT-BiLSTM-CRF model for personal privacy expert models are shown in Tables 5-8: Table 5 Parameters of the BERT layer

[0059] Table 6 Parameters of BiLSTM layer

[0060] Table 7 Parameters of CRF layer

[0061] Table 8 Training parameters of BERT-BiLSTM-CRF model

[0062] The personal information in this example can include four types of entities: name, ID number, phone number, and address. These four types of entities can be mapped to nine tags, as follows: { 'O': 0, 'B-PER': 1, 'I-PER': 2, 'B-ID': 3, 'I-ID': 4, 'B-PHONE': 5, 'I-PHONE': 6, 'B-LOC': 7, 'I-LOC': 8 } Where O represents an insensitive class, B-PER represents a new name, and I-PER represents a continuation of the previous name, such as "Li" or "Si"; B-ID represents a new ID number, and I-ID represents a continuation of the previous ID number; B-PHONE represents a new phone number, and I-PHONE represents a continuation of the previous phone number; B-LOC represents a new address, and I-LOC represents a continuation of the previous address.

[0063] For example, for expert models of power grid facility information, such as Figure 10 As shown, a power domain dictionary combined with a deep learning model is used to identify sensitive information in texts related to power grid facilities. The deep learning model also employs BERT-BiLSTM-CRF. The main difference between this model and the personal privacy information expert model lies in the maximum sequence length of the power grid critical infrastructure expert model. Since information on power grid critical infrastructure may be longer, a longer sequence length is used, resulting in a smaller batch size. All other settings remain the same. The main parameters of the BERT-BiLSTM-CRF model for the power grid facility information expert model are shown in Tables 9-12 below: Table 9 Bert Layer Parameters

[0064] Table 10 BiLSTM Layer Parameters

[0065] Table 11 CRF Layer Parameters

[0066] Table 12 Training Parameters

[0067] The entity type definitions corresponding to the expert model of power grid facilities information can be divided into the following 7 categories: substation name, power equipment, equipment type and model, technical parameters, voltage level, project, and power company. These 7 types of entities can be mapped to 15 tags, namely: { 'O': 0, 'B-SUBSTATION': 1, 'I-SUBSTATION': 2, 'B-EQUIPMENT': 3, 'I-EQUIPMENT': 4, 'B-DEVICE_MODEL': 5, 'I-DEVICE_MODEL': 6, 'B-TECH_PARAM': 7, 'I-TECH_PARAM': 8, 'B-VOLTAGE_LEVEL': 9, 'I-VOLTAGE_LEVEL': 10, 'B-PROJECT': 11, 'I-PROJECT': 12, 'B-COMPANY': 13, 'I-COMPANY': 14 } In this dictionary, 'O' represents insensitive; the 'B' prefix represents a new entity name; the 'I' prefix represents a continuation of the previous entity name; 'SUBSTATION' represents a substation entity; 'EQUIPMENT' represents a power equipment entity; 'DEVICE_MODEL' represents the equipment model; 'TECH_PARAM' represents technical parameters; 'VOLTAGE_LEVEL' represents the voltage level; 'PROJECT' represents an engineering project entity; and 'COMPANY' represents a power company entity. This example's power industry dictionary can include six sub-dictionaries: substation name dictionary, power equipment name dictionary, equipment model dictionary, voltage level dictionary, technical parameter dictionary, and power company name dictionary.

[0068] In some implementations, if the activated expert model includes the enterprise management data expert model, the step of identifying sensitive information from the received multimodal preliminary parsing data through the activated expert model and outputting the identification result includes: The rule matching module of the enterprise operation data expert model is used to perform rule matching on the received text to be identified to obtain the matching entity; The text to be identified is input into the text classification model of the enterprise management data expert model and subjected to multi-scale one-dimensional convolution processing, fully connected and nonlinear mapping to obtain the recognition result output by the text classification model. The final recognition result is determined based on the matching entity and the recognition result output by the text classification model.

[0069] In this example implementation, the enterprise management data expert model identifies sensitive information through rule matching and text classification models. The enterprise management data expert model specifically identifies commercially sensitive information such as amounts, contracts, tenders, and financial statements. Figure 11As shown, the enterprise operation data parsing data is input into the parallel rule module and the TextCNN classification model. The rule module uses regular expressions to match patterned entities such as amounts, bank account numbers, and contract numbers, and quickly outputs the rule matching results. The TextCNN classification model performs sensitive entity recognition. The structure of the TextCNN classification model is similar to that of the gated network, and the recognition process of the TextCNN classification model is basically the same as that of the TextCNN model in the gated network. The difference is that the enterprise operation data expert model outputs the probability that the text is non-financial text, contract agreement, financial report, invoice, purchase order, or payment information, i.e., the probability of each type of entity. The main parameters of the TextCNN classification model in the enterprise operation data expert model are shown in Table 13. Table 13

[0070] In this example, sensitive entity types in enterprise operational data include non-financial text, contracts and agreements, financial reports, invoices, purchase orders, and payment information. Finally, the final recognition result is determined by fusing rule matching results and TextCNN classification probabilities (individual entity types and confidence scores), making it suitable for rapid review of documents in power companies.

[0071] In some implementations, the sensitivity of various entities is determined by fusing the identification results of different expert models for similar entities and the importance of each entity in the power system, including: For each type of entity, a weighted fusion is performed based on the sensitivity probability, number of entities, and number of modalities of the entity in the recognition results output by each expert model to obtain the sensitivity score for each type of entity. Based on the importance of various entities in the power system and the sensitivity score of each entity, the sensitivity of the multimodal data to be output is obtained.

[0072] In this example implementation, for each entity type, the number of entities of that type in the output recognition results of each expert model is counted (i.e., the same entity type is identified as sensitive multiple times). The modality count of that entity type in the output recognition results of each expert model is also counted. For example, if an entity appears in both the text expert model and image expert model recognition results, the modality count is 2; if it also appears in the table expert model recognition results, the modality count is 3. A sensitivity score for each entity type is obtained by weighted fusion of the sensitivity probability (confidence), the number of entities, and the number of modalities for the same entity type. For example, sensitivity scores are calculated using a weighted formula: Sensitivity score for each entity type = Entity base risk score × Entity confidence score × Entity number level × Multimodal validation score. For instance, the entity base risk score is predefined (e.g., a personal privacy score of 80); the entity confidence score comes from the model output; the entity number level is initially 1, and increases by 0.1 for every 10 additional entities; the multimodal validation score is initially 1, and increases by 0.5 for each additional modality identifying an entity of the same type. The importance score of each entity in the pre-built power knowledge graph (more important entities, such as equipment information in core transformer substations, have higher importance scores) is fused with the sensitivity score of each entity. For example, the importance score and sensitivity score are summed to obtain the sensitivity level of each entity type. The sensitivity level of the multimodal data to be output is determined based on the sensitivity level of each entity type. For example, the maximum sensitivity level of each entity type is used as the sensitivity level of the multimodal data to be output. The quantitative fusion in this example takes into account confidence level, quantity, cross-modal verification, and domain knowledge to avoid subjective bias. For example, multimodal verification improves recognition reliability, knowledge graph enhances industry adaptability, makes sensitivity judgment scientific and reasonable, and guides subsequent hierarchical processing.

[0073] In some implementations, after determining the sensitivity of the multimodal data to be output, the method further includes: The multimodal data to be output is controlled to be output in a hierarchical manner based on its sensitivity. The graded output includes at least one of the following: blocking output, desensitized output, post-recording output, and direct output.

[0074] In this example implementation, different levels of sensitivity can be categorized, such as Level 1 sensitive, Level 2 sensitive, Level 3 sensitive, and non-sensitive. Different processing methods are applied to different sensitivity levels. For example, Level 1 sensitive data is immediately blocked (i.e., output is blocked); Level 2 sensitive data undergoes automatic de-identification (e.g., key data masking); Level 3 sensitive data is recorded, including the output time, data content, and access frequency; and non-sensitive data is passed directly (i.e., immediately output to the user).

[0075] Compared to general solutions, this application demonstrates greater adaptability to the power industry and offers better support for recognizing specialized terminology and knowledge in the power sector. Compared to rule-based solutions, pure rule-based solutions have limited generalization capabilities. Compared to single-model solutions, this application has lower computational overhead and higher accuracy; it allows for modular updates during subsequent maintenance, whereas single-model solutions require a complete overhaul. When expansion is needed, this solution allows for plug-in-style extensions. Furthermore, this application can recognize multimodal data such as images and tables, and can interactively verify text content, enabling more effective identification of sensitive risks in multimodal data. The recognition process of this application can be executed on each segment of output data from a large power model.

[0076] Existing sensitive information identification methods based on regular expressions are optimizations of those based on sensitive words. The original sensitive word-based method relied on the setting of sensitive words, matching only strings in the text that were exactly the same as the sensitive words. Regular expression-based sensitive information identification, on the other hand, relies on the setting of matching rules, matching strings in the text that conform to the rules. This method relies on precise pattern definitions and uses a regular expression engine for fast pattern matching. While sensitive word-based sensitive information identification is the simplest method, it suffers from low flexibility and a high false positive rate. Regular expressions can only identify sensitive information with predefined patterns; they are ineffective for pattern variations or non-standard formats. Furthermore, general regular expressions are difficult to adapt to the specific needs and terminology of the power industry, such as the complex equipment models that are difficult to express with regular expressions. Existing sensitive information identification based on NER (Natural Language Processing) is a sequence labeling task in Natural Language Processing, aiming to identify entities with specific meanings in text and classify them into predefined categories. In sensitive information identification scenarios, NER is used to detect and classify various sensitive entities in text. The entire workflow is as follows: text preprocessing, which involves segmenting, vectorizing, and serializing the initial text; feature encoding, which extracts contextual features using a neural network; sequence labeling, which predicts entity labels for each token; entity extraction, which converts continuous label sequences into entity fragments; and result verification, which verifies the rationality of entities based on rules or statistical methods. The power industry contains a wide variety of sensitive content, and a single NER model struggles to handle multiple types of highly diverse sensitive information simultaneously. Furthermore, updating the model is difficult, requiring retraining to incorporate new knowledge. Deep learning models also face performance challenges due to their high computational cost.

[0077] To address the aforementioned issues, this application utilizes the MOE architecture, combining multiple recognition models. A routing layer network selects which experts to activate based on the input, with each expert ultimately responsible for content sensitivity identification. The processing flow is as follows: multimodal raw data → multimodal data parsing module → modality recognition → multimodal feature routing layer → expert model pool → cross-modal result fusion module → sensitivity level determination and execution. This application constructs a specialized expert system by building an expert hybrid model architecture. The MOE (Mixture-of-Experts) model is a machine learning technique that combines multiple "expert" neural network models into a larger model. The goal of MoE is to improve the accuracy and capability of artificial intelligence systems by combining specialized experts, each specializing in a different sub-domain. Each expert performs in-depth optimization for specific types of sensitive information, forming a collaborative recognition system and achieving expert specialization. This solution includes five expert models: a personal privacy information recognition expert, a power grid facility information recognition expert, a business operation information recognition expert, a table information recognition expert, and an image information recognition expert. Each expert model identifies specific domains, effectively improving the ability to identify multimodal power grid sensitive information. Each expert possesses parallel processing capabilities, supporting multiple experts to process different text fragments simultaneously. It also features a modular design, allowing the expert system to be independently updated and expanded. This application designs a cross-column information fusion scheme for tables. Through table information recognition experts, it achieves collaboration between table headers and cell content. The table header vector is injected with cell features through concatenation, enabling the model to clearly define the semantic attribution of cell text. Through attention weight calculation, the model can automatically associate sensitive information split within rows. This application designs an image anti-interference information extraction mechanism. By setting an SE attention layer, it uses a compression-excitation mechanism to enhance useful text region features and suppress background interference such as device casings and reflections. It achieves a text extraction process of "preliminary positioning - enhanced attention - recognition and extraction," effectively improving the recognition effect of text in images. This application designs a multimodal result fusion mechanism, constructing a quantitative method for calculating cross-modal information sensitivity. By integrating multiple recognition results, it achieves a basis for sensitivity determination.

[0078] In practice, with the construction and use of large-scale power data models, some models have outputted sensitive power-related content, posing a potential risk of leakage of sensitive information in the power industry. Therefore, a method for identifying sensitive power data content is needed. Addressing the problems of low flexibility, low accuracy, and low efficiency in traditional sensitive content detection methods, and considering the actual needs of detecting sensitive power-related content in daily operations, this application proposes an intelligent identification method for sensitive power content based on a hybrid expert model. This method uses a routing layer network (such as a gated network, GateNet) to determine which expert model should handle the task for the input sample, and then uses the expert models to identify the content assigned to them. By combining the intelligent allocation of the gated network with expert identification within a specific subset, the efficiency and accuracy of sensitive power data content identification are improved.

[0079] This invention proposes an intelligent identification method for sensitive content in power data based on the MOE (Mean Exchange Expert) model. Compared to traditional single sensitive information models, it represents a significant technological breakthrough. Unlike directly using large pre-trained models, MoE activates experts only when needed, drastically reducing computational costs. Each expert can be deeply optimized in a specific domain without compromising general performance. Compared to traditional ensemble learning schemes, new experts can be easily added to the system without retraining the entire model. Each expert can specialize deeply, rather than pursuing average performance. Experiments on a large power model demonstrate the following overall performance: the accuracy of the personal privacy information expert model reaches 98.0%, recall 97.5%, and F1 score 97.8%; the accuracy of the power grid facility information expert model reaches 95.8%, recall 95.2%, and F1 score 95.2%; the accuracy of the enterprise operation data expert model reaches 96.3%, recall 95.7%, and F1 score 96.0%; and the image expert model achieves 91.3% accuracy and recall 90.8%, with a single image processing speed ≤1 second. The table expert model achieves a token-level annotation accuracy of 94%, a cross-cell entity recognition rate of 91%, and a single table processing speed of ≤1.5s.

[0080] This application is expected to reduce manual review workload by 70%-80%. Taking a provincial power company as an example, it can save millions of RMB in data security audit manpower costs annually. By identifying and blocking the risk of sensitive data leakage in advance, avoiding business interruptions, and reducing losses from security incidents, the potential losses that can be avoided are estimated to be in the tens of millions of RMB annually.

[0081] Example 2 Based on the same inventive concept, the present invention also provides a power system sensitive data identification device, comprising: The parsing unit is used to perform preliminary parsing of the multimodal data to be output in the power system, and to obtain preliminary multimodal parsing data; The routing unit is used to input the multimodal preliminary parsing data into the routing layer network, perform modality recognition and entity type recognition on the multimodal preliminary parsing data, and dynamically route the multimodal preliminary parsing data to at least one expert model based on the recognition results, so as to activate the corresponding expert model. The identification unit is used to identify sensitive information in the multimodal preliminary analysis data it receives through the activated expert model and output the identification results. The fusion unit is used to determine the sensitivity of the multimodal data to be output based on the fusion of the identification results of various expert models for the same type of entity and the importance of various entities in the power system in the identification results, so as to complete the identification of sensitive information in the multimodal data to be output. The expert models include a text expert model, an image expert model, and a table expert model; the text expert model is used to process power text data; the image expert model is used to process power image data; and the table expert model is used to process power table data.

[0082] In one possible implementation, the text expert model includes a personal privacy information expert model, a power grid facility information expert model, and a business operation data expert model; The expert model for personal privacy information is used to identify sensitive text information related to personal privacy. The power grid facility information expert model is used to identify sensitive text information related to key equipment within the power system. The enterprise operation data expert model is used to identify sensitive text information related to the operation data of various power companies.

[0083] In one possible implementation, the routing layer network includes a modality recognition module and a gating network; the multimodal preliminary parsing data includes at least one of power text parsing data, power image parsing data, and power table parsing data; the routing unit includes: A modality recognition subunit is used to input the preliminary multimodal analysis data into the modality recognition module for modality recognition; The text routing subunit is used to perform entity recognition on the power text parsing data whose modality recognition result is text using the gated network, so as to dynamically route the power text data to the corresponding text expert model; The image routing subunit is used to dynamically route the power image parsing data, which is the modality recognition result, to the image expert model; The table routing subunit is used to dynamically route the power table parsing data, whose modality recognition results are tables, to the table expert model.

[0084] In one possible implementation, the text routing subunit is specifically used for: The power text parsing data is segmented and embedded to obtain the corresponding text matrix; The text matrix is ​​subjected to parallel convolution, linear activation and pooling operations using multiple one-dimensional convolution kernels of different sizes to obtain the text feature vectors corresponding to each one-dimensional convolution kernel. The text feature vectors corresponding to each one-dimensional convolution kernel are concatenated and then randomly discarded to obtain a comprehensive text feature vector. The comprehensive text feature vector is processed by full connection and nonlinear mapping to obtain the probability that the power text data belongs to each text expert model; the power text data is then dynamically routed to the text expert model with the highest probability.

[0085] In one possible implementation, if the activated expert model includes an image expert model, the image expert model includes a compressed excitation attention layer; the multimodal preliminary parsing data includes power image parsing data; the recognition unit includes an image recognition subunit, the image recognition subunit being used for: The received power image parsing data is subjected to parallel feature extraction with multi-scale convolution kernels, linear activation, and global average pooling to obtain multi-scale basic feature vectors. The multi-scale basic feature vector is weighted by a compressed excitation attention layer, and channel attention processing is performed on the multi-scale basic feature vector based on the learned weights to obtain a channel attention-adjusted feature vector. The channel attention adjustment feature vector is binarized and segmented to determine the target text region; the target text region is then used to perform text recognition using a convolutional recurrent neural network based on an attention mechanism to obtain structured text. The structured text is semantically encoded to obtain the corresponding semantic encoding vector; Sensitive entity matching is performed based on the similarity between the semantic encoding vector and each entity in the pre-constructed power knowledge graph, and the sensitive text recognition result of the power image parsing data is output. In the channel attention processing, a preset channel weight threshold is used to filter background interference features.

[0086] In one possible implementation, if the activated expert model includes a table expert model; the multimodal preliminary parsing data includes electricity table parsing data; the identification unit includes a table identification subunit, the table identification subunit comprising: The encoding submodule is used to extract the header text and the corresponding cell text from the parsed data of the power table; to perform semantic encoding on the header text and the corresponding cell text to generate a header semantic vector and a cell text feature vector; to concatenate the header semantic vector and the corresponding cell text feature vector to form a fused feature vector; and to perform layer normalization processing on the fused feature vector. The temporal classification submodule is used to extract bidirectional temporal dependencies, perform cross-cell semantic association, and linear classification sequentially on the normalized fused feature vector through the bidirectional long short-term memory network in the table expert model to obtain the table entity classification result. The optimized labeling submodule is used to decode the globally optimal label sequence based on the transition scores of each entity category label learned from the conditional random field layer in the table expert model and the table entity classification results, using the Viterbi decoding algorithm, and thus obtain the recognition result of the power table parsing data.

[0087] In one possible implementation, the bidirectional long short-term memory network in the table expert model includes a bidirectional LSTM layer and an in-row attention layer connected in sequence; the temporal classification submodule is specifically used for: The time-series correlation feature vectors output by the bidirectional LSTM layer are concatenated column by column to form a row feature matrix, so that the time-series correlation feature vectors corresponding to each cell in the same row of the power table parsing data are located in the same row of the row feature matrix; Based on the row feature matrix, the intra-row attention weights are determined using a single-head attention mechanism, and the row feature matrix is ​​weighted using the intra-row attention weights to obtain an intra-row association enhancement feature matrix. The feature matrix with enhanced in-row correlation is split by column to restore the matrix dimension.

[0088] In one possible implementation, the recognition unit includes a first text recognition subunit, which is used for: If the activated expert model includes the personal privacy information expert model and / or the power grid facility information expert model, the target rule module of the activated expert model is used to perform rule matching on the text to be identified, and the rule recognition result is obtained. The text to be recognized is input into the encoding layer of the activated expert model and then deep encoded to obtain a deep encoded vector. The bidirectional long short-term memory network in the activated expert model is used to extract bidirectional temporal dependencies and perform linear classification on the normalized deep encoding vector to obtain the text entity classification result. Based on the transition scores of each entity category label learned by the conditional random field layer in the activated expert model and the text entity classification results, the Viterbi decoding algorithm is used to decode the globally optimal label sequence, thereby obtaining the recognition result output by the activated expert model. Based on the rule recognition results and the model output recognition results, the final recognition result of the text to be recognized is determined; Wherein, if the activated expert model is a personal privacy information expert model, the target rule module is a regular expression for personal privacy information; if the activated expert model is a power grid facility information expert model, the target rule module is constructed based on various power field dictionaries.

[0089] In one possible implementation, the recognition unit includes a second text recognition subunit, which is used for: If the activated expert model includes the enterprise management data expert model, the rule matching module of the enterprise management data expert model is used to perform rule matching on the received text to be identified to obtain the matching entity; The text to be identified is input into the text classification model of the enterprise management data expert model and subjected to multi-scale one-dimensional convolution processing, fully connected and nonlinear mapping to obtain the recognition result output by the text classification model. The final recognition result is determined based on the matching entity and the recognition result output by the text classification model.

[0090] In one possible implementation, the fusion unit is specifically used for: For each type of entity, a weighted fusion is performed based on the sensitivity probability, number of entities, and number of modalities of the entity in the recognition results output by each expert model to obtain the sensitivity score for each type of entity. Based on the importance of various entities in the power system and the sensitivity score of each entity, the sensitivity of the multimodal data to be output is obtained.

[0091] In one possible implementation, it also includes: The processing module is used to control the hierarchical output of the multimodal data to be output based on the sensitivity of the multimodal data to be output; The graded output includes at least one of the following: blocking output, desensitized output, post-recording output, and direct output.

[0092] Taking a large-scale power model as an example, while it effectively improves individual work efficiency and information acquisition capabilities, its black-box nature means that sensitive content may appear in the output. Therefore, it's necessary to add sensitivity checks at the output boundaries to prevent the leakage of sensitive information and potential harm to the company's interests.

[0093] This application aims to construct an intelligent sensitive information identification and control system for the power industry. Power industry data is characterized by its multimodality, specialization, and high sensitivity, making it difficult to effectively address using traditional single-technology approaches. Therefore, this application innovatively proposes a core architecture of "multimodal hybrid experts." Through an intelligent scheduling hub, different types of data are automatically routed to the most specialized analysis engine for processing, and evidence from different sources is integrated for joint judgment, ultimately achieving accurate, interpretable, and adaptive sensitive data discovery and security control.

[0094] This invention enables efficient and accurate detection of sensitive content, which not only has good direct and indirect economic benefits, but also has good social benefits after large-scale application.

[0095] In the future, with the construction of new power systems and the increasing informatization of power grids, the number of nodes where sensitive power data may be leaked will also increase, making protection requirements more stringent. It is essential to ensure both accurate detection of sensitive content and sufficiently efficient detection without disrupting existing business operations. After successful application in power business scenarios, this invention also has the potential to be extended to other industries. Therefore, the future application prospects of this invention are broad, and it is of great significance for protecting the vital interests of enterprises and maintaining the safe development of the economy and society.

[0096] This application can safeguard citizens' personal information rights, protect personal information, prevent "excessive collection" and "illegal use", effectively implement personal information protection, and safeguard the legitimate rights and interests of electricity users.

[0097] This application can promote the legal and compliant circulation of data elements, and under the premise of ensuring security, enable high-value electricity data that was previously "untouchable and unusable" to be safely de-identified, developed and utilized, empowering electricity market transactions, load forecasting, renewable energy consumption and other businesses, promoting the digital transformation of energy, and breaking down "data silos".

[0098] This application establishes a foundation of trust for the data element market: it provides a reliable technical means to ensure the compliance of data in the process of sharing and trading, laying the foundation for building a safe and trustworthy power data element market.

[0099] This application promotes the integration of industry and research and independent technological innovation, setting an industry benchmark. It serves as a prime example of the deep integration of artificial intelligence technology with the power industry, demonstrating AI's ability to solve core pain points and establishing a technological benchmark for the "AI + Energy" field. It promotes independent controllability by employing open-source models and frameworks, avoiding reliance on foreign commercial models and aligning with the strategic direction of technological self-reliance. Building upon existing technologies, and addressing the specific needs of the power industry, a hybrid sensitive information identification scheme based on the MoE architecture is proposed. Compared to existing technologies, it offers better industry adaptability, higher identification accuracy, and superior computational efficiency, making it an ideal solution for data security governance in the power industry. This application organically combines expert knowledge (rules, dictionaries) with deep learning models, achieving efficient and accurate sensitive information identification through an intelligent routing mechanism, providing a tailored data security protection solution for the power industry.

[0100] Example 3 Based on the same inventive concept, the present invention also provides a power information terminal, including: a power system sensitive data identification device according to any one of embodiments 2.

[0101] In this exemplary embodiment, the sensitive data identification device of this application can be deployed on various power information terminals to identify and process sensitive content in the power data to be output by the power information terminals. Power information terminals can be output terminals of large-scale power models, user information collection terminals, power grid operation log query terminals, etc. Figure 12 As shown, the sensitive data identification device can be deployed between power information terminals and users. It detects data on power information terminals and can be deployed between power large-scale models, user information collection terminals, power grid operation log query terminals, and users. The identification results can also be displayed on the front end of the power information terminal, and identified risk items can be recorded. This application can strengthen the security of critical infrastructure and safeguard the power lifeline: This application provides a powerful "immune system" for core power data, effectively protecting the core secrets of national critical infrastructure such as power grid structure, core equipment parameters, and real-time operating status, combating cyberattacks and data theft, and contributing to social stability and national security. It also raises the overall security level of the industry: The successful implementation of this application can provide reusable technical blueprints and construction experience for other critical information infrastructure industries such as oil, gas, and transportation, generating significant spillover effects.

[0102] Example 4 like Figure 13As shown, the present invention also provides an electronic device, which may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.

[0103] The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the storage medium to implement the corresponding method flow or corresponding function, so as to implement the steps of the power system sensitive data identification method in the above embodiments.

[0104] Example 5 Based on the same inventive concept, this invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). An electronic device readable storage medium is a memory device within an electronic device used to store programs and data. It is understood that the storage medium here can include both built-in storage media within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. Loading and executing one or more instructions stored in the storage medium by the processor can implement the steps of the power system sensitive data identification method described in the above embodiments.

[0105] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0106] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0107] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0108] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the application, but these changes, modifications or equivalent substitutions are all within the scope of protection of the claims pending approval.

Claims

1. A method for identifying sensitive data in a power system, characterized in that, include: Preliminary analysis of the multimodal data to be output from the power system is performed to obtain preliminary multimodal analysis data; The multimodal preliminary parsing data is input into the routing layer network, where modality recognition and entity recognition are performed on the multimodal preliminary parsing data, and the multimodal preliminary parsing data is dynamically routed to at least one expert model based on the recognition results to activate the corresponding expert model. The activated expert model identifies sensitive information from the received multimodal preliminary analysis data and outputs the identification results. Based on the fusion of the identification results of various expert models for similar entities and the importance of various entities in the power system in the identification results, the sensitivity of the multimodal data to be output is determined in order to complete the identification of sensitive information in the multimodal data to be output. The expert models include a text expert model, an image expert model, and a table expert model; the text expert model is used to process power text data; the image expert model is used to process power image data; and the table expert model is used to process power table data.

2. The method according to claim 1, characterized in that, The text expert models include expert models for personal privacy information, expert models for power grid facility information, and expert models for enterprise operational data. The expert model for personal privacy information is used to identify sensitive text information related to personal privacy. The power grid facility information expert model is used to identify sensitive text information related to key equipment within the power system. The enterprise operation data expert model is used to identify sensitive text information related to the operation data of various power companies.

3. The method according to claim 2, characterized in that, The routing layer network includes a modality recognition module and a gating network; the multimodal preliminary parsing data includes at least one of power text parsing data, power image parsing data, and power table parsing data; the multimodal preliminary parsing data is input into the routing layer network, modality recognition and entity recognition are performed on the multimodal preliminary parsing data, and based on the recognition results, the multimodal preliminary parsing data is dynamically routed to at least one expert model, including: The preliminary multimodal analysis data is input into the modality recognition module for modality recognition. The gated network is used to perform entity recognition on the parsed power text data whose modality recognition result is text, so as to dynamically route the power text data to the corresponding text expert model; The power image parsing data, which is the modality recognition result, is dynamically routed to the image expert model; The power table parsing data, which is a table of modality recognition results, is dynamically routed to the table expert model.

4. The method according to claim 3, characterized in that, The gated network is used to perform entity recognition on power text parsing data whose modality recognition results are text, so as to dynamically route the power text parsing data to the corresponding text expert model, including: The power text parsing data is segmented and embedded to obtain the corresponding text matrix; The text matrix is ​​subjected to parallel convolution, linear activation and pooling operations using multiple one-dimensional convolution kernels of different sizes to obtain the text feature vectors corresponding to each one-dimensional convolution kernel. The text feature vectors corresponding to each one-dimensional convolution kernel are concatenated and then randomly discarded to obtain a comprehensive text feature vector. The comprehensive text feature vector is processed by full connection and nonlinear mapping to obtain the probability that the power text data belongs to each text expert model; the power text data is then dynamically routed to the text expert model with the highest probability.

5. The method according to claim 1, characterized in that, If the activated expert model includes an image expert model, the image expert model includes a compressed excitation attention layer; the multimodal preliminary parsing data includes power image parsing data; The process involves using an activated expert model to identify sensitive information from the received multimodal preliminary analysis data and outputting the identification results, including: The received power image parsing data is subjected to parallel feature extraction with multi-scale convolution kernels, linear activation, and global average pooling to obtain multi-scale basic feature vectors. The multi-scale basic feature vector is weighted by a compressed excitation attention layer, and channel attention processing is performed on the multi-scale basic feature vector based on the learned weights to obtain a channel attention-adjusted feature vector. The channel attention adjustment feature vector is binarized and segmented to determine the target text region; the target text region is then used to perform text recognition using a convolutional recurrent neural network based on an attention mechanism to obtain structured text. The structured text is semantically encoded to obtain the corresponding semantic encoding vector; Sensitive entity matching is performed based on the similarity between the semantic encoding vector and each entity in the pre-constructed power knowledge graph, and the sensitive text recognition result of the power image parsing data is output. In the channel attention processing, a preset channel weight threshold is used to filter background interference features.

6. The method according to claim 1, characterized in that, If the activated expert model includes a tabular expert model; the preliminary multimodal parsing data includes electricity tabular parsing data; The process involves using an activated expert model to identify sensitive information from the received multimodal preliminary analysis data and outputting the identification results, including: Extract the header text and corresponding cell text from the parsed data of the electricity table; Semantic encoding is performed on the header text and the corresponding cell text respectively to generate a header semantic vector and a cell text feature vector; The semantic vector of the table header is concatenated with the corresponding text feature vector of the cell to form a fused feature vector; the fused feature vector is then subjected to layer normalization. The bidirectional long short-term memory network in the table expert model is used to sequentially extract bidirectional temporal dependencies, perform cross-cell semantic association and linear classification on the normalized fused feature vector to obtain the table entity classification result. Based on the transition scores of each entity category label learned by the conditional random field layer in the table expert model and the table entity classification results, the Viterbi decoding algorithm is used to decode the globally optimal label sequence, thereby obtaining the recognition result of the power table parsing data.

7. The method according to claim 6, characterized in that, The bidirectional long short-term memory network in the table expert model includes a sequentially connected bidirectional LSTM layer and an in-row attention layer; the cross-cell semantic association includes: The time-series correlation feature vectors output by the bidirectional LSTM layer are concatenated column by column to form a row feature matrix, so that the time-series correlation feature vectors corresponding to each cell in the same row of the power table parsing data are located in the same row of the row feature matrix; Based on the row feature matrix, the intra-row attention weights are determined using a single-head attention mechanism, and the row feature matrix is ​​weighted using the intra-row attention weights to obtain an intra-row association enhancement feature matrix. The feature matrix with enhanced in-row correlation is split by column to restore the matrix dimension.

8. The method according to claim 2, characterized in that, If the activated expert model includes the personal privacy information expert model and / or the power grid facility information expert model, the step of identifying sensitive information from the received multimodal preliminary parsing data through the activated expert model and outputting the identification result includes: The target rule module of the activated expert model is used to perform rule matching on the text to be recognized, and the rule recognition result is obtained. The text to be recognized is input into the encoding layer of the activated expert model and then deep encoded to obtain a deep encoded vector. The bidirectional long short-term memory network in the activated expert model is used to extract bidirectional temporal dependencies and perform linear classification on the normalized deep encoding vector to obtain the text entity classification result. Based on the transition scores of each entity category label learned by the conditional random field layer in the activated expert model and the text entity classification results, the Viterbi decoding algorithm is used to decode the globally optimal label sequence, thereby obtaining the recognition result output by the activated expert model. Based on the rule recognition results and the model output recognition results, the final recognition result of the text to be recognized is determined; Wherein, if the activated expert model is a personal privacy information expert model, the target rule module is a regular expression for personal privacy information; if the activated expert model is a power grid facility information expert model, the target rule module is constructed based on various power field dictionaries.

9. The method according to claim 2, characterized in that, If the activated expert model includes the enterprise operation data expert model, the step of identifying sensitive information from the received multimodal preliminary parsing data through the activated expert model and outputting the identification result includes: The rule matching module of the enterprise operation data expert model is used to perform rule matching on the received text to be identified to obtain the matching entity; The text to be identified is input into the text classification model of the enterprise management data expert model and subjected to multi-scale one-dimensional convolution processing, fully connected and nonlinear mapping to obtain the recognition result output by the text classification model. The final recognition result is determined based on the matching entity and the recognition result output by the text classification model.

10. The method according to claim 2, characterized in that, Based on the fusion of the identification results of various expert models for similar entities and the importance of each entity in the power system, the sensitivity of each entity is determined, including: For each type of entity, a weighted fusion is performed based on the sensitivity probability, number of entities, and number of modalities of the entity in the recognition results output by each expert model to obtain the sensitivity score for each type of entity. Based on the importance of various entities in the power system and the sensitivity score of each entity, the sensitivity of the multimodal data to be output is obtained.

11. The method according to any one of claims 1-10, characterized in that, After determining the sensitivity of the multimodal data to be output, the following steps are also included: The multimodal data to be output is controlled to be output in a hierarchical manner based on its sensitivity. The graded output includes at least one of the following: blocking output, desensitized output, post-recording output, and direct output.

12. A power system sensitive data identification device, characterized in that, include: The parsing unit is used to perform preliminary parsing of the multimodal data to be output in the power system, and to obtain preliminary multimodal parsing data; The routing unit is used to input the multimodal preliminary parsing data into the routing layer network, perform modality recognition and entity type recognition on the multimodal preliminary parsing data, and dynamically route the multimodal preliminary parsing data to at least one expert model based on the recognition results, so as to activate the corresponding expert model. The identification unit is used to identify sensitive information in the multimodal preliminary analysis data it receives through the activated expert model and output the identification results. The fusion unit is used to determine the sensitivity of the multimodal data to be output based on the fusion of the identification results of various expert models for the same type of entity and the importance of various entities in the power system in the identification results, so as to complete the identification of sensitive information in the multimodal data to be output. The expert models include a text expert model, an image expert model, and a table expert model; the text expert model is used to process power text data; the image expert model is used to process power image data; and the table expert model is used to process power table data.

13. The apparatus according to claim 12, characterized in that, The text expert models include expert models for personal privacy information, expert models for power grid facility information, and expert models for enterprise operational data. The expert model for personal privacy information is used to identify sensitive text information related to personal privacy. The power grid facility information expert model is used to identify sensitive text information related to key equipment within the power system. The enterprise operation data expert model is used to identify sensitive text information related to the operation data of various power companies.

14. The apparatus according to claim 13, characterized in that, The routing layer network includes a modality recognition module and a gating network; the preliminary multimodal parsing data includes at least one of power text parsing data, power image parsing data, and power table parsing data; The routing unit includes: A modality recognition subunit is used to input the preliminary multimodal analysis data into the modality recognition module for modality recognition; The text routing subunit is used to perform entity recognition on the power text parsing data whose modality recognition result is text using the gated network, so as to dynamically route the power text data to the corresponding text expert model; The image routing subunit is used to dynamically route the power image parsing data, which is the modality recognition result, to the image expert model; The table routing subunit is used to dynamically route the power table parsing data, whose modality recognition results are tables, to the table expert model.

15. The apparatus according to claim 14, characterized in that, The text routing subunit is specifically used for: The power text parsing data is segmented and embedded to obtain the corresponding text matrix; The text matrix is ​​subjected to parallel convolution, linear activation and pooling operations using multiple one-dimensional convolution kernels of different sizes to obtain the text feature vectors corresponding to each one-dimensional convolution kernel. The text feature vectors corresponding to each one-dimensional convolution kernel are concatenated and then randomly discarded to obtain a comprehensive text feature vector. The comprehensive text feature vector is processed by full connection and nonlinear mapping to obtain the probability that the power text data belongs to each text expert model; the power text data is then dynamically routed to the text expert model with the highest probability.

16. The apparatus according to claim 12, characterized in that, If the activated expert model includes an image expert model, the image expert model includes a compressed excitation attention layer; the multimodal preliminary parsing data includes power image parsing data; the recognition unit includes an image recognition subunit, the image recognition subunit being used for: The received power image parsing data is subjected to parallel feature extraction with multi-scale convolution kernels, linear activation, and global average pooling to obtain multi-scale basic feature vectors. The multi-scale basic feature vector is weighted by a compressed excitation attention layer, and channel attention processing is performed on the multi-scale basic feature vector based on the learned weights to obtain a channel attention-adjusted feature vector. The channel attention adjustment feature vector is binarized and segmented to determine the target text region; the target text region is then used to perform text recognition using a convolutional recurrent neural network based on an attention mechanism to obtain structured text. The structured text is semantically encoded to obtain the corresponding semantic encoding vector; Sensitive entity matching is performed based on the similarity between the semantic encoding vector and each entity in the pre-constructed power knowledge graph, and the sensitive text recognition result of the power image parsing data is output. In the channel attention processing, a preset channel weight threshold is used to filter background interference features.

17. The apparatus according to claim 12, characterized in that, If the activated expert model includes a table expert model; the multimodal preliminary parsing data includes electricity table parsing data; the identification unit includes a table identification subunit, which includes: The encoding submodule is used to extract the header text and the corresponding cell text from the parsed data of the power table; to perform semantic encoding on the header text and the corresponding cell text to generate a header semantic vector and a cell text feature vector; to concatenate the header semantic vector and the corresponding cell text feature vector to form a fused feature vector; and to perform layer normalization processing on the fused feature vector. The temporal classification submodule is used to extract bidirectional temporal dependencies, perform cross-cell semantic association, and linear classification sequentially on the normalized fused feature vector through the bidirectional long short-term memory network in the table expert model to obtain the table entity classification result. The optimized labeling submodule is used to decode the globally optimal label sequence based on the transition scores of each entity category label learned from the conditional random field layer in the table expert model and the table entity classification results, using the Viterbi decoding algorithm, and thus obtain the recognition result of the power table parsing data.

18. The apparatus according to claim 17, characterized in that, The bidirectional long short-term memory network in the table expert model includes a sequentially connected bidirectional LSTM layer and an in-row attention layer; the temporal classification submodule is specifically used for: The time-series correlation feature vectors output by the bidirectional LSTM layer are concatenated column by column to form a row feature matrix, so that the time-series correlation feature vectors corresponding to each cell in the same row of the power table parsing data are located in the same row of the row feature matrix; Based on the row feature matrix, the intra-row attention weights are determined using a single-head attention mechanism, and the row feature matrix is ​​weighted using the intra-row attention weights to obtain an intra-row association enhancement feature matrix. The feature matrix with enhanced in-row correlation is split by column to restore the matrix dimension.

19. The apparatus according to claim 13, characterized in that, The recognition unit includes a first text recognition subunit, which is used for: If the activated expert model includes the personal privacy information expert model and / or the power grid facility information expert model, the target rule module of the activated expert model is used to perform rule matching on the text to be identified, and the rule recognition result is obtained. The text to be recognized is input into the encoding layer of the activated expert model and then deep encoded to obtain a deep encoded vector. The bidirectional long short-term memory network in the activated expert model is used to extract bidirectional temporal dependencies and perform linear classification on the normalized deep encoding vector to obtain the text entity classification result. Based on the transition scores of each entity category label learned by the conditional random field layer in the activated expert model and the text entity classification results, the Viterbi decoding algorithm is used to decode the globally optimal label sequence, thereby obtaining the recognition result output by the activated expert model. Based on the rule recognition results and the model output recognition results, the final recognition result of the text to be recognized is determined; Wherein, if the activated expert model is a personal privacy information expert model, the target rule module is a regular expression for personal privacy information; if the activated expert model is a power grid facility information expert model, the target rule module is constructed based on various power field dictionaries.

20. The apparatus according to claim 13, characterized in that, The recognition unit includes a second text recognition subunit, which is used for: If the activated expert model includes the enterprise management data expert model, the rule matching module of the enterprise management data expert model is used to perform rule matching on the received text to be identified to obtain the matching entity; The text to be identified is input into the text classification model of the enterprise management data expert model and subjected to multi-scale one-dimensional convolution processing, fully connected and nonlinear mapping to obtain the recognition result output by the text classification model. The final recognition result is determined based on the matching entity and the recognition result output by the text classification model.

21. The apparatus according to claim 13, characterized in that, The fusion unit is specifically used for: For each type of entity, a weighted fusion is performed based on the sensitivity probability, number of entities, and number of modalities of the entity in the recognition results output by each expert model to obtain the sensitivity score for each type of entity. Based on the importance of various entities in the power system and the sensitivity score of each entity, the sensitivity of the multimodal data to be output is obtained.

22. The apparatus according to any one of claims 12-21, characterized in that, Also includes: The processing module is used to control the hierarchical output of the multimodal data to be output based on the sensitivity of the multimodal data to be output; The graded output includes at least one of the following: blocking output, desensitized output, post-recording output, and direct output.

23. A power information terminal, characterized in that, include: The power system sensitive data identification device according to any one of claims 12-22.

24. A readable storage medium, characterized in that, It contains an executable program, which, when executed, implements the method as described in any one of claims 1 to 11.

Images