An edge intelligent encryption strategy selection method for power multi-mode data transmission credibility

Abstract

This invention discloses a reliable edge-based intelligent encryption strategy selection method for multi-modal power data transmission. It collects multi-source power data, including image frames, time-series measurements, and log text, and preprocesses, encodes, and maps these data to a unified semantic space. Enhanced embeddings and semantic labels are generated using cosine attention and graph attention networks. Real-time resource status, such as edge node CPU, memory, network round-trip latency, and encryption coprocessor utilization, is collected to obtain resource levels. Based on the semantic labels, enhanced embeddings, and resource levels, the optimal encryption action is determined using a static knowledge matrix and an online scoring model, and a plug-in encryption framework is invoked to output reliable ciphertext. Finally, adaptive strategy migration is performed based on encryption latency and security event feedback. This invention combines multi-modal semantic understanding results with edge resource status for encryption strategy selection, improving edge node resource utilization efficiency and security resilience while ensuring the reliability of power data transmission.

Description

Technical Field

[0001] This invention relates to the fields of secure transmission of power data, edge computing, multimodal data processing, and intelligent encryption decision-making, and particularly to a method, system, edge computing device, and computer-readable storage medium for selecting a trusted edge intelligent encryption strategy for multimodal power data transmission. Background Technology

[0002] With the deepening of digital transformation, power systems are rapidly evolving towards digital twin power grids characterized by intelligent dispatch, ubiquitous sensing, and real-time interaction. Numerous terminal devices are deployed in distribution networks, microgrids, energy consumption sides, and renewable energy access sides, continuously generating multi-source heterogeneous sensing data, including structured measurement data, semi-structured equipment logs, and unstructured images and other multi-modal information. Power grid operation places higher demands on the real-time performance and security of data, especially at the edge. How to efficiently and securely process and transmit multi-modal data has become a key issue affecting system performance and security capabilities.

[0003] Currently, most mainstream edge security protection methods in industrial scenarios rely on static rules or template-based encryption. Some solutions first assess the sensitivity level of fields in multi-source heterogeneous data, and then apply fixed, differentiated encryption templates. For example, attribute-based encryption is used uniformly for highly sensitive fields, while low-sensitivity fields are only masked. These solutions are essentially still based on a preset rule base, and the sensitivity determination and algorithm selection are determined during the deployment phase, lacking the ability to adapt to subsequent semantic changes and fluctuations in edge computing power.

[0004] On the other hand, existing semantic consistency modeling schemes for multi-source heterogeneous data typically address cross-modal semantic conflicts, content propagation optimization, or data representation consistency issues. They fail to further map semantic results to dynamic encryption strategies at the edge, and do not fully consider the balance between security control strength and performance under resource constraints. Therefore, facing the complex conditions of diverse sensing data structures, dynamically changing sensitivity, limited edge resources, and high real-time requirements in power scenarios, existing technologies still have the following shortcomings: a lack of deep semantic understanding capabilities, making it unable to identify the privacy sensitivity behind multi-modal data such as images, time-series waveforms, and log text; decoupling encryption strategies from node computing power, easily creating edge computing bottlenecks during peak load periods; inability to switch encryption strategies online based on sudden changes in grid status, alarm event triggering, or changes in task priority; and a lack of mechanism for converting cross-modal semantic fusion results into security control strategies.

[0005] Therefore, it is necessary to provide an edge intelligent encryption strategy selection method with multi-modal semantic awareness, resource status self-monitoring and policy self-adaptation capabilities, so as to improve the resource utilization efficiency of edge nodes and the system response capability while ensuring the reliability of power data transmission. Summary of the Invention

[0006] The technical problem this invention aims to solve is to address the issues of existing edge security protection schemes relying on static rules, lacking multi-modal semantic understanding, not linking edge resource states, and being difficult to adapt online. This invention provides a reliable edge intelligent encryption strategy selection method for power multi-modal data transmission, enabling the system to dynamically select encryption strategies by combining data semantic features with edge node operational status information, and to achieve continuous adaptive migration through execution feedback.

[0007] The key points of this invention are: First, by using a multimodal semantic understanding mechanism, image frames, time-series measurements, and log text are converted into unified semantic representations and semantic tags that can be used for security decision-making; Second, by using a joint mapping of semantic tags, enhanced embedding, and edge resource levels, encryption strategies that match data semantics and resource status are dynamically selected; Third, by updating the online scoring model through encryption delay and security event feedback, the encryption strategy can adaptively migrate with changes in power grid operating status, resource load, and security risks.

[0008] To address the aforementioned technical problems, this invention provides a reliable edge intelligent encryption strategy selection method for multi-mode power data transmission. The method includes the following steps: collecting multi-source power data and performing preprocessing and multi-mode feature encoding to obtain a normalized multi-mode feature set; constructing a semantic graph based on the normalized multi-mode feature set, and performing semantic reasoning through cosine attention and graph attention networks to generate enhanced embeddings and a set of semantic labels; collecting real-time resource status of edge nodes, performing short-term trend prediction and resource health scoring on the real-time resource status, and discretizing to obtain resource levels; determining the optimal encryption action from an algorithm library based on the enhanced embeddings, the set of semantic labels, and the resource levels through a semantic-resource jointly driven encryption strategy decision model; calling the corresponding encryption algorithm plugin based on the optimal encryption action to encrypt the data to be transmitted, outputting reliable ciphertext, and adaptively transferring the encryption strategy decision model based on encryption latency and security event feedback.

[0009] Compared with existing technologies, the present invention has at least the following beneficial effects: First, the present invention does not select encryption algorithms solely based on field names, data formats, or static sensitivity levels, but rather identifies the true business semantics and sensitivity of data through multimodal semantic coding, semantic graphs, and semantic tags; Second, the present invention incorporates edge resource status into encryption decisions, enabling a dynamic balance between encryption strength and resource load; Third, the present invention achieves policy migration through encryption delay and security event feedback, enabling the system to automatically adjust encryption strategies when there are sudden changes in power grid load or increased attack risks. Attached Figure Description

[0010] Figure 1A schematic diagram of the system architecture selection for a trusted edge-intelligent encryption strategy for multi-mode power data transmission.

[0011] Figure 2 Flowchart of a method for selecting a trusted edge intelligent encryption strategy for multimode power data transmission.

[0012] Figure 3 This diagram illustrates the semantic-resource joint-driven encryption strategy decision-making and adaptive migration. Detailed Implementation

[0013] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Where there is no conflict, the following embodiments and the technical features thereof can be combined with each other.

[0014] like Figure 1 As shown, the edge intelligent encryption strategy selection system of the present invention includes a data acquisition and multi-modal feature encoding unit S1, a semantic graph modeling and cosine attention unit S2, an edge resource perception and state analysis unit S3, a semantic-policy mapping and encryption decision-making unit S4, and an encryption execution and adaptive policy migration unit S5. This system takes multi-source power data as input, outputs trusted encrypted data, and continuously updates the policy model through encryption execution feedback.

[0015] In this embodiment, the main innovation of the patent focuses on the mapping relationship between multimodal semantic understanding results and encryption strategy selection. Therefore, the following implementation methods prioritize multimodal data encoding, semantic graph modeling, resource level assessment, semantic-resource joint decision-making, and feedback transfer.

[0016] In phase S1, the system synchronously accesses multi-source data streams from cameras, smart meters, and protection devices via an edge gateway, where image frames are represented as... Time series measurement is represented as The log text is represented as To ensure consistency between different modalities in the temporal and semantic domains, the system sequentially performs identifier parsing, timestamp alignment, and wavelet thresholding for denoising. Subsequently, image frames are processed by a convolutional neural network to output spatial features, temporal measurements are processed by a gated recurrent unit to output dynamic embeddings, and log text is used to generate context vectors through a pre-trained language model.

[0017]

[0018] The above formula is used to obtain the initial feature representations for the three modalities: image, time series, and log. Wherein, Indicates the first The sampling period or the first sampling period One data sample; Indicates the first Image frame data acquired in one sampling period; Indicates the first Time series measurement data acquired in each sampling period; Indicates the first Log text data acquired in each sampling period; The feature vector representing the image modality, with superscript Represents image modality; The feature vector representing the temporal mode, with superscript Indicates the timing measurement mode; The feature vector representing the log text modality, with superscript Indicates log text modality; The parameter is Convolutional neural network encoder, subscript Indicates its corresponding image modality; The parameter is The gated loop unit encoder, subscript Indicates its corresponding timing mode; The parameter is The pre-trained language model encoder subscript This indicates the corresponding log text modality.

[0019] To eliminate the differences in dimensionality and scale of features across different modalities, the system maps the above vectors to a unified semantic space via a multilayer perceptron and performs Z-score normalization:

[0020]

[0021] The above formula is used to transform feature vectors from different modalities into the same semantic space and eliminate the influence of dimensions. Wherein, This represents the modal index, which is the set of values ​​it can take. These correspond to the image modality, time series measurement modality, and log text modality, respectively. Indicates the first Each sample in modality The original encoded features below; Indicates for modal The multilayer perceptron mapping function, superscript This indicates that the mapping function is related to the mode. correspond; Representing modes The mean of the sample features; Representing modes The standard deviation of the sample characteristics; Representing modes The feature vector after unified semantic space mapping and normalization is indicated by the hat symbol, which means that the feature has been normalized.

[0022] The normalized multimodal feature set is represented as:

[0023]

[0024] The above formula is used to form the first... A unified multimodal feature set corresponding to each sampling period. Among them, Indicates the first The multimodal feature set of each sample; square brackets indicate that the normalized features of the three modalities of image, time series and log are combined into a joint representation; , and They represent the first The normalized features of each sample across image, time-series, and log modalities. This set provides multimodal embeddings at the same scale for subsequent semantic graph construction.

[0025] In phase S2, the system is based on A semantic graph is constructed, and each data point is assigned a measurable and secure semantic label using cosine attention and graph attention networks. Specifically, the system uses normalized multimodal features as the node set. An initial semantic graph is constructed using the cosine similarity between nodes as edge weights. .node With nodes The edge weights between them are represented as follows:

[0026]

[0027] The above formula is used to measure the semantic similarity between different nodes. Wherein, and These represent the node numbers in the semantic graph; Represents a node With nodes The boundary weight between; and Representing nodes respectively and nodes The normalized eigenvectors; the dot product symbol represents the vector dot product; This represents the vector norm. The larger the edge weight, the more semantically similar the data corresponding to the two nodes.

[0028] After constructing the semantic graph, the system introduces a graph attention network to iteratively propagate node features:

[0029]

[0030] The above formula is used to update the node embedding based on the semantic association strength of neighboring nodes. Wherein, Represents a node In the Embedded after layer update; Indicates the layer number of the graph attention network; Representing neighboring nodes In the Layer embedding; Represents a node The set of neighboring nodes; Indicates the first Layer nodes For nodes Attention weights; Indicates the first The learnable transformation matrix of the layer; Represents a non-linear activation function; Indicates a node The index variable used when summing the neighboring nodes; Represents a node With neighboring nodes The right to the side.

[0031] After multiple propagation layers, the system outputs semantic labels through a linear classifier:

[0032]

[0033] The above formula is used to determine the nodes. The semantic tags corresponding to the data. Among them, Represents a node semantic tags; Represents a set of semantic tags; Indicates the candidate categories in the label set; Indicate category The corresponding classification weights; Represents a node The enhanced embedding obtained after the final layer of the graph attention network, where the superscript... This indicates the final layer, not the log modality. The semantic tag set can include tags for high sensitivity, low real-time performance, critical equipment alarms, general measurements, and tags related to the context of power business tasks. Enhanced embedding. With semantic tags It is the core input for subsequent encryption strategy selection.

[0034] In phase S3, the system transforms instantaneous resource samples into resource levels usable for decision-making. The system continuously collects real-time load data from edge nodes at preset intervals.

[0035]

[0036] The above formula is used to represent the edge node at time t. The resource state vector. Where, Indicates the time of resource sampling; Indicates time Real-time resource status; Indicates time CPU utilization; Indicates time Memory usage; Indicates time Network round-trip latency; Indicates time The utilization rate of the encryption coprocessor. These metrics collectively reflect the resource capacity of edge nodes when performing encryption tasks.

[0037] To assess short-term load trends, the system introduces a weighted sliding window forecast:

[0038]

[0039] The above formula is used to predict resource trends in the next moment based on the resource status over a recent period. Wherein, Indicates time Predicted resource status; Indicates the length of the sliding window; Indicates the lagging index within the window; Indicates time The historical resource status; Indicates the first Exponential decay weighting of historical resource states; This represents the attenuation coefficient, and The range of values ​​is arrive The resource status at more recent times has a greater impact on the prediction results.

[0040] The system further calculates the resource health score:

[0041]

[0042] The above formula is used to compress multidimensional resource status into a single comparable score. Wherein, Indicates time Resource health score; Represents the weight vector; , , and These represent the weights corresponding to CPU utilization, memory usage, network round-trip latency, and encryption coprocessor utilization, respectively. This weight vector can be set manually based on experience or obtained through optimized searching using historical data.

[0043] The system discretizes the data based on the resource health score:

[0044]

[0045] The above formula is used to map continuous resource health scores to resource levels. Wherein, Indicates time Resource level; This indicates that the resource status is good and can withstand high-overhead encryption strategies; This indicates a medium resource status. This indicates that resources are strained, and a lightweight encryption strategy should be prioritized. and This represents the discretization threshold, which can be adaptively adjusted based on the 95th percentile of historical resource health scores.

[0046] In phase S4, the system, focusing on the core protection priorities of this invention, maps multimodal semantic results and edge resource levels together into encryption actions. Specifically, the system employs a two-layer mapping mechanism, firstly through a static knowledge matrix... Provide rule priors, and then use an online scoring model. Dynamically fine-tune the prior rules:

[0047]

[0048] The above formula is used to output the optimal encryption action. Wherein, Represents a node The corresponding algorithmic score vector; This represents a static knowledge matrix, used to determine the knowledge based on semantic tags. and resource level Provide rules prior; The parameter is An online scoring model for use based on enhanced embeddings and resource level Correcting the prior rules; This represents candidate encryption actions in the algorithm library; Represents an algorithm library; Indicates candidate encryption action Corresponding rating; Represented as nodes The optimal encryption action selected for the corresponding data. Algorithm library. It can include algorithms such as AES, SM4, ChaCha20, RSA, ECC, ABE, PRE, PHE, FHE, and secure hash signatures.

[0049] To balance security levels and system latency, the online scoring model... The loss function is designed as follows:

[0050]

[0051] The above formula is used to ensure that the encryption strategy simultaneously satisfies semantic security requirements and resource overhead constraints. Wherein, Represents the loss function; and These represent two trade-off factors, used to control the importance of security fit and resource cost, respectively. Indicates the selected encryption action With semantic tags The degree of security matching; Indicates the selected encryption action In resource level The system can select different encryption strategies based on the resource cost or load time under different conditions. This loss function allows the system to choose between highly sensitive data, ordinary measurement data, resource-scarce conditions, and resource-sufficient conditions.

[0052] In phase S5, the system employs a plug-in encryption framework to execute the encryption actions decided in phase S4. For the data to be transmitted... The system calls the unified encryption interface:

[0053]

[0054] The above formula is used to generate trusted ciphertext. Wherein, Representing data The ciphertext obtained after encryption; Represents a node or the One piece of data to be transmitted; This refers to the key material used in the encryption process, which may include a symmetric key. Attribute key set Public key or private key ; This represents the optimal encryption action output by S4; This represents the unified interface provided by the plug-in encryption framework. The system, based on... It automatically selects the key type and completes key acquisition, encryption operation, ciphertext labeling, and ciphertext output.

[0055] The system records encryption delays during operation. Security incidents And calculate the sliding index after the set self-check cycle:

[0056]

[0057] The above formula is used to evaluate the performance and security of the current encryption strategy over a period of time. Wherein, Indicates the number of data entries or the length of the statistical window within a self-test cycle; Indicates the first Encryption delay of data; Indicates the first Security event tags or security event counts associated with each data item; This indicates the average encryption latency during the self-test period; This indicates the average security incident rate within the self-inspection period.

[0058] when Exceeding the performance threshold or When the risk threshold is exceeded, the system triggers a policy migration, calling the incremental learner to update the online scoring model in reverse:

[0059]

[0060] The above formula is used to update the parameters of the online scoring model. Wherein, Indicates online rating model Parameters; Indicates the learning rate; Represents the loss function For parameters The gradient is calculated. After the update is completed, the system immediately reloads the mapping model without stopping the service. Subsequently, the encrypted data carrying the security tag and security measurement information is sent to the cloud or local data pool.

[0061] like Figure 2As shown, the method flow of the present invention may include steps 101 to 108. Step 101: Collect multi-source power data; Step 102: Perform preprocessing and multi-mode feature encoding; Step 103: Perform unified semantic space mapping and normalization; Step 104: Construct a semantic graph and perform graph attention inference; Step 105: Collect the resource status of edge nodes and perform health scoring; Step 106: Select an encryption strategy based on semantic labels and resource levels; Step 107: Perform encryption and output trusted ciphertext; Step 108: Collect latency and security event feedback and update the strategy model. Step 108 feeds back to step 106 through adaptive strategy transfer to form a closed-loop optimization.

[0062] like Figure 3 As shown, semantic branch output enhances embedding. and semantic tag set Resource branches output real-time resource status. and resource level The encryption strategy decision engine utilizes a static knowledge matrix. With online rating models Generate algorithm score vectors and extract them from the algorithm library. Select the optimal encryption action After the encryption execution module outputs trusted ciphertext, it feeds back the encryption latency and security event rate to the feedback evaluation module and the incremental learner, and then returns to update the online scoring model, realizing parameter updates and model reloading.

[0063] As can be seen from the above embodiments, the core of this invention is not simply listing various encryption algorithms into an algorithm library, but rather identifying data security semantics through multimodal semantic understanding, selecting encryption actions in conjunction with edge resource states, and continuously adapting the strategy model to changes in the power grid operating environment through feedback migration. This technology chain can improve the reliability, security resilience, and edge processing efficiency of multimodal power data transmission.

Claims

1. A method for selecting a trusted edge-intelligent encryption strategy for multi-mode power data transmission, characterized in that, Includes the following steps: S1. Collect multi-source power data, and preprocess and encode the multi-source power data to obtain a normalized multi-mode feature set; S2, construct a semantic graph based on the normalized multimodal feature set, and perform semantic reasoning on the semantic graph through cosine attention and graph attention networks to generate an enhanced embedding and semantic label set; S3, collect the real-time resource status of edge nodes, perform short-term trend prediction, resource health scoring and level discretization on the real-time resource status to obtain the resource level; S4. Based on the enhanced embedding, the semantic tag set, and the resource level, the optimal encryption action is determined from the algorithm library through a semantic-resource jointly driven encryption strategy decision model. S5, based on the optimal encryption action, call the corresponding encryption algorithm plugin to encrypt the data to be transmitted, output trusted ciphertext, and perform adaptive policy migration on the encryption policy decision model according to encryption delay and security event feedback.

2. The method for selecting a trusted edge intelligent encryption strategy for multi-mode power data transmission according to claim 1, characterized in that, The multi-source power data in S1 includes image frames. Time series measurement and log text ;in, Indicates the sampling period or sample number; the multimodal feature encoding satisfies at least the following relationship: ； in, , and Represent the encoded features of the image modality, time series measurement modality, and log text modality, respectively, with superscripts. , , These represent the corresponding modes. , and These represent the parameters of the corresponding encoder.

3. The method for selecting a trusted edge intelligent encryption strategy for multi-mode power data transmission according to claim 2, characterized in that, S1 further includes mapping the modality-coded features to a unified semantic space and performing normalization processing to obtain the following multimodal feature set: ； in, Indicates modal index, Representing modes The normalization characteristics, and Representing modes The characteristic means and standard deviation, Indicates the first The multi-modal feature set corresponding to each sampling period.

4. The method for selecting a trusted edge intelligent encryption strategy for multi-mode power data transmission according to claim 1, characterized in that, S2 includes constructing a semantic graph using normalized multimodal features as nodes and cosine similarity between nodes as edge weights, and performing graph attention propagation on the semantic graph; wherein, nodes and nodes The edge weights and graph attention propagation between them must satisfy at least the following relationship: ； in, Represents a node With nodes semantic similarity edge weights Represents a node In the Layer embedding, Represents a node The set of neighboring nodes, Indicates the first Layer nodes For nodes Attention weights Indicates the first The learnable transformation matrix of the layer.

5. The method for selecting a trusted edge intelligent encryption strategy for multi-mode power data transmission according to claim 1, characterized in that, The real-time resource status in S3 includes CPU utilization, memory usage, network round-trip latency, and encryption coprocessor utilization. The resource level is obtained by weighted sliding window prediction, resource health scoring, and threshold discretization of the real-time resource status, and is used to constrain the resource overhead of subsequent encryption strategy selection.

6. The method for selecting a trusted edge intelligent encryption strategy for multi-mode power data transmission according to claim 1, characterized in that, The encryption strategy decision model in S4 includes a static knowledge matrix. and online rating models The static knowledge matrix is ​​used to provide prior knowledge of encryption policy rules based on semantic tags and resource levels. The online scoring model is used to dynamically score and correct the prior knowledge of rules based on enhanced embedding and resource levels to form an algorithm scoring vector and determine the optimal encryption action, satisfying at least the following relationship: ； in, Represents a node The corresponding data's algorithmic scoring vector, Indicates semantic tags, Indicates time Resource level, This indicates the last layer of semantically enhanced embedding. Represents an algorithm library. Represented as nodes The optimal encryption action selected for the corresponding data.

7. The method for selecting a trusted edge intelligent encryption strategy for multi-mode power data transmission according to claim 6, characterized in that, The algorithm library It includes at least one of the following: symmetric encryption algorithm, public-key encryption algorithm, attribute-based encryption algorithm, proxy re-encryption algorithm, partially homomorphic encryption algorithm, fully homomorphic encryption algorithm, and secure hash signature algorithm; wherein, the symmetric encryption algorithm includes AES, SM4, or ChaCha20, and the public-key encryption algorithm includes RSA or ECC.

8. The method for selecting a trusted edge intelligent encryption strategy for multi-mode power data transmission according to claim 1, characterized in that, S5 includes encapsulating different encryption algorithms using a pluggable encryption framework, automatically selecting key type and algorithm pluggins based on the optimal encryption action, performing key acquisition, encryption operation, ciphertext labeling, and trusted ciphertext output on the data to be transmitted, and updating the online scoring model according to the following relationship when the encryption delay exceeds the performance threshold or the security event rate exceeds the risk threshold: ； in, Indicates the parameters of the online scoring model. Indicates the learning rate. This represents the loss function used to combine security matching degree and resource cost.

9. A trusted edge-intelligent encryption strategy selection system for multi-mode power data transmission, characterized in that, include: The system comprises a data acquisition and multimodal feature encoding unit, a semantic graph modeling and cosine attention unit, an edge resource perception and state analysis unit, a semantic-policy mapping and encryption decision-making unit, and an encryption execution and adaptive policy migration unit; wherein each unit is configured to perform the method described in any one of claims 1 to 8.

10. An edge computing device, characterized in that, The device includes a processor, a memory, and a communication interface. The memory stores a computer program, which, when executed by the processor, causes the edge computing device to perform the method described in any one of claims 1 to 8.

Images