Cross-domain data collaborative management method based on edge computing

Abstract

The invention discloses a cross-domain data collaborative management method based on edge computing. The method comprises the following steps: S1, acquiring data to form a time series data set; s2, extracting features of time sequence data at edge nodes, and generating a state vector sequence; s3, inputting the state vector sequence into an edge decision network for weighted aggregation to generate a collaborative state sequence; s4, detecting a behavior change point of the collaborative state sequence by adopting a cumulative sum control chart method, and constructing a state mapping sequence; s5, based on the state mapping sequence, constructing a region index set through discrete wavelet transform; s6, inputting the regional index set into an edge decision network to generate a management instruction sequence; and S7, executing the management instruction sequence at the edge node, and updating the edge decision network based on an execution result. According to the method, multiple intelligent algorithms are fused to realize data collaborative management of cross-domain edge nodes, and the response efficiency and decision-making ability in an edge environment are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cross-domain data management, and particularly relates to a cross-domain data collaborative management method based on edge computing. BACKGROUND

[0002] Under the background of wide deployment of distributed intelligent terminals, Internet of Things devices and edge nodes, real-time collaboration and fine management of data in a cross-domain environment are gradually becoming an important research direction of edge computing. Traditional data collaboration methods mainly rely on central cloud for unified scheduling and processing, which has the problems of high response delay, large communication bandwidth occupation and difficult privacy risk control, and is difficult to meet the real-time and controllability requirements under the heterogeneous collaboration of multiple edge nodes.

[0003] In the prior art, some methods attempt to implement data processing by deploying lightweight models on the edge side to reduce the burden on the cloud, but still face the following key challenges: first, the edge nodes in the cross-domain environment have diverse functions and complex data types, and lack unified modeling and classification mechanisms, resulting in low data fusion efficiency; second, in the process of data change, there is a lack of effective time series behavior change detection means, which cannot accurately identify the mutation point and link the access strategy adjustment; third, in the regional collaboration process, there is a lack of efficient feature extraction and index modeling mechanism, which makes it difficult to dynamically depict the data evolution trend; fourth, in the generation and management instruction stage, the strategy optimization path is single, and lacks a feedback-driven adaptive updating mechanism, which affects the convergence efficiency and execution accuracy of the overall strategy.

[0004] In addition, the existing strategy decision process relies on fixed rules or simple network structures, which is difficult to achieve dynamic learning and efficient action selection in a complex state space, especially in the resource-limited scenario of edge nodes, which is more likely to cause problems such as strategy overfitting and model update delay. Therefore, in the complete process of "feature extraction - behavior recognition - index construction - strategy execution", the existing scheme still lacks a unified, efficient and scalable collaborative management mechanism.

[0005] Therefore, how to provide a cross-domain data collaborative management method based on edge computing is a problem that those skilled in the art need to solve. SUMMARY

[0006] An object of the present application is to provide a cross-domain data collaborative management method based on edge computing. The present application combines convolution feature extraction, attention mechanism, behavior detection and strategy gradient optimization, realizes intelligent collaboration of data between cross-domain edge nodes, has the advantages of fast response, accurate recognition and strong decision optimization capability, and improves the data management and collaboration efficiency in the edge environment.

[0007] According to the cross-domain data collaborative management method based on edge computing according to the embodiment of the present application, the following steps are included. S1, acquire original data, node state and access control parameter of each edge node in cross-domain environment, and form time series data set in time sequence; S2, perform one-dimensional convolution operation on the time series data of each time step at the edge node to extract features, one-hot encode the preset label, and connect the encoding result and the convolution result according to the feature dimension to generate a state vector sequence; S3, input the state vector sequence into the edge decision network, calculate the correlation weight between each time step through the multi-head attention mechanism, and generate a collaborative state sequence by weighting and aggregating the state vector sequence according to the weight; S4, perform a first-order difference operation on the collaborative state sequence, detect the behavior change point using the cumulative sum control chart method, and time-align the detection result with the access control parameter to construct a state mapping sequence; S5, divide the state mapping sequence into multiple regional state sequences according to the spatial distribution of the edge nodes, generate wavelet feature sequences using discrete wavelet transform, and calculate the mean and variance of the wavelet feature sequences to construct a regional index set; S6, input the regional index set into the edge decision network, and use the policy gradient algorithm to evaluate the policy and determine the action to generate a management instruction sequence; S7, execute the management instruction sequence at the edge node, and update the gradient clipping of the edge decision network parameters based on the execution result.

[0008] Optionally, the original data includes sensor sampling data, network communication data, log information, task execution state and user request record, the node state represents the running parameter of the edge node at a specific time step, the access control parameter represents the configuration content of the data sharing permission between the edge nodes, the preset label represents discrete identification information for distinguishing the edge node category, function role or domain, and the behavior change point represents a mutation position in the state evolution process.

[0009] Optionally, the S2 specifically includes: S21, divide the time series data set into multiple data segments in the edge node according to the time step, and arrange each data segment in a vector according to a preset dimension to form a structural feature data; S22, extract local time series features of the structural feature data using a multi-scale one-dimensional convolution method, perform parallel convolution processing according to different convolution kernel sizes and fixed sliding steps during the convolution process, and perform batch normalization processing on the convolution result to construct a convolution feature representation; S23, perform one-hot encoding operation on the preset label associated with the edge node, convert the encoding result into a fixed-length compact vector through embedding mapping to form a label feature representation; S24. The convolutional feature representation and label feature representation at each time step are structurally concatenated according to the feature dimension, and the dominant component of the concatenation result is extracted by principal component analysis to form a fused feature vector. S25. Arrange the fused feature vectors in chronological order to construct a state vector sequence.

[0010] Optionally, the principal component analysis specifically includes: Arrange the splicing results in chronological order to construct a sample data matrix. Calculate the covariance relationship between all feature dimensions based on this matrix and output the feature covariance matrix. The Jacobi method is used to perform eigenvalue decomposition on the feature covariance matrix, extracting the principal direction vectors and eigenvalues ​​corresponding to all feature dimensions, and then sorting the principal direction vectors by dimensionality reduction according to the eigenvalues. Based on the set feature retention threshold, select the main direction vector that meets the conditions and construct the feature mapping matrix; Perform matrix multiplication on the sample data matrix and the feature mapping matrix to generate a dimension-compressed feature vector as the dominant component.

[0011] Optionally, S3 specifically includes: S31. For each time step of the state vector sequence, construct an encoding vector representing the time order, and concatenate the encoding vector with the corresponding state vector according to the feature dimension to generate an enhanced representation sequence. S32. Input the augmented representation sequence into the edge decision network, and construct multiple attention channels in the edge decision network. Each channel performs an affine transformation operation on the augmented representation sequence to obtain the corresponding query vector, key vector and value vector. S33. Perform element-wise dot product operation on the query vector and key vector of all time steps in each attention channel to generate the correlation score matrix between time steps. Perform numerical scaling on the correlation score matrix according to the set scaling factor, and calculate the attention weight through the Softmax function. S34. In each attention channel, the value vector is weighted and summed according to the attention weight to obtain the attention representation corresponding to the channel. S35. Concatenate the attention representations extracted from all attention channels according to the feature dimension, and perform feature cross operation on the concatenation result to generate a co-state sequence.

[0012] Optionally, S4 specifically includes: S41. Perform a first-order difference operation on the cooperative state sequence to generate a difference result sequence in chronological order; S42. The cumulative sum control chart method is used to detect the behavioral change points of the difference result sequence. The difference result sequence is processed by time step, and the cumulative processing results are compared with the set control limits. The time steps that exceed the control limits are identified and extracted as behavioral change points. S43. Arrange the behavior change points in chronological order to generate an event sequence, retrieve the access control parameters corresponding to the time step in the time series data set, and align them with the event sequence in time to construct a state mapping sequence.

[0013] Optionally, the cumulative sum control chart method specifically includes: Perform cumulative summation operation based on the difference result sequence, and construct a cumulative statistical sequence according to time steps; The mean and standard deviation of the difference results are calculated at multiple time steps in the difference result sequence. The mean is set as the baseline. The upper and lower control limits are constructed by adding or subtracting the standard deviation of the mean by a preset multiple, forming the control limit interval. The statistical values ​​at each time step in the cumulative statistics sequence are compared with the control limit interval, the time step that first deviates from the control limit interval is marked, and the time step is organized into a set of candidate change points. Construct a set of feature vectors containing the difference results for each candidate change point at the corresponding time step, the cumulative statistics for that time step and adjacent time steps, and call the isolated forest algorithm to calculate the anomaly score for each feature vector. Time steps with abnormal scores exceeding a set threshold are selected as behavioral change points.

[0014] Optionally, the computation process of the isolated forest algorithm specifically includes: Each feature vector in the feature vector set is used as a sample node, and multiple random subsets of equal size are constructed by random sampling without replacement. For each random subset, a corresponding random tree structure is constructed. Based on the value range of the random subset in each dimension of features, the partition dimension and partition threshold are randomly selected. The sample is divided into left and right child nodes according to the selection result, and the partition operation is recursively executed. In each recursive partition, when the number of samples in a node is less than a set threshold or the preset maximum tree depth is reached, the partitioning stops and the node is retained as a leaf node, thus completing the construction of the random tree structure. The path depth of each sample node in all random tree structures is statistically analyzed, and the average path length of that sample node in all random trees is calculated. An anomaly scoring function is constructed based on the average path length of the sample nodes and the total number of samples, and anomaly scores are calculated for all sample nodes.

[0015] Optionally, S5 specifically includes: S51. Divide the state mapping sequence into multiple regional state sequences according to the spatial distribution of edge nodes, and organize the state sequence of each region in chronological order. S52. Perform discrete wavelet transform on the state sequence of each region to generate a multi-scale wavelet feature sequence; S53. Calculate the mean and variance in the time dimension for the wavelet feature sequences respectively, and combine the mean and variance according to the feature dimension to form the regional index of the corresponding region. S54. Arrange the regional indicators of each region in a predetermined order to construct a set of regional indicators.

[0016] Optionally, S6 specifically includes: S61. Input the set of regional indicators into the edge decision network, and organize each indicator vector into a strategy vector sequence according to the region. S62. Perform multi-layer perceptual transformation operation on each policy vector. In the first layer, perform affine transformation and ReLU function activation operation on the policy vector in sequence and output the result. S63. Starting from the second layer, affine transformation and ReLU function activation operations are performed sequentially based on the output of the previous layer, and the activation result is residually connected with the output of the previous layer as the current output. The output of the last layer is used as the policy feature representation. S64. Construct a policy probability distribution based on policy feature representation, construct a policy loss function based on the deviation between the policy distribution and the target action, and use a policy gradient algorithm to determine the policy update direction. The target action represents the expected action instruction pre-set in the cross-domain data collaboration process. S65. The policy gradient is weighted according to the action selection probability automatically allocated by the edge decision network, and the management instructions corresponding to each region are generated based on the updated policy probability distribution to form a management instruction sequence.

[0017] The beneficial effects of this invention are: First, by deploying a feature extraction network based on one-dimensional convolution and label encoding on the edge node side, this invention can effectively capture local dynamic features in time-series data and fuse node category information to generate a state vector sequence, thereby improving the expressive power and context awareness of edge data processing.

[0018] Secondly, by using a multi-head attention mechanism to perform temporal correlation modeling of the state vector sequence, and combining first-order difference and cumulative sum control charts to achieve accurate detection of behavioral change points, and then using the isolated forest algorithm to complete mutation screening, the response of cross-domain data in behavior monitoring and access control is more sensitive and accurate.

[0019] Finally, by introducing discrete wavelet transform to extract local fluctuation features, constructing a set of regional indicators, and combining multilayer sensing structure and policy gradient algorithm for policy optimization, the accurate generation and adaptive updating of management instructions are realized, improving the efficiency and stability of collaborative decision-making between cross-domain edge nodes, thereby effectively enhancing the data intelligent collaborative management capability in the edge computing environment. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Fig. 1 This is a flowchart of a cross-domain data collaborative management method based on edge computing proposed in this invention; Fig. 2 This is a flowchart of feature fusion and principal component analysis for a cross-domain data collaborative management method based on edge computing proposed in this invention. Fig. 3 This is a flowchart illustrating the behavior detection and anomaly recognition process of a cross-domain data collaborative management method based on edge computing proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figs. 1-3 A cross-domain data collaborative management method based on edge computing includes the following steps: S1. Obtain the raw data, node status and access control parameters of each edge node in the cross-domain environment, and assemble them into a time-series data set in chronological order; S2. At the edge nodes, perform one-dimensional convolution operation on the time series data of each time step to extract features, perform one-hot encoding on the preset labels, and connect the encoding results with the convolution results according to the feature dimension to generate a state vector sequence. S3. Input the state vector sequence into the edge decision network, calculate the correlation weight between each time step through the multi-head attention mechanism, and perform weighted aggregation of the state vector sequence based on the weight to generate a collaborative state sequence. S4. Perform a first-order difference operation on the cooperative state sequence, use the cumulative sum control chart method to detect behavioral change points, and align the detection results with the access control parameters in time to construct a state mapping sequence. S5. Divide the state mapping sequence into multiple regional state sequences according to the spatial distribution of edge nodes, use discrete wavelet transform to generate wavelet feature sequences, and calculate the mean and variance of the wavelet feature sequences to construct a set of regional indicators. S6. Input the set of regional indicators into the edge decision network, use the policy gradient algorithm to evaluate the policy and determine the action, and generate a sequence of management instructions. S7. Execute the management instruction sequence at the edge node, and perform gradient pruning and updating of the edge decision network parameters based on the execution results.

[0023] In this embodiment, the raw data includes sensor sampling data, network communication data, log information, task execution status and user request records. The node status represents the running parameters of the edge node at a specific time step. The access control parameters represent the configuration content that controls the data sharing permissions between edge nodes. The preset label represents discrete identification information used to distinguish edge node categories, functional roles or domains. The behavior change point represents the abrupt change position that occurs during the state evolution process.

[0024] In this embodiment, S2 specifically includes: S21. Divide the time series data set into multiple data segments in the edge nodes according to time steps, and arrange each data segment into vectors according to a preset dimension to form structural feature data. S22. Local temporal features of structural feature data are extracted using a multi-scale one-dimensional convolution method. During the convolution process, parallel convolution processing is carried out according to different set convolution kernel sizes and fixed sliding strides. Batch normalization processing is performed on the convolution results to construct convolution feature representation. S23. Perform one-hot encoding on the preset labels associated with the edge nodes, and convert the encoding results into a compact vector of fixed length through embedding mapping to form a label feature representation; S24. The convolutional feature representation and label feature representation at each time step are structurally concatenated according to the feature dimension, and the dominant component of the concatenation result is extracted by principal component analysis to form a fused feature vector. S25. Arrange the fused feature vectors in chronological order to construct a state vector sequence.

[0025] In this embodiment, the principal component analysis specifically includes: Arrange the splicing results in chronological order to construct a sample data matrix. Calculate the covariance relationship between all feature dimensions based on this matrix and output the feature covariance matrix. The Jacobi method is used to perform eigenvalue decomposition on the feature covariance matrix, extracting the principal direction vectors and eigenvalues ​​corresponding to all feature dimensions, and then sorting the principal direction vectors by dimensionality reduction according to the eigenvalues. Based on the set feature retention threshold, select the main direction vector that meets the conditions and construct the feature mapping matrix; Perform matrix multiplication on the sample data matrix and the feature mapping matrix to generate a dimension-compressed feature vector as the dominant component.

[0026] In this embodiment, the Jacobi method specifically includes: The two feature dimensions are determined by sequentially searching for the off-diagonal elements with the largest absolute values ​​in the feature covariance matrix. A rotation factor is constructed based on two feature dimensions, and an orthogonal transformation in the two-dimensional plane is performed to weaken the corresponding feature covariance terms in the matrix. The feature covariance terms represent the elements in the feature covariance matrix located outside the main diagonal. Repeat the rotation factor update and matrix adjustment to gradually reduce the absolute value of all off-diagonal elements; After all off-diagonal elements are less than a set threshold, the elements on the main diagonal are extracted as feature values, and the cumulative transformation result of the rotation factor in each round is used as the corresponding main direction vector.

[0027] In this embodiment, S3 specifically includes: S31. For each time step of the state vector sequence, construct an encoding vector representing the time order, and concatenate the encoding vector with the corresponding state vector according to the feature dimension to generate an enhanced representation sequence. S32. Input the augmented representation sequence into the edge decision network, and construct multiple attention channels in the edge decision network. Each channel performs an affine transformation operation on the augmented representation sequence to obtain the corresponding query vector, key vector and value vector. S33. Perform element-wise dot product operation on the query vector and key vector of all time steps in each attention channel to generate the correlation score matrix between time steps. Perform numerical scaling on the correlation score matrix according to the set scaling factor, and calculate the attention weight through the Softmax function. S34. In each attention channel, the value vector is weighted and summed according to the attention weight to obtain the attention representation corresponding to the channel. S35. Concatenate the attention representations extracted from all attention channels according to the feature dimension, and perform feature cross operation on the concatenation result to generate a co-state sequence.

[0028] In this embodiment, the feature crossing operation specifically includes: Construct a cross structure containing multiple cross layers in the edge nodes, and use the splicing result as the input of the first layer of the cross structure; In each cross layer, the input of that layer is set as the output of the previous layer, and the following processing is performed: Perform an element-wise product operation on the input and concatenation result of this layer to obtain the cross term; The cross terms are linearly weighted using trainable weight vectors and then added to a trainable bias vector to obtain the cross output vector of the current layer, which serves as the input to the next layer. The weight vectors represent the parameter vectors used to construct feature interaction relationships in the cross structure. The initial values ​​of the weight vectors are set through random initialization and iteratively updated in the edge nodes through the parameter optimization process of the edge decision network. The trainable bias vectors are the parameter terms corresponding to each layer in the cross structure and are iteratively updated in the edge nodes through the parameter optimization process of the edge decision network to correct the numerical deviation of the cross features. After all cross layers have been executed, the output of the last layer is used as the cooperative state sequence.

[0029] In this embodiment, S4 specifically includes: S41. Perform a first-order difference operation on the cooperative state sequence to generate a difference result sequence in chronological order; S42. The cumulative sum control chart method is used to detect the behavioral change points of the difference result sequence. The difference result sequence is processed by time step, and the cumulative processing results are compared with the set control limits. The time steps that exceed the control limits are identified and extracted as behavioral change points. S43. Arrange the behavior change points in chronological order to generate an event sequence, retrieve the access control parameters corresponding to the time step in the time series data set, and align them with the event sequence in time to construct a state mapping sequence.

[0030] In this embodiment, the cumulative sum control chart method specifically includes: Perform cumulative summation operation based on the difference result sequence, and construct a cumulative statistical sequence according to time steps; The mean and standard deviation of the difference results are calculated at multiple time steps in the difference result sequence. The mean is set as the baseline. The upper and lower control limits are constructed by adding or subtracting the standard deviation of the mean by a preset multiple, forming the control limit interval. The statistical values ​​at each time step in the cumulative statistics sequence are compared with the control limit interval, the time step that first deviates from the control limit interval is marked, and the time step is organized into a set of candidate change points. Construct a set of feature vectors containing the difference results for each candidate change point at the corresponding time step, the cumulative statistics for that time step and adjacent time steps, and call the isolated forest algorithm to calculate the anomaly score for each feature vector. Time steps with abnormal scores exceeding a set threshold are selected as behavioral change points.

[0031] In this embodiment, the calculation process of the isolated forest algorithm specifically includes: Each feature vector in the feature vector set is used as a sample node, and multiple random subsets of equal size are constructed by random sampling without replacement. For each random subset, a corresponding random tree structure is constructed. Based on the value range of the random subset in each dimension of features, the partition dimension and partition threshold are randomly selected. The sample is divided into left and right child nodes according to the selection result, and the partition operation is recursively executed. In each recursive partition, when the number of samples in a node is less than a set threshold or the preset maximum tree depth is reached, the partitioning stops and the node is retained as a leaf node, thus completing the construction of the random tree structure. The path depth of each sample node in all random tree structures is statistically analyzed, and the average path length of that sample node in all random trees is calculated. An anomaly scoring function is constructed based on the average path length of the sample nodes and the total number of samples, and anomaly scores are calculated for all sample nodes.

[0032] In this embodiment, S5 specifically includes: S51. Divide the state mapping sequence into multiple regional state sequences according to the spatial distribution of edge nodes, and organize the state sequence of each region in chronological order. S52. Perform discrete wavelet transform on the state sequence of each region to generate a multi-scale wavelet feature sequence; S53. Calculate the mean and variance in the time dimension for the wavelet feature sequences respectively, and combine the mean and variance according to the feature dimension to form the regional index of the corresponding region. S54. Arrange the regional indicators of each region in a predetermined order to construct a set of regional indicators.

[0033] In this embodiment, the discrete wavelet transform specifically includes: In the edge nodes, a preset wavelet filter bank is invoked to perform high-pass filtering on the state sequence of each region in chronological order to extract detailed feature sequences. High-pass filtering is recursively performed on the detail feature sequence to obtain detail feature sequences at multiple scales. The detailed feature sequences at each scale are spliced ​​together by time step, and the splicing results are standardized in the time dimension to form a multi-scale wavelet feature sequence.

[0034] In this embodiment, S6 specifically includes: S61. Input the set of regional indicators into the edge decision network, and organize each indicator vector into a strategy vector sequence according to the region. S62. Perform multi-layer perceptual transformation operation on each policy vector. In the first layer, perform affine transformation and ReLU function activation operation on the policy vector in sequence and output the result. S63. Starting from the second layer, affine transformation and ReLU function activation operations are performed sequentially based on the output of the previous layer, and the activation result is residually connected with the output of the previous layer as the current output. The output of the last layer is used as the policy feature representation. S64. Construct a policy probability distribution based on policy feature representation, construct a policy loss function based on the deviation between the policy distribution and the target action, and use a policy gradient algorithm to determine the policy update direction. The target action represents the expected action instruction pre-set in the cross-domain data collaboration process. S65. The policy gradient is weighted according to the action selection probability automatically allocated by the edge decision network, and the management instructions corresponding to each region are generated based on the updated policy probability distribution to form a management instruction sequence.

[0035] Example 1: To verify the feasibility of this invention in practice, it was applied to a multi-edge node collaborative management scenario in an industrial edge computing environment. This scenario consists of several heterogeneous edge devices distributed across different network domains, used for cross-node joint monitoring and decision control of device operating status, business data changes, and access permission policies. This scenario presents typical cross-domain data management challenges, such as diverse data sources, limited inter-node communication, and inconsistent management policies. Conventional centralized processing methods are insufficient to achieve efficient response and local optimization.

[0036] In this scenario, each edge node collects local raw data, such as device operating temperature, voltage fluctuations, task load, bandwidth usage, and number of connection sessions. This data, combined with node operating status and access permission policies, forms a time-step granular data input. The method proposed in this invention performs a one-dimensional convolution operation at each edge node to extract local features. Simultaneously, node type labels are one-hot encoded and fused to generate a state vector sequence. A multi-head attention mechanism is constructed between nodes through an edge decision network to capture temporal dependencies and identify key behavioral change points.

[0037] During behavioral analysis, cumulative sum control charts are used to detect behavioral mutations in edge data, and the Isolation Forest algorithm is invoked to filter out anomalous change points, forming an event sequence. Subsequently, discrete wavelet transform is used to extract local fluctuations from the regional state sequences, forming a set of regional indicators. A policy gradient optimization algorithm is then used to evaluate the policies for each regional indicator, ultimately outputting specific management instructions, such as dynamic bandwidth scheduling, task migration, and node cooling policy deployment. After the system executes the management instructions, the feedback results are used to optimize the edge decision network parameters, forming a stable edge collaborative closed loop.

[0038] After seven consecutive days of simulated operation testing, the system has achieved significant improvements in data behavior response, anomaly identification, and policy hit rate.

[0039] The table below shows the comparison results of specific test data.

[0040] Table 1. Comparison of the effects of edge collaborative management methods As shown in Table 1, the method of this invention outperforms traditional methods in terms of response speed, policy accuracy, node collaboration efficiency, and anomaly detection performance. Specifically, the policy hit rate is improved by over 22%, and the anomaly response latency is reduced by over 50%, demonstrating good stability and practicality in complex cross-domain environments. Furthermore, this invention achieves faster convergence during the policy training phase, reducing the average number of convergence rounds to 38, significantly shortening deployment initialization time. Experiments also revealed that after using the method of this invention, the resource utilization rate of edge nodes remained stable at over 75%, indicating more refined and efficient management and scheduling.

[0041] The above embodiments demonstrate that the cross-domain data collaborative management method based on edge computing proposed in this invention can effectively improve the level of multi-node collaborative intelligence, achieve agile identification and strategy response to cross-domain data changes, and has good engineering application prospects.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A cross-domain data collaborative management method based on edge computing, characterized in that, Includes the following steps: S1. Obtain the raw data, node status and access control parameters of each edge node in the cross-domain environment, and assemble them into a time-series data set in chronological order; S2. At the edge nodes, perform one-dimensional convolution operation on the time series data of each time step to extract features, perform one-hot encoding on the preset labels, and connect the encoding results with the convolution results according to the feature dimension to generate a state vector sequence. S3. Input the state vector sequence into the edge decision network, calculate the correlation weight between each time step through the multi-head attention mechanism, and perform weighted aggregation of the state vector sequence based on the weight to generate a collaborative state sequence. S4. Perform a first-order difference operation on the cooperative state sequence, use the cumulative sum control chart method to detect behavioral change points, and align the detection results with the access control parameters in time to construct a state mapping sequence. S5. Divide the state mapping sequence into multiple regional state sequences according to the spatial distribution of edge nodes, use discrete wavelet transform to generate wavelet feature sequences, and calculate the mean and variance of the wavelet feature sequences to construct a set of regional indicators. S6. Input the set of regional indicators into the edge decision network, use the policy gradient algorithm to evaluate the policy and determine the action, and generate a sequence of management instructions. S7. Execute the management instruction sequence at the edge node, and perform gradient pruning and updating of the edge decision network parameters based on the execution results.

2. The cross-domain data collaborative management method based on edge computing according to claim 1, characterized in that, The raw data includes sensor sampling data, network communication data, log information, task execution status and user request records. The node status represents the running parameters of the edge node at a specific time step. The access control parameters represent the configuration content that controls the data sharing permissions between edge nodes. The preset tags represent discrete identification information used to distinguish edge node categories, functional roles or domains. The behavior change points represent the mutation locations that occur during state evolution. The edge decision network represents the neural network structure deployed on the edge nodes for performing strategy evaluation and action determination.

3. The cross-domain data collaborative management method based on edge computing according to claim 1, characterized in that, S2 specifically includes: S21. Divide the time series data set into multiple data segments in the edge nodes according to time steps, and arrange each data segment into vectorized arrangement according to a preset dimension to form structural feature data. S22. Local temporal features of structural feature data are extracted using a multi-scale one-dimensional convolution method. During the convolution process, parallel convolution processing is carried out according to different set convolution kernel sizes and fixed sliding strides. Batch normalization processing is performed on the convolution results to construct convolution feature representation. S23. Perform one-hot encoding on the preset labels associated with the edge nodes, and convert the encoding results into a compact vector of fixed length through embedding mapping to form a label feature representation; S24. The convolutional feature representation and label feature representation at each time step are structurally concatenated according to the feature dimension, and the dominant component of the concatenation result is extracted by principal component analysis to form a fused feature vector. S25. Arrange the fused feature vectors in chronological order to construct a state vector sequence.

4. The cross-domain data collaborative management method based on edge computing according to claim 3, characterized in that, The principal component analysis specifically includes: Arrange the splicing results in chronological order to construct a sample data matrix. Calculate the covariance relationship between all feature dimensions based on this matrix and output the feature covariance matrix. The Jacobi method is used to perform eigenvalue decomposition on the feature covariance matrix, extracting the principal direction vectors and eigenvalues ​​corresponding to all feature dimensions, and then sorting the principal direction vectors by dimensionality reduction according to the eigenvalues. Based on the set feature retention threshold, select the main direction vector that meets the conditions and construct the feature mapping matrix; Perform matrix multiplication on the sample data matrix and the feature mapping matrix to generate a dimension-compressed feature vector as the dominant component.

5. The cross-domain data collaborative management method based on edge computing according to claim 1, characterized in that, S3 specifically includes: S31. For each time step of the state vector sequence, construct an encoding vector representing the time order, and concatenate the encoding vector with the corresponding state vector according to the feature dimension to generate an enhanced representation sequence. S32. Input the augmented representation sequence into the edge decision network, and construct multiple attention channels in the edge decision network. Each channel performs an affine transformation operation on the augmented representation sequence to obtain the corresponding query vector, key vector and value vector. S33. Perform element-wise dot product operation on the query vector and key vector of all time steps in each attention channel to generate the correlation score matrix between time steps. Perform numerical scaling on the correlation score matrix according to the set scaling factor, and calculate the attention weight through the Softmax function. S34. In each attention channel, the value vector is weighted and summed according to the attention weight to obtain the attention representation corresponding to the channel. S35. Concatenate the attention representations extracted from all attention channels according to the feature dimension, and perform feature cross operation on the concatenation result to generate a co-state sequence.

6. The cross-domain data collaborative management method based on edge computing according to claim 1, characterized in that, S4 specifically includes: S41. Perform a first-order difference operation on the cooperative state sequence to generate a difference result sequence in chronological order; S42. The cumulative sum control chart method is used to detect the behavioral change points of the difference result sequence. The difference result sequence is processed by time step, and the cumulative processing results are compared with the set control limits. The time steps that exceed the control limits are identified and extracted as behavioral change points. S43. Arrange the behavior change points in chronological order to generate an event sequence, retrieve the access control parameters corresponding to the time step in the time series data set, and align them with the event sequence in time to construct a state mapping sequence.

7. The cross-domain data collaborative management method based on edge computing according to claim 6, characterized in that, The cumulative sum control chart method specifically includes: Perform cumulative summation operation based on the difference result sequence, and construct a cumulative statistical sequence according to time steps; The mean and standard deviation of the difference results are calculated at multiple time steps in the difference result sequence. The mean is set as the baseline. The upper and lower control limits are constructed by adding or subtracting the standard deviation of the mean by a preset multiple, forming the control limit interval. The statistical values ​​at each time step in the cumulative statistics sequence are compared with the control limit interval, the time step that first deviates from the control limit interval is marked, and the time step is organized into a set of candidate change points. Construct a set of feature vectors containing the difference results for each candidate change point at the corresponding time step, the cumulative statistics for that time step and adjacent time steps, and call the isolated forest algorithm to calculate the anomaly score for each feature vector. Time steps with abnormal scores exceeding a set threshold are selected as behavioral change points.

8. The cross-domain data collaborative management method based on edge computing according to claim 7, characterized in that, The calculation process of the isolated forest algorithm specifically includes: Each feature vector in the feature vector set is used as a sample node, and multiple random subsets of equal size are constructed by random sampling without replacement. For each random subset, a corresponding random tree structure is constructed. Based on the value range of the random subset in each dimension of features, the partition dimension and partition threshold are randomly selected. The sample is divided into left and right child nodes according to the selection result, and the partition operation is recursively executed. In each recursive partition, when the number of samples in a node is less than a set threshold or the preset maximum tree depth is reached, the partitioning stops and the node is retained as a leaf node, thus completing the construction of the random tree structure. The path depth of each sample node in all random tree structures is statistically analyzed, and the average path length of that sample node in all random trees is calculated. An anomaly scoring function is constructed based on the average path length of the sample nodes and the total number of samples, and anomaly scores are calculated for all sample nodes.

9. The cross-domain data collaborative management method based on edge computing according to claim 1, characterized in that, S5 specifically includes: S51. Divide the state mapping sequence into multiple regional state sequences according to the spatial distribution of edge nodes, and organize the state sequence of each region in chronological order. S52. Perform discrete wavelet transform on the state sequence of each region to generate a multi-scale wavelet feature sequence; S53. Calculate the mean and variance in the time dimension for the wavelet feature sequences respectively, and combine the mean and variance according to the feature dimension to form the regional index of the corresponding region. S54. Arrange the regional indicators of each region in a predetermined order to construct a set of regional indicators.

10. The cross-domain data collaborative management method based on edge computing according to claim 1, characterized in that, S6 specifically includes: S61. Input the set of regional indicators into the edge decision network, and organize each indicator vector into a strategy vector sequence according to the region. S62. Perform multi-layer perceptual transformation operation on each policy vector. In the first layer, perform affine transformation and ReLU function activation operation on the policy vector in sequence and output the result. S63. Starting from the second layer, affine transformation and ReLU function activation operations are performed sequentially based on the output of the previous layer, and the activation result is residually connected with the output of the previous layer as the current output. The output of the last layer is used as the policy feature representation. S64. Construct a policy probability distribution based on policy feature representation, construct a policy loss function based on the deviation between the policy distribution and the target action, and use a policy gradient algorithm to determine the policy update direction. The target action represents the expected action instruction pre-set in the cross-domain data collaboration process. S65. The policy gradient is weighted according to the action selection probability automatically allocated by the edge decision network, and the management instructions corresponding to each region are generated based on the updated policy probability distribution to form a management instruction sequence.

Images