Collaborative Control Method and System for Intelligent Converged Terminals Based on Edge Computing

Abstract

This invention discloses a collaborative control method and system for intelligent converged terminals based on edge computing, comprising: acquiring raw terminal data, performing preprocessing and generating a standard terminal dataset; constructing dual-track executable digital shadows for each terminal based on the standard terminal dataset; predicting state evolution and generating risk parameters using an improved NGBoost model; constructing a reconfigurable collaborative temporal cube and generating candidate collaborative control sections; searching for paths and generating differential control tokens based on the candidate collaborative control sections; issuing differential control tokens, performing control corrections, and feeding back to update the model. This invention achieves low-latency, adaptive collaborative control of multiple intelligent converged terminals in edge computing scenarios through dual-track executable digital shadows, improved NGBoost probabilistic prediction, a reconfigurable collaborative temporal cube, and differential control tokens.

Description

Technical Field

[0001] This invention relates to the field of edge collaborative state data processing technology, and in particular to a collaborative control method and system for intelligent fusion terminals based on edge computing. Background Technology

[0002] With the development of edge computing and smart terminal technologies, various types of intelligent converged terminals, such as camera terminals, sensor terminals, edge gateways, and control execution terminals, are widely deployed in scenarios such as park management, industrial sites, equipment monitoring, and intelligent control. Existing collaborative control methods typically involve edge nodes collecting terminal resource status and allocating tasks or issuing control commands to multiple terminals based on current data such as CPU utilization, memory utilization, network bandwidth, and online status.

[0003] While existing technologies can reduce latency caused by centralized cloud control to some extent, their control decisions largely rely on static resource states or fixed scheduling rules, lacking prediction of the state evolution after the terminal executes control actions. In scenarios where multiple heterogeneous intelligent fusion terminals participate in the same control task, existing methods struggle to assess terminal load fluctuations, link risks, control conflicts, and the likelihood of takeover triggers before control execution, leading to inaccurate collaborative path selection.

[0004] Existing methods typically employ a process of issuing control commands, executing them at the terminal, and rescheduling after an anomaly. The terminal side lacks the ability to autonomously correct for deviations between the predicted state and the actual execution state. When terminal response times out, link jitter occurs, or coordination conflicts arise, edge nodes need to recalculate and reissue control strategies, which can easily lead to takeover delays, redundant scheduling, and decreased control continuity, making it difficult to meet the stability requirements of low-latency, multi-terminal collaborative control scenarios.

[0005] Therefore, how to provide a collaborative control method and system for intelligent converged terminals based on edge computing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a collaborative control method and system for intelligent fusion terminals based on edge computing. This invention fully utilizes dual-track executable digital shadows, improved NGBoost probabilistic prediction models, reconfigurable collaborative timing cubes, and differential control token technology to predict and dynamically correct the state evolution, collaborative link risks, and control deviations of multiple heterogeneous intelligent fusion terminals. It has the advantages of high collaborative control stability, fast takeover response, less redundant scheduling, and strong real-time edge control.

[0007] The collaborative control method for an edge computing-based intelligent fusion terminal according to an embodiment of the present invention includes:

[0008] Obtain raw terminal data from multiple intelligent converged terminals, preprocess the raw terminal data, and obtain a standard terminal dataset;

[0009] Based on the standard terminal dataset, a dual-track executable digital shadow is constructed for each intelligent fusion terminal. The dual-track executable digital shadow includes a main shadow body, a counterfactual shadow body, and a shadow coupling gating unit.

[0010] Based on each dual-track executable digital shadow, the improved NGBoost model is used to predict the state evolution of the master shadow and counterfactual shadow under different control actions, and obtain the state distribution parameters, shadow divergence value, link risk value, control conflict value and takeover trigger value.

[0011] A reconfigurable collaborative time-series cube is constructed based on the control task to be executed. The state distribution parameters, shadow divergence values, link risk values, control conflict values, and takeover trigger values ​​are written into the corresponding cube cells. Based on the shadow divergence values, the reconfigurable collaborative time-series cube is expanded along the time slice dimension, pruned along the terminal dimension, and switched along the collaborative layer dimension to generate candidate collaborative control sections.

[0012] Based on each candidate cooperative control section, a cooperative path search is performed to generate a target cooperative control path, and a differential control token is generated based on the target cooperative control path.

[0013] Differential control tokens are sent to the corresponding intelligent fusion terminals. Each intelligent fusion terminal executes the corresponding control action. When the deviation between the actual execution state and the predicted state exceeds the allowable deviation range, a deviation correction strategy is executed. The corresponding dual-track executable digital shadow and reconfigurable collaborative timing cube are updated based on the execution data fed back by each intelligent fusion terminal.

[0014] Optionally, the original terminal data includes terminal identification data, terminal type data, terminal computing resource data, terminal storage resource data, terminal communication resource data, terminal power status data, terminal control interface data, terminal historical response latency data, terminal historical control success rate data, terminal abnormal interruption data, and inter-terminal collaboration relationship data.

[0015] Optionally, the preprocessing of the raw terminal data to obtain a standard terminal dataset includes:

[0016] The original terminal data is subjected to timestamp unification, sampling frequency alignment, duplicate data deletion, outlier removal, missing value imputation, unit normalization, terminal identifier mapping, and cooperative relationship encoding to obtain a standard terminal dataset.

[0017] Optionally, the construction of a dual-track executable digital shadow for each intelligent fusion terminal based on a standard terminal dataset includes:

[0018] Within each edge node, traverse the standard terminal dataset one by one according to the terminal identifier;

[0019] For each intelligent fusion terminal, the acquisition capability, computing capability and control interface are encapsulated as capability shadows, the current resource occupancy, communication status and power status are encapsulated as status shadows, the historical response latency, historical control success rate and abnormal interruption records are encapsulated as behavior shadows, and the dependency relationship, conflict relationship and takeover relationship are encapsulated as relationship shadows. The four types of shadows are combined to generate the main shadow body.

[0020] Three types of preset control actions are set for each master shadow body: candidate control action, takeover action, and rollback action. By calling the built-in state transition rules, each type of preset control action is quickly deduced to obtain the corresponding predicted resource changes, predicted communication changes, and predicted control results. The deduction results are associated with the corresponding actions to generate a counterfactual shadow body.

[0021] A shadow coupling gating unit is configured for each main shadow entity and its corresponding counterfactual shadow entity. The shadow coupling gating unit dynamically calculates the synchronization update triggering conditions of the main shadow entity and the counterfactual shadow entity based on the real-time state change amplitude of the terminal, the historical behavior fluctuation amplitude, and the frequency of changes in the collaborative relationship. When the triggering conditions are met, the shadow fields of the two shadows are refreshed synchronously.

[0022] The master shadow body, counterfactual shadow body, and shadow coupling gating unit are bound together. After writing a unified identifier, version number, and update timestamp, a dual-track executable digital shadow of the intelligent fusion terminal is generated and stored in the digital shadow library of the edge node.

[0023] Optionally, obtaining the state distribution parameters, shadow divergence value, link risk value, control conflict value, and takeover trigger value includes:

[0024] For each dual-track executable digital shadow, read its acquisition capability, computing capability, control interface, current CPU utilization, current memory utilization, current network bandwidth usage, current remaining power, average response latency of the last 100 times, control success rate of the last 100 times, count of the last 100 abnormal interruptions, dependency flag, conflict flag, and takeover flag, and concatenate them into a terminal feature vector according to the preset field order.

[0025] After one-hot encoding of candidate control actions, takeover actions and rollback actions, they are input into the action embedding layer to obtain a sixteen-dimensional action embedding vector. The action embedding vector is then concatenated bitwise with the terminal feature vector that has been normalized to zero mean to form the model input vector.

[0026] The model input vector is sequentially fed into the feature normalization layer, action embedding layer, gradient boosting forest layer, natural gradient update layer, and multi-task distribution output layer of the improved NGBoost model. The gradient boosting forest layer is set with 400 regression trees of depth 3 and iterates with a learning rate of 0.05. The natural gradient update layer updates the distribution parameters at the end of each iteration. The multi-task distribution output layer is a branch output module connected in series after the original NGBoost output end, which is used to simultaneously output the mean and variance of CPU utilization distribution, memory utilization distribution, network bandwidth usage distribution, response latency distribution, and energy consumption distribution.

[0027] The improved NGBoost model is input into the main shadow and the corresponding counterfactual shadow respectively to obtain two sets of state distribution parameters. The state distribution parameters include the mean and variance of CPU utilization, the mean and variance of memory utilization, the mean and variance of network bandwidth usage, the mean and variance of response latency, and the mean and variance of energy consumption.

[0028] The difference between the mean values ​​of the state distributions corresponding to the primary shadow body and the counterfactual shadow body is used as the shadow divergence value. The link risk value is calculated by comparing the mean response delay corresponding to the primary shadow body with the maximum allowable delay threshold and the network bandwidth usage variance. The control conflict value is calculated by using the conflict marker and the response delay variance. The takeover trigger value is calculated by using the takeover marker and the CPU utilization variance.

[0029] Write the state distribution parameters, shadow divergence values, link risk values, control conflict values, and takeover trigger values ​​into the corresponding dual-track executable digital shadow.

[0030] Optionally, generating candidate cooperative control sections includes:

[0031] The control task to be executed is analyzed to obtain the task identifier, task control target, terminal type required by the task, maximum allowable response time, target control action, takeover constraint and rollback constraint. Based on the terminal type required by the task, the set of participating terminals is selected from the dual-track executable digital shadows corresponding to each intelligent fusion terminal.

[0032] The terminal dimension is formed by the terminal identifier corresponding to the participating terminal set, the time slice dimension is formed by the continuous time slices obtained according to the preset time granularity, and the collaboration layer dimension is formed by the main control layer, the co-control layer, the verification layer, the takeover layer and the rollback layer. A reconfigurable collaborative time sequence cube is generated based on the terminal dimension, the time slice dimension and the collaboration layer dimension.

[0033] For each participating terminal, the state distribution parameters, shadow divergence values, link risk values, control conflict values, and takeover trigger values ​​obtained by the participating terminal under the target control action, takeover action, and rollback action are written into the corresponding cube cell.

[0034] The reconfigurable collaborative timing cube is expanded according to the shadow divergence value. When the shadow divergence value of the participating terminal under the current control action exceeds the preset divergence threshold, a takeover judgment time slice or a rollback judgment time slice is added after the current time slice. The newly added time slice is associated with the participating terminal and the corresponding collaborative layer.

[0035] Based on the link risk value, control conflict value, and takeover trigger value, the reconfigurable collaborative time sequence cube is subjected to terminal dimension pruning and collaborative layer dimension switching. Terminal units whose link risk value exceeds the preset risk threshold and whose takeover trigger value does not reach the preset takeover threshold are deleted. Terminal units whose control conflict value exceeds the preset conflict threshold are switched from the current collaborative layer to the rollback layer or suppression state.

[0036] From the reconfigurable collaborative time sequence cube after time slice dimension expansion, terminal dimension pruning, and collaborative layer dimension switching, candidate collaborative control sections that satisfy task control objectives, maximum allowable response latency, takeover constraints, and rollback constraints are extracted.

[0037] Optionally, generating differential control tokens based on the target cooperative control path includes:

[0038] For each candidate cooperative control section, all cube cells are read sequentially, and the terminal identifier, time slice number, cooperative layer type, control action type, predicted response delay, shadow divergence value, link risk value, control conflict value, and takeover trigger value corresponding to each cube cell are recorded.

[0039] According to the ascending order of time slice numbers and the priority order of the collaborative layer, the cube units in the same candidate collaborative control section are connected in series to form a candidate collaborative control path. When connecting in series, only the main control layer unit is allowed to be followed by the co-control layer unit, the co-control layer unit is allowed to be followed by the verification layer unit, and the verification layer unit is allowed to be followed by the main control layer unit or the rollback layer unit.

[0040] For each candidate cooperative control path, the path cost is calculated. The path cost is equal to the sum of the predicted response delay, shadow divergence value, link risk value and control conflict value of all cube cells in the path, multiplied by a preset weight. Candidate cooperative control paths with predicted response delay exceeding the maximum allowable response delay or containing takeover trigger values ​​that have not reached the preset takeover threshold are filtered out.

[0041] The path with the lowest path cost among the remaining candidate cooperative control paths is selected as the target cooperative control path. The order of issuance of differential control tokens is determined according to the time slice number and cooperative layer type of each cube unit in the target cooperative control path.

[0042] Differential control tokens are generated sequentially for each cube unit in the target collaborative control path. Each differential control token includes terminal identifier, collaborative layer type, control action, predicted state mean, allowable deviation range, deviation correction strategy, takeover trigger condition, rollback action and feedback rule. The generated differential control tokens are stored in the token distribution queue.

[0043] Optionally, each intelligent fusion terminal executes a corresponding control action, and when the deviation between the actual execution state and the predicted state exceeds the allowable deviation range, a deviation correction strategy is executed, including:

[0044] Edge nodes send differential control tokens to the corresponding smart fusion terminals according to the token distribution queue. After receiving the differential control token, the smart fusion terminal performs matching and verification of the terminal identifier, collaboration layer type and control action. After the matching and verification is passed, the control action is executed and real status data during the execution process is collected.

[0045] The intelligent fusion terminal calculates the deviation between the real state data and the predicted state mean in the differential control token to obtain the execution deviation value. The execution deviation value satisfies the following condition: the execution deviation value is equal to the sum of the CPU utilization deviation, memory utilization deviation, network bandwidth usage deviation, response latency deviation and energy consumption deviation multiplied by their respective state weights.

[0046] When the execution deviation value exceeds the allowable deviation range, the intelligent fusion terminal performs takeover action, rollback action, control action downgrade or control action suppression according to the deviation correction strategy in the differential control token, and generates feedback data containing the execution deviation value, deviation correction strategy type and correction execution result.

[0047] Edge nodes receive execution data from each intelligent fusion terminal, and update the master shadow body, counterfactual shadow body, and shadow coupling gating unit in the corresponding dual-track executable digital shadow based on the execution data. They also update the link risk value, control conflict value, takeover trigger value, and path weight of the corresponding cube unit in the reconfigurable collaborative temporal cube.

[0048] According to an embodiment of the present invention, a collaborative control system for an edge computing-based intelligent fusion terminal includes:

[0049] The data preprocessing module is used to acquire and preprocess raw terminal data from multiple smart converged terminals to generate a standard terminal dataset.

[0050] The dual-track shadow construction module is used to construct dual-track executable digital shadows for each intelligent fusion terminal based on a standard terminal dataset.

[0051] The probability prediction module is used to predict the state evolution under different control actions based on the dual-track executable digital shadow and the improved NGBoost model, so as to obtain the state distribution parameters and risk parameters.

[0052] The cube reconstruction module is used to construct and reconstruct a collaborative time-series cube based on the control task to be executed and the risk parameters, and generate candidate collaborative control sections.

[0053] The path token generation module is used to search for the target cooperative control path based on the candidate cooperative control cross-section and generate differential control tokens;

[0054] The execution feedback update module is used to issue differential control tokens to control the intelligent fusion terminal to perform control actions, and to update the dual-track executable digital shadow and collaborative timing cube based on the execution feedback.

[0055] The beneficial effects of this invention are:

[0056] This invention constructs a dual-track executable digital shadow and combines it with an improved NGBoost model to probabilistically predict the state evolution of intelligent fusion terminals under different control actions. It can obtain the terminal state distribution, shadow divergence value, link risk value, control conflict value, and takeover trigger value before the control command is issued, thereby improving the edge node's ability to predict the results of multi-terminal collaborative control.

[0057] This invention constructs a reconfigurable collaborative time-series cube and dynamically adjusts the time slice dimension, terminal dimension, and collaborative layer dimension based on prediction results. This enables the search for target collaborative control paths in multiple candidate collaborative control sections, reducing the probability of high-risk terminal combinations and conflicting control paths being selected, and improving the stability of collaborative control of multi-heterogeneous intelligent fusion terminals.

[0058] This invention generates differential control tokens, enabling intelligent fusion terminals to execute deviation correction strategies based on the deviation between the actual execution state and the predicted state when performing control actions. It also feeds back execution data to update the dual-track executable digital shadow and the reconfigurable collaborative timing cube, reducing takeover lag and repeated scheduling, and improving the real-time performance and continuity of collaborative control in edge computing scenarios. Attached Figure Description

[0059] 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:

[0060] Figure 1 This is a flowchart of the collaborative control method for an intelligent fusion terminal based on edge computing proposed in this invention;

[0061] Figure 2 This is a schematic diagram of the collaborative control system of the intelligent fusion terminal based on edge computing proposed in this invention. Detailed Implementation

[0062] 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.

[0063] refer to Figure 1 A collaborative control method for intelligent converged terminals based on edge computing includes:

[0064] Obtain raw terminal data from multiple intelligent converged terminals, preprocess the raw terminal data, and obtain a standard terminal dataset;

[0065] Based on the standard terminal dataset, a dual-track executable digital shadow is constructed for each intelligent fusion terminal. The dual-track executable digital shadow includes a main shadow body, a counterfactual shadow body, and a shadow coupling gating unit.

[0066] Based on each dual-track executable digital shadow, the improved NGBoost model is used to predict the state evolution of the master shadow and counterfactual shadow under different control actions, and obtain the state distribution parameters, shadow divergence value, link risk value, control conflict value and takeover trigger value.

[0067] A reconfigurable collaborative time-series cube is constructed based on the control task to be executed. The state distribution parameters, shadow divergence values, link risk values, control conflict values, and takeover trigger values ​​are written into the corresponding cube cells. Based on the shadow divergence values, the reconfigurable collaborative time-series cube is expanded along the time slice dimension, pruned along the terminal dimension, and switched along the collaborative layer dimension to generate candidate collaborative control sections.

[0068] Based on each candidate cooperative control section, a cooperative path search is performed to generate a target cooperative control path, and a differential control token is generated based on the target cooperative control path.

[0069] Differential control tokens are sent to the corresponding intelligent fusion terminals. Each intelligent fusion terminal executes the corresponding control action. When the deviation between the actual execution state and the predicted state exceeds the allowable deviation range, a deviation correction strategy is executed. The corresponding dual-track executable digital shadow and reconfigurable collaborative timing cube are updated based on the execution data fed back by each intelligent fusion terminal.

[0070] In this embodiment, the original terminal data includes terminal identification data, terminal type data, terminal computing resource data, terminal storage resource data, terminal communication resource data, terminal power status data, terminal control interface data, terminal historical response latency data, terminal historical control success rate data, terminal abnormal interruption data, and inter-terminal collaboration relationship data.

[0071] In this embodiment, the preprocessing of the raw terminal data to obtain a standard terminal dataset includes:

[0072] The original terminal data is subjected to timestamp unification, sampling frequency alignment, duplicate data deletion, outlier removal, missing value imputation, unit normalization, terminal identifier mapping, and cooperative relationship encoding to obtain a standard terminal dataset.

[0073] In this embodiment, the construction of a dual-track executable digital shadow for each intelligent fusion terminal based on a standard terminal dataset includes:

[0074] Within each edge node, traverse the standard terminal dataset one by one according to the terminal identifier;

[0075] For each intelligent fusion terminal, the acquisition capability, computing capability, and control interface are encapsulated as capability shadows; the current resource occupancy, communication status, and power status are encapsulated as status shadows; historical response latency, historical control success rate, and abnormal interruption records are encapsulated as behavior shadows; and dependency relationships, conflict relationships, and takeover relationships are encapsulated as relationship shadows. These four types of shadows are combined to generate the main shadow body. Specifically, the combination of these four types of shadows to generate the main shadow body is as follows:

[0076] Read the acquisition capability field, computing capability field, and control interface field from the capability shadow, and write them into the capability segment in the order of acquisition capability-computing capability-control interface; read the CPU utilization, memory utilization, network bandwidth usage, and remaining power from the status shadow, and write them into the status segment in the order of CPU-memory-bandwidth-power; read the average response latency, average control success rate, and abnormal interruption count from the behavior shadow, and write them into the behavior segment in the order of latency-success rate-abnormal count; read the dependency flag, conflict flag, and takeover flag from the relation shadow, and write them into the relation segment in the order of dependency-conflict-takeover; concatenate the capability segment, status segment, behavior segment, and relation segment into a fixed field sequence, add the terminal identifier at the beginning of the sequence, and append the current timestamp and shadow version number to the end of the sequence to generate a shadow record of fixed length, and write the shadow record into the shadow cache table of the edge node as the main shadow body of the terminal;

[0077] For each primary shadow entity, three types of preset control actions are assigned: candidate control actions, takeover actions, and rollback actions. Each preset control action is rapidly simulated by invoking built-in state transition rules, yielding corresponding predicted resource changes, predicted communication changes, and predicted control results. The simulation results are then correlated with the corresponding actions to generate a counterfactual shadow entity. Specifically, the rapid simulation of each preset control action using built-in state transition rules is as follows:

[0078] Read the current CPU utilization, current memory utilization, current network bandwidth usage, current remaining power, average response latency, average control success rate, and abnormal interruption count from the master shadow entity; read the action resource consumption table corresponding to the preset control action to obtain the CPU increment, memory increment, bandwidth increment, power consumption, and basic execution latency of the action; add the current CPU utilization to the CPU increment to obtain the predicted CPU utilization; add the current memory utilization to the memory increment to obtain the predicted memory utilization; add the current network bandwidth usage to the bandwidth increment to obtain the predicted network bandwidth usage; subtract the power consumption from the current remaining power to obtain the predicted remaining power; add the average response latency to the basic execution latency to obtain the predicted response latency; determine the predicted control result field based on the average control success rate and abnormal interruption count.

[0079] The counterfactual shadow entity is generated by associating the deduction results with the corresponding actions, specifically as follows:

[0080] Action records are created for candidate control actions, takeover actions, and rollback actions. Each action record contains the following fields: action type, action identifier, predicted CPU utilization, predicted memory utilization, predicted network bandwidth usage, predicted remaining battery power, predicted response latency, and predicted control result. The three action records are then concatenated in the order of candidate control action record – takeover action record – rollback action record. The terminal identifier of the corresponding master shadow body is written at the beginning of the concatenated result, and the generation timestamp and counterfactual shadow version number are written at the end, generating the counterfactual shadow body corresponding to this intelligent fused terminal.

[0081] A shadow coupling gating unit is configured for each primary shadow entity and its corresponding counterfactual shadow entity. This gating unit dynamically calculates the synchronization update trigger conditions for the primary and counterfactual shadow entities based on the terminal's real-time state change amplitude, historical behavior fluctuation amplitude, and frequency of collaborative relationship changes. When the trigger conditions are met, the shadow fields of both shadows are refreshed synchronously. Specifically, the dynamic calculation of the synchronization update trigger conditions for the primary and counterfactual shadow entities is as follows:

[0082] Read the CPU utilization, memory utilization, network bandwidth usage, and power reserve of the main shadow body in the latest 10 sampling periods, and calculate the percentage difference between the current period and the average of the 10 periods as the state change amplitude; read the response latency and control success rate in the latest 50 control logs of the main shadow body, and calculate the proportion of the standard deviation of this batch of data to the mean as the historical behavior fluctuation amplitude; count the number of additions and deletions of dependency markers, conflict markers, and takeover markers in the relation shadow in the last 60 seconds, and calculate the proportion of the total number of relation shadow fields as the frequency of collaborative relationship changes; multiply the state change amplitude, historical behavior fluctuation amplitude, and collaborative relationship change frequency by weight coefficients of 0.5, 0.3, and 0.2 respectively, and sum them to obtain the synchronization update score. When the synchronization update score is greater than 0.4 or any one of the three indicators exceeds 0.25, the trigger condition is considered met, and the corresponding fields of the main shadow body and the counterfactual shadow body are immediately refreshed and a new timestamp is written.

[0083] The master shadow body, counterfactual shadow body, and shadow coupling gating unit are bound together. After writing a unified identifier, version number, and update timestamp, a dual-track executable digital shadow of the intelligent fusion terminal is generated and stored in the digital shadow library of the edge node.

[0084] In this embodiment, obtaining the state distribution parameters, shadow divergence value, link risk value, control conflict value, and takeover trigger value includes:

[0085] For each dual-track executable digital shadow, the following data is collected: acquisition capability, computing power, control interface, current CPU utilization, current memory utilization, current network bandwidth usage, current remaining battery power, average response latency of the last 100 responses, control success rate of the last 100 responses, count of the last 100 abnormal interruptions, dependency flags, conflict flags, and takeover flags. These data are then concatenated into a terminal feature vector according to a preset field order. Specifically, the pre-set field order for concatenating the terminal feature vector is as follows:

[0086] First, a 4-byte terminal identifier is entered at the beginning of the vector. Then, the acquisition capability code, computing capability code, and control interface code are written in sequence, each occupying 2 bytes. Next, the current CPU utilization, current memory utilization, and current network bandwidth usage are written, each represented by a 4-byte floating-point number. Then, the current remaining battery power is written, using a 2-byte unsigned integer to represent the percentage value. Next, the average response latency of the last 100 times and the control success rate of the last 100 times are written, each occupying 4 bytes of floating-point number. Then, the count of the last 100 abnormal interruptions is written, stored as a 2-byte unsigned integer. Finally, the dependency flag, conflict flag, and takeover flag are written, each encoded with a 1-byte boolean bit. All fields are concatenated in the above order without delimiters to form a fixed-length terminal feature vector, with a 4-byte timestamp appended to the end to record the sampling time.

[0087] After one-hot encoding of candidate control actions, takeover actions and rollback actions, they are input into the action embedding layer to obtain a sixteen-dimensional action embedding vector. The action embedding vector is then concatenated bitwise with the terminal feature vector that has been normalized to zero mean to form the model input vector.

[0088] The model input vector is sequentially fed into the feature normalization layer, action embedding layer, gradient boosting forest layer, natural gradient update layer, and multi-task distribution output layer of the improved NGBoost model. The gradient boosting forest layer is set with 400 regression trees of depth 3 and iterates with a learning rate of 0.05. The natural gradient update layer updates the distribution parameters at the end of each iteration. The multi-task distribution output layer is a branch output module connected in series after the original NGBoost output end, which is used to simultaneously output the mean and variance of CPU utilization distribution, memory utilization distribution, network bandwidth usage distribution, response latency distribution, and energy consumption distribution.

[0089] The feature normalization layer includes:

[0090] Field validator: Detects whether the input vector is arranged in the order of terminal identifier, capability code, status value, behavior statistics, relationship tag, and action embedding. If it does not match, an alarm is triggered and the sample is discarded.

[0091] Miss filler: Performs historical median filling for CPU utilization and memory utilization, and zero filling for Boolean flags;

[0092] Linear standardizer: Logarithmic fields are calculated as (original value - field mean) / field standard deviation; the mean and standard deviation are derived from a rolling statistical buffer of the past 10,000 samples;

[0093] Normalized output buffer: The normalized result is written in 32-bit floating-point format to a 128-bit aligned circular buffer for sequential reading by the next layer;

[0094] The action embedding layer includes:

[0095] One-hot encoding unit: maps the three types of actions—candidate, takeover, and rollback—to 3×8-bit Boolean vectors respectively;

[0096] Sparse compressor: Uses bitmap indexes to remove fields with all zeros to prevent invalid writes;

[0097] Dense projector: Right-multiply the sparse Boolean vector with a 256×16 trainable matrix to obtain a 16-dimensional action embedding vector; the matrix is ​​initially initialized using Xavier;

[0098] Concatenation register: The normalized numerical segment, relational tag segment, and action embedding segment are concatenated in sequence to form the final model input, with a fixed length of 96 dimensions;

[0099] Gradient boosting forest layers include:

[0100] Tree set buffer pool: pre-allocates an array of 400 regression tree nodes of depth 3. Each node stores a floating-point threshold, left and right child pointers, and leaf value index.

[0101] Node splitting evaluator: For each leaf node, take 256 candidate feature thresholds, calculate the squared error gain, and select the node with the highest gain for splitting;

[0102] Learning rate regulator: The tree output is multiplied by 0.05 and written to the residual buffer; the learning rate is a global constant and can be fine-tuned in 0.01 steps on the rolling validation set;

[0103] Residual accumulator: Saves the current residual distribution at the end of each iteration, providing input for the natural gradient update layer;

[0104] The natural gradient update layer includes:

[0105] Score cache table: stores the log-likelihood gradient vector of each sample under the current distribution;

[0106] Fisher approximator: uses the outer product moving average of the score vectors as an approximation of the Fisher information matrix; window width 512 lines;

[0107] Parameter recalibrator: Solve the Fisher-score linear equations to obtain the natural gradient step size, and update the mean and variance parameter vectors in the multi-task distributed output layer;

[0108] Stabilization clamp: Performs truncation on entries with updated variance parameters less than 1e-4 or greater than 1e3 to prevent numerical explosion;

[0109] The multi-task distributed output layer includes:

[0110] Branch router: Copy the shared hidden vector to the five parallel branches based on the target field identifier;

[0111] Parameter expander: Each branch is mapped to mean and variance nodes using two fully connected layers, with a hidden dimension of 32 and an activation function of ReLU;

[0112] Distributed encoder: Combines mean and variance nodes into target field-mean-variance triples; fields include CPU utilization, memory utilization, network bandwidth usage, response latency, and energy consumption;

[0113] Output buffer queue: Writes the five triples into the inference result queue in a fixed order;

[0114] The improved NGBoost model is input into the main shadow and the corresponding counterfactual shadow respectively to obtain two sets of state distribution parameters. The state distribution parameters include the mean and variance of CPU utilization, the mean and variance of memory utilization, the mean and variance of network bandwidth usage, the mean and variance of response latency, and the mean and variance of energy consumption.

[0115] The improved NGBoost model is constructed as follows:

[0116] A feature normalization layer is added before the input of the traditional NGBoost model to perform missing data imputation and zero-mean normalization on CPU utilization, memory utilization, network bandwidth usage, remaining battery power, response latency, and control success rate in the terminal feature vector. An action embedding layer is added after the feature normalization layer and before the gradient boosting forest layer of the traditional NGBoost model. This layer maps the one-hot encoding of candidate control actions, takeover actions, and rollback actions into 16-dimensional action embedding vectors, which are then concatenated with the normalized terminal feature vectors in a fixed order before being input into the gradient boosting forest layer. The traditional NGBoost model is retained. The improved NGBoost model consists of a gradient boosting forest layer and a natural gradient update layer. The gradient boosting forest layer is configured with 400 regression trees of depth 3 and a learning rate of 0.05. The natural gradient update layer updates the distribution parameters after each iteration. The original single-task distribution output of the traditional NGBoost is replaced with a multi-task distribution output layer. The shared prediction features output by the natural gradient update layer are fed into five parallel output branches: CPU utilization, memory utilization, network bandwidth usage, response latency, and energy consumption. Each branch outputs the mean and variance of the corresponding index, resulting in the improved NGBoost model of this invention.

[0117] The difference between the mean values ​​of the state distributions corresponding to the primary shadow body and the counterfactual shadow body is used as the shadow divergence value. The link risk value is calculated by comparing the mean response delay corresponding to the primary shadow body with the maximum allowable delay threshold and the network bandwidth usage variance. The control conflict value is calculated by using the conflict flag and the response delay variance. The takeover trigger value is calculated by using the takeover flag and the CPU utilization variance.

[0118] The link risk value is calculated by comparing the mean response delay corresponding to the primary shadow body with the maximum allowable delay threshold and the network bandwidth usage variance. Specifically:

[0119] Read the predicted average response delay from the master shadow body and the maximum allowable response delay threshold from the control task; if the predicted average response delay does not exceed the threshold, set the delay over-limit ratio to zero; if the predicted average response delay exceeds the threshold, subtract the threshold from the predicted average response delay and then divide by the threshold to obtain the delay over-limit ratio; read the network bandwidth usage variance from the master shadow body, take the square root to obtain the bandwidth standard deviation, and then divide by the link bandwidth limit recorded in the edge node configuration table to obtain the bandwidth uncertainty. Weight the delay over-limit ratio and the bandwidth uncertainty with a weight of 6:4 and sum them up to three decimal places. The result is the link risk value.

[0120] The control conflict value is calculated using conflict markers and response delay variance, specifically as follows:

[0121] Read the conflict marker in the relation shadow; if the marker is no conflict, set the basic conflict term to zero; if the marker is conflict, set the basic conflict term to one. Read the response delay variance in the main shadow body, take the square root to get the response delay standard deviation, and then divide it by the above threshold to get the delay fluctuation coefficient. Multiply the basic conflict term and the delay fluctuation coefficient by one and one-half respectively, add them together, and round up to two decimal places. The result is the control conflict value.

[0122] The takeover trigger value is calculated based on the takeover flag and CPU utilization variance, specifically as follows:

[0123] Read the takeover flag in the relation shadow; if it is marked as no takeover required, set the basic takeover item to zero; if it is marked as takeoverable, set the basic takeover item to one. Read the CPU utilization variance in the main shadow body, take the square root to get the CPU utilization standard deviation, divide by one hundred to get the CPU fluctuation coefficient, multiply the basic takeover item by the unit value plus the CPU fluctuation coefficient and keep two decimal places. The result is the takeover trigger value.

[0124] Write the state distribution parameters, shadow divergence values, link risk values, control conflict values, and takeover trigger values ​​into the corresponding dual-track executable digital shadow;

[0125] In the training process of the improved NGBoost model, the multi-task negative log-likelihood loss function is adopted as the model loss function. This multi-task negative log-likelihood loss function corresponds one-to-one with the five output branches of the multi-task distributed output layer: CPU utilization branch, memory utilization branch, network bandwidth usage branch, response latency branch, and energy consumption branch. For each training sample, the actual CPU utilization, actual memory utilization, actual network bandwidth usage, actual response latency, and actual energy consumption are used as supervision labels, and the corresponding mean and variance of the multi-task distributed output layer are used as prediction distribution parameters. For any output branch, the single-item loss function is:

[0126] The single-item loss value is equal to half multiplied by the natural logarithm of the prediction variance, plus the square of the difference between the true label value and the predicted mean divided by twice the prediction variance. The prediction variance is obtained by processing the variance output node through the Softplus function and then adding 0.0001.

[0127] Among them, the true label value is the measured value of the corresponding state index in the training sample, the predicted mean is the mean of the state index output by the output branch, the predicted variance is the variance of the state index output by the output branch, the true label value of the CPU utilization branch is the true CPU utilization, the true label value of the memory utilization branch is the true memory utilization, the true label value of the network bandwidth usage branch is the true network bandwidth usage, the true label value of the response latency branch is the true response latency, and the true label value of the energy consumption branch is the true energy consumption.

[0128] The total loss value for a single training sample is obtained by multiplying the individual loss values ​​of the CPU utilization branch, memory utilization branch, network bandwidth usage branch, response latency branch, and energy consumption branch by their respective task weights and then summing them. The weights of the CPU utilization branch, memory utilization branch, network bandwidth usage branch, response latency branch, and energy consumption branch are 0.20, 0.15, 0.20, 0.30, and 0.15 respectively. The batch loss value is obtained by averaging the total loss values ​​of all training samples within a training batch.

[0129] In each training round, the gradient boosting forest layer calculates the residuals of each regression tree leaf node based on the batch loss value, and the natural gradient update layer recalibrates the gradient direction of the predicted mean and predicted variance based on the batch loss value. The recalibrated natural gradient is used as the fitting target of the regression tree in the next round, so that the improved NGBoost model reduces the prediction distribution error of the five state indicators round by round. When the batch loss value decreases by less than 0.001 for ten consecutive training rounds, or when the number of training rounds reaches 400, training is stopped and the model parameters are saved.

[0130] In this embodiment, generating candidate collaborative control sections includes:

[0131] The control task to be executed is analyzed to obtain the task identifier, task control target, terminal type required by the task, maximum allowable response time, target control action, takeover constraint and rollback constraint. Based on the terminal type required by the task, the set of participating terminals is selected from the dual-track executable digital shadows corresponding to each intelligent fusion terminal.

[0132] The terminal dimension is formed by the terminal identifiers corresponding to the participating terminal set; the time slice dimension is formed by the continuous time slices obtained according to the preset time granularity; and the collaboration layer dimension is formed by the main control layer, co-control layer, verification layer, takeover layer, and rollback layer. A reconfigurable collaborative time sequence cube is generated based on the terminal dimension, time slice dimension, and collaboration layer dimension. Specifically, the time slice dimension is formed by the continuous time slices obtained according to the preset time granularity.

[0133] Read the time granularity configuration from the task parameter table and set it to fifty milliseconds; read the maximum allowed response time of the task and set it to three seconds; divide the time granularity by the maximum allowed response time to get the number of time slices, and add one to get the last compensation slice; number each time slice sequentially starting from zero, and write the number, start millisecond, and end millisecond into the time slice index table to form a continuous time slice dimension;

[0134] A reconfigurable collaborative time-series cube is generated based on the terminal dimension, time-slice dimension, and collaboration layer dimension, specifically as follows:

[0135] The dimensions of the three-dimensional array are calculated in the order of terminal dimension length, time slice dimension length, and collaboration layer dimension length. A contiguous block of memory is allocated within the edge node and logically divided into a three-dimensional structure of terminal index, time slice index, and collaboration layer index. A mapping table is generated for each dimension to record the relationship between terminal identifier and row number, time slice number and column number, and collaboration layer type and layer number, respectively. All dimension combinations are traversed, and pointers to empty structures are filled into the corresponding cells of the three-dimensional array. The empty structures are used to write state distribution parameters, shadow divergence values, link risk values, control conflict values, and takeover trigger values. The pointers to the three-dimensional array, the pointers to the three mapping tables, and the memory size information are registered in the cube metadata table to complete the initialization of the reconfigurable collaborative time-series cube.

[0136] For each participating terminal, the state distribution parameters, shadow divergence values, link risk values, control conflict values, and takeover trigger values ​​obtained by the participating terminal under the target control action, takeover action, and rollback action are written into the corresponding cube cell.

[0137] The reconfigurable collaborative time-series cube is expanded according to the shadow divergence value. When the shadow divergence value of the participating terminal under the current control action exceeds the preset divergence threshold, a takeover judgment time slice or a rollback judgment time slice is added after the current time slice. The newly added time slice is associated with the participating terminal and the corresponding collaborative layer. The preset divergence threshold is set to 0.25. When the shadow divergence value is greater than 0.25, it is determined that there is a state evolution deviation between the main shadow body and the counterfactual shadow body, and a takeover judgment time slice or a rollback judgment time slice is added after the current time slice.

[0138] Based on link risk values, control conflict values, and takeover trigger values, the reconfigurable collaborative timing cube undergoes terminal dimension pruning and collaborative layer dimension switching. Terminals with link risk values ​​exceeding a preset risk threshold but takeover trigger values ​​not reaching a preset takeover threshold are deleted. Terminals with control conflict values ​​exceeding a preset conflict threshold are switched from the current collaborative layer to a rollback layer or a suppressed state. Specifically, the terminal dimension pruning and collaborative layer dimension switching for the reconfigurable collaborative timing cube based on link risk values, control conflict values, and takeover trigger values ​​are as follows:

[0139] Read the link risk value, control conflict value, and takeover trigger value of each unit corresponding to the terminal in the reconfigurable collaborative timing cube one by one; compare the link risk value with the preset risk threshold, the takeover trigger value with the preset takeover threshold, and the control conflict value with the preset conflict threshold; when the link risk value exceeds the preset risk threshold and the takeover trigger value does not reach the preset takeover threshold, delete the corresponding unit of the terminal in the current time slice and subsequent time slices; when the control conflict value exceeds the preset conflict threshold, move the corresponding unit of the terminal from the current collaborative layer to the rollback layer. If the terminal does not have a rollback operation, mark the corresponding unit of the terminal as a suppressed state; the preset risk threshold is 0.65, the preset takeover threshold is 0.70, and the preset conflict threshold is 0.60.

[0140] From the reconfigurable collaborative time sequence cube after time slice dimension expansion, terminal dimension pruning, and collaborative layer dimension switching, candidate collaborative control sections that satisfy task control objectives, maximum allowable response latency, takeover constraints, and rollback constraints are extracted.

[0141] In this embodiment, generating differential control tokens based on the target cooperative control path includes:

[0142] For each candidate cooperative control section, all cube cells are read sequentially, and the terminal identifier, time slice number, cooperative layer type, control action type, predicted response delay, shadow divergence value, link risk value, control conflict value, and takeover trigger value corresponding to each cube cell are recorded.

[0143] According to the ascending order of time slice numbers and the priority order of the collaborative layer, the cube units in the same candidate collaborative control section are connected in series to form a candidate collaborative control path. When connecting in series, only the main control layer unit is allowed to be followed by the co-control layer unit, the co-control layer unit is allowed to be followed by the verification layer unit, and the verification layer unit is allowed to be followed by the main control layer unit or the rollback layer unit.

[0144] For each candidate cooperative control path, a path cost is calculated. The path cost is equal to the sum of the predicted response delay, shadow divergence value, link risk value, and control conflict value corresponding to all cube cells in the path, multiplied by preset weights. Candidate cooperative control paths with predicted response delays exceeding the maximum allowable response delay or containing takeover trigger values ​​that do not reach the preset takeover threshold are filtered out. The preset weights include the predicted response delay weight, shadow divergence value weight, link risk value weight, and control conflict value weight. The predicted response delay weight is 0.35, the shadow divergence value weight is 0.25, the link risk value weight is 0.25, and the control conflict value weight is 0.15.

[0145] Among the remaining candidate cooperative control paths, the one with the lowest path cost is selected as the target cooperative control path. The order of differential control token issuance is determined according to the time slice number and cooperative layer type of each cube unit in the target cooperative control path. Specifically, the order of differential control token issuance is determined as follows:

[0146] Read the time slice number and coordination layer type of each cube unit in the target coordinated control path, and sort them in ascending order of time slice number; for multiple cube units in the same time slice, sort them in the order of master control layer, co-control layer, verification layer, takeover layer, and rollback layer; if there are multiple cube units in the same time slice and the same coordination layer, sort them in ascending order of link risk value; number the sorted cube units in sequence, and generate a differential control token distribution queue in the order of numbering;

[0147] Differential control tokens are generated sequentially for each cube unit in the target collaborative control path. Each differential control token includes a terminal identifier, collaboration layer type, control action, predicted state mean, allowable deviation range, deviation correction strategy, takeover trigger condition, rollback action, and feedback rules. The generated differential control tokens are stored in the token distribution queue. Specifically, the generation of differential control tokens for each cube unit in the target collaborative control path is as follows:

[0148] Read the cube cells in the target collaborative control path one by one according to the differential control token issuance queue; read the terminal identifier, collaboration layer type, control action, predicted state mean, state distribution variance, takeover trigger value, and rollback action flag from the cube cells; write the terminal identifier to the token receive field, the collaboration layer type to the token role field, the control action to the token action field, and the predicted state mean to the token predicted state field; calculate the allowable deviation range based on the state distribution variance and write it to the token deviation field; when the takeover trigger value reaches the preset takeover threshold, write the takeover action to the token takeover field, otherwise set the token takeover field to empty; when the rollback action flag is valid, write the rollback action to the token rollback field, otherwise set the token rollback field to empty; write the deviation correction strategy and feedback rules to the token control field, generate the corresponding differential control token, and write it to the token issuance queue according to the issuance order.

[0149] In this embodiment, each intelligent fusion terminal executes a corresponding control action, and when the deviation between the actual execution state and the predicted state exceeds the allowable deviation range, a deviation correction strategy is executed, including:

[0150] Edge nodes send differential control tokens to the corresponding intelligent converged terminals according to the token distribution queue. After receiving the differential control tokens, the intelligent converged terminals match and verify the terminal identifier, collaboration layer type, and control actions. After the matching and verification are passed, the control actions are executed, and real-state data during the execution process is collected. The real-state data includes real CPU utilization, real memory utilization, real network bandwidth usage, real response latency, and real energy consumption.

[0151] The intelligent fusion terminal calculates the deviation between the actual state data and the predicted state mean in the differential control token to obtain the execution deviation value. The execution deviation value satisfies the following: the execution deviation value is equal to the sum of the CPU utilization deviation, memory utilization deviation, network bandwidth usage deviation, response latency deviation, and energy consumption deviation multiplied by their respective state weights. Among them, the CPU utilization deviation is the absolute value of the difference between the actual CPU utilization and the predicted average CPU utilization; the memory utilization deviation is the absolute value of the difference between the actual memory utilization and the predicted average memory utilization; the network bandwidth usage deviation is the absolute value of the difference between the actual network bandwidth usage and the predicted average network bandwidth usage; the response latency deviation is the absolute value of the difference between the actual response latency and the predicted average response latency; and the energy consumption deviation is the absolute value of the difference between the actual energy consumption and the predicted average energy consumption.

[0152] When the execution deviation value exceeds the allowable deviation range, the intelligent fusion terminal performs takeover action, rollback action, control action downgrade or control action suppression according to the deviation correction strategy in the differential control token, and generates feedback data containing the execution deviation value, deviation correction strategy type and correction execution result.

[0153] Edge nodes receive execution data from each intelligent fusion terminal, and update the master shadow body, counterfactual shadow body, and shadow coupling gating unit in the corresponding dual-track executable digital shadow based on the execution data. They also update the link risk value, control conflict value, takeover trigger value, and path weight of the corresponding cube unit in the reconfigurable collaborative temporal cube.

[0154] refer to Figure 2 A collaborative control system for intelligent converged terminals based on edge computing, comprising:

[0155] The data preprocessing module is used to acquire and preprocess raw terminal data from multiple smart converged terminals to generate a standard terminal dataset.

[0156] The dual-track shadow construction module is used to construct dual-track executable digital shadows for each intelligent fusion terminal based on a standard terminal dataset.

[0157] The probability prediction module is used to predict the state evolution under different control actions based on the dual-track executable digital shadow and the improved NGBoost model, so as to obtain the state distribution parameters and risk parameters.

[0158] The cube reconstruction module is used to construct and reconstruct a collaborative time-series cube based on the control task to be executed and the risk parameters, and generate candidate collaborative control sections.

[0159] The path token generation module is used to search for the target cooperative control path based on the candidate cooperative control cross-section and generate differential control tokens;

[0160] The execution feedback update module is used to issue differential control tokens to control the intelligent fusion terminal to perform control actions, and to update the dual-track executable digital shadow and collaborative timing cube based on the execution feedback.

[0161] Example 1: In a device linkage monitoring area, the system connects to six types of intelligent fusion terminals, including image acquisition terminals, environmental sensing terminals, edge inference terminals, audio-visual execution terminals, backup control terminals, and communication gateway terminals. Within one control cycle, the edge node receives 1800 raw terminal data entries, including CPU utilization, memory utilization, network bandwidth usage, remaining power, historical response latency, historical control success rate, abnormal interruption count, dependency flags, conflict flags, and takeover flags. The raw data contains 18 duplicate entries, 31 missing fields, and 14 abnormal fields. The highest abnormal value for network bandwidth usage reaches 132%, which is outside the data range.

[0162] The system preprocesses the raw terminal data, unifying the sampling granularity of data uploaded from different terminals to 50ms, deleting duplicate data, filling missing CPU utilization and network bandwidth usage fields with the median of the corresponding terminal's most recent 10 sampling periods, removing resource usage fields exceeding 100% as an anomaly, and remapping terminal identifiers to fixed numbers. After processing, 1737 standard terminal data entries are obtained, improving field completeness from 96.8% to 100%, and reducing data time alignment error from a maximum of 210ms to within 50ms.

[0163] In the standard terminal dataset, the image acquisition terminal has a CPU utilization of 42%, a memory utilization of 51%, a network bandwidth utilization of 38%, and an average latency of 118ms for the last 100 responses; the edge inference terminal has a CPU utilization of 58%, a memory utilization of 63%, a network bandwidth utilization of 34%, an average latency of 146ms for the last 100 responses, and an abnormal interruption count of 2; the backup control terminal has a CPU utilization of 33%, a memory utilization of 41%, a network bandwidth utilization of 22%, an average latency of 92ms for the last 100 responses, and a takeover flag of 1. The system reads the above fields one by one according to the terminal identifier, generates a capability shadow, a status shadow, a behavior shadow, and a relationship shadow for each terminal, and concatenates the four types of shadows in the order of capability segment, status segment, behavior segment, and relationship segment to generate the main shadow body.

[0164] The system generates a counterfactual shadow entity for each primary shadow entity. Taking an edge inference terminal as an example, the CPU increment corresponding to a candidate control action is 18 percentage points, the memory increment is 11 percentage points, the bandwidth increment is 8 percentage points, the power consumption is 2 percentage points, and the basic execution latency is 95ms. The system adds 18% to the current CPU utilization of 58% to obtain a predicted CPU utilization of 76%; adds 11% to the current memory utilization of 63% to obtain a predicted memory utilization of 74%; adds 8% to the current bandwidth usage of 34% to obtain a predicted bandwidth usage of 42%; and adds 95ms to the average response latency of 146ms to obtain a predicted response latency of 241ms. The system generates three action records in the order of candidate control action, takeover action, and rollback action, and writes them into the counterfactual shadow entity.

[0165] The shadow coupling gating unit calculates the state change amplitude based on the latest 10 sampling periods, the historical behavior fluctuation amplitude based on the latest 50 control logs, and the frequency of collaborative relationship changes based on the changes in relationship markers within the last 60 seconds. The state change amplitude of the edge inference terminal is 0.18, the historical behavior fluctuation amplitude is 0.11, and the collaborative relationship change frequency is 0.06. After weighting by 0.5, 0.3, and 0.2, the synchronization update score is 0.135, and synchronization update is not triggered. The state change amplitude of the standby control terminal is 0.31, exceeding 0.25. The system immediately refreshes its primary shadow body and counterfactual shadow body and writes the new shadow version number.

[0166] The improved NGBoost model was trained using 12,000 historical control samples and validated using 3,000 samples. In the training samples, Sample 1 represents a candidate control action from an edge inference terminal, with inputs of CPU utilization (58%), memory utilization (63%), bandwidth usage (34%), response latency (146ms), historical success rate (94.8%), and abnormal interruption count (2). The outputs are: actual CPU utilization (78%), actual memory utilization (75%), actual bandwidth usage (43%), actual response latency (258ms), and actual energy consumption (4.8%). Sample 2 represents a takeover action from a backup control terminal, with inputs of CPU utilization (33%), memory utilization (41%), bandwidth usage (22%), response latency (92ms), historical success rate (96.9%), and abnormal interruption count (1). The outputs are: actual CPU utilization (49%), actual memory utilization (53%), actual bandwidth usage (32%), actual response latency (188ms), and actual energy consumption (3.9%). Sample 3 is a candidate control action of the audio-visual execution terminal. The input CPU utilization is 27%, memory utilization is 35%, bandwidth usage is 18%, and response latency is 73ms. The output actual response latency is 96ms and actual power consumption is 1.7%.

[0167] After the model input vector enters the feature normalization layer, the CPU, memory, bandwidth, and response latency fields are converted into zero-mean normalized values. Candidate control actions, takeover actions, and rollback actions are processed by the action embedding layer to generate 16-dimensional action embedding vectors. These vectors then enter a gradient boosting forest layer consisting of 400 regression trees of depth 3, with a learning rate of 0.05. The natural gradient update layer updates the distribution parameters after each iteration. The multi-task distribution output layer outputs the mean and variance of CPU utilization, memory utilization, network bandwidth usage, response latency, and energy consumption, respectively. In the validation set, the mean absolute error of response latency is 21.4 ms, the mean error of CPU utilization is 4.7 percentage points, and the mean error of network bandwidth usage is 5.1 percentage points.

[0168] In a control task to be executed, the maximum allowable response latency is 3 seconds, with a time granularity of 50 milliseconds. The system constructs a reconfigurable collaborative timing cube, forming a terminal dimension with 6 terminals, dividing the time from 0 milliseconds to 3000 milliseconds into 61 time slices, and forming a collaborative layer dimension with the main control layer, co-control layer, verification layer, takeover layer, and rollback layer, generating a total of 1830 cube units. The shadow divergence value of the edge inference terminal is 0.29, exceeding the preset divergence threshold of 0.25. The system adds a takeover judgment time slice after the current time slice and associates it with a backup control terminal. The link risk value of the communication gateway terminal is 0.68, exceeding the preset risk threshold of 0.65, and the takeover trigger value is 0.42, which does not reach the preset takeover threshold of 0.70. Therefore, the system deletes the corresponding unit of this terminal in the current and subsequent time slices. The control conflict value of the audio-visual execution terminal is 0.64, exceeding the preset conflict threshold of 0.60. Therefore, it is switched to the rollback layer to await confirmation.

[0169] The system generates three candidate cooperative control paths. Path 1 consists of an edge inference terminal as the main controller, an image acquisition terminal as a co-controller, an environmental sensor terminal as a verifier, and a backup control terminal as the takeover point. The total predicted response latency is 642ms, with a shadow divergence value of 0.29, a link risk value of 0.41, and a control conflict value of 0.22. The path cost, calculated using weights of 0.35, 0.25, 0.25, and 0.15, is 0.319. Path 2 includes a communication gateway terminal, with a link risk value of 0.68, and is therefore filtered out. Path 3 consists of a backup control terminal as the main controller, an image acquisition terminal as a co-controller, an environmental sensor terminal as a verifier, and an edge inference terminal as the takeover point. The total predicted response latency is 698ms, with a shadow divergence value of 0.17, a link risk value of 0.36, a control conflict value of 0.20, and a path cost of 0.297. The system selects Path 3 as the target cooperative control path.

[0170] The system generates a differential control token based on the target path. The backup control terminal's predicted average response latency is 186ms, with a response latency variance of 361 (square root yields 19ms), allowing a response latency deviation of 38ms. The average CPU utilization is 48.7%, with a CPU variance of 14.2 (square root yields approximately 3.8 percentage points), allowing a CPU deviation of 7.6 percentage points. Upon receiving the token, the terminal executes the master control action. The actual response latency is 211ms, deviating from the predicted value by 25ms, which is within the 38ms limit, and execution continues. When the edge inference terminal acts as a takeover candidate, the actual response latency reaches 331ms, exceeding the predicted value by 284ms, with a deviation of 47ms, exceeding the allowable deviation. The terminal immediately executes a degraded inference strategy, switching from full inference to keyframe inference, completing the correction feedback within 18ms.

[0171] In the comparative experiment, the traditional static scheduling method selects the terminal only based on the current CPU, memory, and bandwidth, without constructing a dual-track shadow, predicting the distribution, or generating differential tokens. Both the method of this invention and the traditional method use 12,000 training samples and 3,000 validation samples, and are tested in 500 simulated control tasks.

[0172] The average first response latency of the traditional method is 348ms, while that of this invention is 216ms;

[0173] The traditional method has a 912ms response time at the 95th percentile, while this invention has a response time of 438ms.

[0174] The average takeover trigger latency of the traditional method is 312ms, while that of this invention is 74ms;

[0175] The traditional method has an average of 1.08 rescheduling times per task, while the present invention has an average of 0.21 times.

[0176] Traditional methods control the conflict occurrence rate at 14.3%, while this invention achieves 3.1%.

[0177] The success rate of takeover using traditional methods is 82.6%, while that of this invention is 96.4%.

[0178] The traditional method achieves a task completion rate of 87.3%, while this invention achieves 97.7%.

[0179] 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 method for cooperative control of an intelligent converged terminal based on edge computing, characterized in that, include: Obtain raw terminal data from multiple intelligent converged terminals, preprocess the raw terminal data, and obtain a standard terminal dataset; Based on the standard terminal dataset, a dual-track executable digital shadow is constructed for each intelligent fusion terminal. The dual-track executable digital shadow includes a main shadow body, a counterfactual shadow body, and a shadow coupling gating unit. Based on each dual-track executable digital shadow, the improved NGBoost model is used to predict the state evolution of the master shadow and counterfactual shadow under different control actions, and obtain the state distribution parameters, shadow divergence value, link risk value, control conflict value and takeover trigger value. A reconfigurable collaborative time-series cube is constructed based on the control task to be executed. The state distribution parameters, shadow divergence values, link risk values, control conflict values, and takeover trigger values ​​are written into the corresponding cube cells. Based on the shadow divergence values, the reconfigurable collaborative time-series cube is expanded along the time slice dimension, pruned along the terminal dimension, and switched along the collaborative layer dimension to generate candidate collaborative control sections. Based on each candidate cooperative control section, a cooperative path search is performed to generate a target cooperative control path, and a differential control token is generated based on the target cooperative control path. Differential control tokens are sent to the corresponding intelligent fusion terminals. Each intelligent fusion terminal executes the corresponding control action. When the deviation between the actual execution state and the predicted state exceeds the allowable deviation range, a deviation correction strategy is executed. The corresponding dual-track executable digital shadow and reconfigurable collaborative timing cube are updated based on the execution data fed back by each intelligent fusion terminal. The generation of candidate collaborative control sections includes: The control task to be executed is analyzed to obtain the task identifier, task control target, terminal type required by the task, maximum allowable response time, target control action, takeover constraint and rollback constraint. Based on the terminal type required by the task, the set of participating terminals is selected from the dual-track executable digital shadows corresponding to each intelligent fusion terminal. The terminal dimension is formed by the terminal identifier corresponding to the participating terminal set, the time slice dimension is formed by the continuous time slices obtained according to the preset time granularity, and the collaboration layer dimension is formed by the main control layer, the co-control layer, the verification layer, the takeover layer and the rollback layer. A reconfigurable collaborative time sequence cube is generated based on the terminal dimension, the time slice dimension and the collaboration layer dimension. For each participating terminal, the state distribution parameters, shadow divergence values, link risk values, control conflict values, and takeover trigger values ​​obtained by the participating terminal under the target control action, takeover action, and rollback action are written into the corresponding cube cell. The reconfigurable collaborative timing cube is expanded according to the shadow divergence value. When the shadow divergence value of the participating terminal under the current control action exceeds the preset divergence threshold, a takeover judgment time slice or a rollback judgment time slice is added after the current time slice. The newly added time slice is associated with the participating terminal and the corresponding collaborative layer. Based on the link risk value, control conflict value, and takeover trigger value, the reconfigurable collaborative time sequence cube is subjected to terminal dimension pruning and collaborative layer dimension switching. Terminal units whose link risk value exceeds the preset risk threshold and whose takeover trigger value does not reach the preset takeover threshold are deleted. Terminal units whose control conflict value exceeds the preset conflict threshold are switched from the current collaborative layer to the rollback layer or suppression state. From the reconfigurable collaborative time sequence cube after time slice dimension expansion, terminal dimension pruning, and collaborative layer dimension switching, candidate collaborative control sections that satisfy task control objectives, maximum allowable response latency, takeover constraints, and rollback constraints are extracted. 2.The method according to claim 1, wherein, The original terminal data includes terminal identification data, terminal type data, terminal computing resource data, terminal storage resource data, terminal communication resource data, terminal power status data, terminal control interface data, terminal historical response latency data, terminal historical control success rate data, terminal abnormal interruption data, and inter-terminal collaboration relationship data. 3.The method according to claim 1, wherein, The preprocessing of the raw terminal data to obtain a standard terminal dataset includes: The original terminal data is subjected to timestamp unification, sampling frequency alignment, duplicate data deletion, outlier removal, missing value imputation, unit normalization, terminal identifier mapping, and cooperative relationship encoding to obtain a standard terminal dataset. 4.The method according to claim 1, wherein, The construction of a dual-track executable digital shadow for each intelligent fusion terminal based on a standard terminal dataset includes: Within each edge node, traverse the standard terminal dataset one by one according to the terminal identifier; For each intelligent fusion terminal, the acquisition capability, computing capability and control interface are encapsulated as capability shadows, the current resource occupancy, communication status and power status are encapsulated as status shadows, the historical response latency, historical control success rate and abnormal interruption records are encapsulated as behavior shadows, and the dependency relationship, conflict relationship and takeover relationship are encapsulated as relationship shadows. The four types of shadows are combined to generate the main shadow body. Three types of preset control actions are set for each master shadow body: candidate control action, takeover action, and rollback action. By calling the built-in state transition rules, each type of preset control action is quickly deduced to obtain the corresponding predicted resource changes, predicted communication changes, and predicted control results. The deduction results are associated with the corresponding actions to generate a counterfactual shadow body. A shadow coupling gating unit is configured for each main shadow entity and its corresponding counterfactual shadow entity. The shadow coupling gating unit dynamically calculates the synchronization update triggering conditions of the main shadow entity and the counterfactual shadow entity based on the real-time state change amplitude of the terminal, the historical behavior fluctuation amplitude, and the frequency of changes in the collaborative relationship. When the triggering conditions are met, the shadow fields of the two shadows are refreshed synchronously. The master shadow body, counterfactual shadow body, and shadow coupling gating unit are bound together. After writing a unified identifier, version number, and update timestamp, a dual-track executable digital shadow of the intelligent fusion terminal is generated and stored in the digital shadow library of the edge node. 5.The method of claim 1, wherein, The obtained state distribution parameters, shadow divergence values, link risk values, control conflict values, and takeover trigger values ​​include: For each dual-track executable digital shadow, read its acquisition capability, computing capability, control interface, current CPU utilization, current memory utilization, current network bandwidth usage, current remaining power, average response latency of the last 100 times, control success rate of the last 100 times, count of the last 100 abnormal interruptions, dependency flag, conflict flag, and takeover flag, and concatenate them into a terminal feature vector according to the preset field order. After one-hot encoding of candidate control actions, takeover actions and rollback actions, they are input into the action embedding layer to obtain a sixteen-dimensional action embedding vector. The action embedding vector is then concatenated bitwise with the terminal feature vector that has been normalized to zero mean to form the model input vector. The model input vector is sequentially fed into the feature normalization layer, action embedding layer, gradient boosting forest layer, natural gradient update layer, and multi-task distribution output layer of the improved NGBoost model. The gradient boosting forest layer is set with 400 regression trees of depth 3 and iterates with a learning rate of 0.

05. The natural gradient update layer updates the distribution parameters at the end of each iteration. The multi-task distribution output layer is a branch output module connected in series after the original NGBoost output end, which is used to simultaneously output the mean and variance of CPU utilization distribution, memory utilization distribution, network bandwidth usage distribution, response latency distribution, and energy consumption distribution. The improved NGBoost model is input into the main shadow and the corresponding counterfactual shadow respectively to obtain two sets of state distribution parameters. The state distribution parameters include the mean and variance of CPU utilization, the mean and variance of memory utilization, the mean and variance of network bandwidth usage, the mean and variance of response latency, and the mean and variance of energy consumption. The difference between the mean values ​​of the state distributions corresponding to the primary shadow body and the counterfactual shadow body is used as the shadow divergence value. The link risk value is calculated by comparing the mean response delay corresponding to the primary shadow body with the maximum allowable delay threshold and the network bandwidth usage variance. The control conflict value is calculated by using the conflict marker and the response delay variance. The takeover trigger value is calculated by using the takeover marker and the CPU utilization variance. Write the state distribution parameters, shadow divergence values, link risk values, control conflict values, and takeover trigger values ​​into the corresponding dual-track executable digital shadow. 6.The method of claim 1, wherein, The step of generating differential control tokens based on the target collaborative control path includes: For each candidate cooperative control section, all cube cells are read sequentially, and the terminal identifier, time slice number, cooperative layer type, control action type, predicted response delay, shadow divergence value, link risk value, control conflict value, and takeover trigger value corresponding to each cube cell are recorded. According to the ascending order of time slice numbers and the priority order of the collaborative layer, the cube units in the same candidate collaborative control section are connected in series to form a candidate collaborative control path. When connecting in series, only the main control layer unit is allowed to be followed by the co-control layer unit, the co-control layer unit is allowed to be followed by the verification layer unit, and the verification layer unit is allowed to be followed by the main control layer unit or the rollback layer unit. For each candidate cooperative control path, the path cost is calculated. The path cost is equal to the sum of the predicted response delay, shadow divergence value, link risk value and control conflict value of all cube cells in the path, multiplied by a preset weight. Candidate cooperative control paths with predicted response delay exceeding the maximum allowable response delay or containing takeover trigger values ​​that have not reached the preset takeover threshold are filtered out. The path with the lowest path cost among the remaining candidate cooperative control paths is selected as the target cooperative control path. The order of issuance of differential control tokens is determined according to the time slice number and cooperative layer type of each cube unit in the target cooperative control path. Differential control tokens are generated sequentially for each cube unit in the target collaborative control path. Each differential control token includes terminal identifier, collaborative layer type, control action, predicted state mean, allowable deviation range, deviation correction strategy, takeover trigger condition, rollback action and feedback rule. The generated differential control tokens are stored in the token distribution queue. 7.The method according to claim 1, wherein, Each intelligent fusion terminal executes a corresponding control action, and when the deviation between the actual execution state and the predicted state exceeds the allowable deviation range, it executes a deviation correction strategy, including: Edge nodes send differential control tokens to the corresponding smart fusion terminals according to the token distribution queue. After receiving the differential control token, the smart fusion terminal performs matching and verification of the terminal identifier, collaboration layer type and control action. After the matching and verification is passed, the control action is executed and real status data during the execution process is collected. The intelligent fusion terminal calculates the deviation between the real state data and the predicted state mean in the differential control token to obtain the execution deviation value. The execution deviation value satisfies the following condition: the execution deviation value is equal to the sum of the CPU utilization deviation, memory utilization deviation, network bandwidth usage deviation, response latency deviation and energy consumption deviation multiplied by their respective state weights. When the execution deviation value exceeds the allowable deviation range, the intelligent fusion terminal performs takeover action, rollback action, control action downgrade or control action suppression according to the deviation correction strategy in the differential control token, and generates feedback data containing the execution deviation value, deviation correction strategy type and correction execution result. Edge nodes receive execution data from each intelligent fusion terminal, and update the master shadow body, counterfactual shadow body, and shadow coupling gating unit in the corresponding dual-track executable digital shadow based on the execution data. They also update the link risk value, control conflict value, takeover trigger value, and path weight of the corresponding cube unit in the reconfigurable collaborative temporal cube.

8. A collaborative control system for an edge-computing-based intelligent converged terminal, comprising executing the collaborative control method for an edge-computing-based intelligent converged terminal as described in any one of claims 1 to 7, characterized in that, include: The data preprocessing module is used to acquire and preprocess raw terminal data from multiple smart converged terminals to generate a standard terminal dataset. The dual-track shadow construction module is used to construct dual-track executable digital shadows for each intelligent fusion terminal based on a standard terminal dataset. The probability prediction module is used to predict the state evolution under different control actions based on the dual-track executable digital shadow and the improved NGBoost model, so as to obtain the state distribution parameters and risk parameters. The cube reconstruction module is used to construct and reconstruct a collaborative time-series cube based on the control task to be executed and the risk parameters, and generate candidate collaborative control sections. The path token generation module is used to search for the target cooperative control path based on the candidate cooperative control cross-section and generate differential control tokens; The execution feedback update module is used to issue differential control tokens to control the intelligent fusion terminal to perform control actions, and to update the dual-track executable digital shadow and collaborative timing cube based on the execution feedback.

Images