A large model-based data fusion intelligent device linkage management and control system

Abstract

The present application relates to the technical field of data processing, and more particularly to a data fusion intelligent device linkage control system based on a large model, comprising a data acquisition module that outputs standardized data carrying a privacy classification identifier; a data processing module that outputs a synchronization-bound optimization algorithm identifier and a collaborative strategy identifier; a privacy algorithm fusion scheduling module that generates atomized computing graph data and strategy audit logs for fusion privacy operations; a multiple optimization resource module that executes computing graph data and drives a parallel simulation environment to pre-play resource schemes; an arbitrator that activates a double closed-loop feedback module when detecting a deviation that exceeds a standard for three consecutive times; a security instruction generation module that fuses output results and audit logs to generate a digital signature device control instruction; and a double closed-loop feedback module that updates knowledge base weight parameters through a data feedback loop, and controls a feedback loop to adjust resource allocation thresholds based on simulation confidence levels, thereby achieving a dynamic balance between resource efficiency and privacy strength.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a data fusion intelligent device linkage management and control system based on a large model. BACKGROUND

[0002] The intelligent device linkage management and control system is a comprehensive management platform based on Internet of Things technology and artificial intelligence algorithms, aiming to realize collaborative control and optimized operation among multiple intelligent devices; the system collects environmental data from various sensors in real time through a central processing unit, dynamically analyzes the data in combination with preset logical rules or machine learning models, and outputs accurate control commands, so that lighting, security, temperature control and other devices can respond to external condition changes in coordination, significantly reducing the demand for manpower in the automated process, and improving the safety protection level and energy utilization efficiency in smart home or industrial scenarios.

[0003] In the cross-scene data fusion of the intelligent device linkage management and control system, the algorithm adaptive scheduling and privacy security collaborative optimization technology problems occur, which leads to the inability of efficient integration of algorithm dynamic scheduling process and privacy protection mechanism, resulting in conflicts in resource allocation and data processing; specifically, in the smart community scenario, the system needs to schedule a face recognition algorithm in real time based on environmental factors to optimize access control, but the privacy security layer is not dynamically coupled with the scheduling decision, for example, when the access control and video stream data are fused, the algorithm adjustment needs to recalculate the feature vector, while the privacy encryption calculation is carried out independently, increasing the calculation overhead and data transmission exposure risk, causing response delay and privacy leakage hazards, and limiting the robust deployment of the system under variable conditions. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides an intelligent device linkage management and control system based on data fusion of a large model, which solves the technical problem of algorithm adaptive scheduling and privacy security collaborative optimization in cross-scene data fusion.

[0005] To solve the above technical problems, the specific technical solutions of the present application are as follows:

[0006] The present application provides an intelligent device linkage management and control system based on data fusion of a large model, which comprises:

[0007] A data acquisition module acquires the original data stream of the terminal device and outputs standardized data carrying a privacy classification identifier;

[0008] A data processing module receives environmental parameters in the standardized data, queries a dynamic knowledge base to output a synchronized binding of an optimization algorithm identifier and a collaborative strategy identifier;

[0009] The privacy algorithm fusion scheduling module parses the synchronization bound optimization algorithm identifier and the cooperation strategy identifier, compiles the privacy operation corresponding to the cooperation strategy identifier into a subgraph node of the pre-training algorithm calculation graph, generates atomized calculation graph data of the fused privacy operation, and outputs the atomized calculation graph data and the associated strategy audit log;

[0010] The multiple optimization resource module allocates computing resources to execute the atomized calculation graph data, and simultaneously drives the parallel simulation environment to pre-play the resource scheme.

[0011] The security instruction generation module receives the output result of the atomized calculation graph data, receives the strategy audit log output by the privacy algorithm fusion scheduling module, attaches the strategy audit log to the output result to generate to-be-signed data, and executes digital signature on the to-be-signed data to generate device control instructions.

[0012] The double closed-loop feedback module receives the delay index generated by the device execution control instruction, updates the dynamic knowledge base weight parameter through a data feedback loop, and adjusts the resource allocation threshold according to the simulation confidence level through a control feedback loop.

[0013] The data acquisition module outputs standardized data to the data processing module, the data processing module outputs the bound identifier to the privacy algorithm fusion scheduling module, the privacy algorithm fusion scheduling module outputs the atomized calculation graph data and the strategy audit log to the multiple optimization resource module and the security instruction generation module, the multiple optimization resource module activates the double closed-loop feedback module after triggering the threshold determination of the arbitrator, the double closed-loop feedback module outputs the updated dynamic knowledge base weight parameter to the data processing module and outputs the resource allocation threshold to the multiple optimization resource module, and the security instruction generation module outputs the device control instruction to the terminal device and feeds back the execution delay index to the double closed-loop feedback module.

[0014] Further, the data fusion intelligent device linkage management and control system based on a large model comprises a protocol adaptation layer and an identifier injection unit.

[0015] The protocol adaptation layer converts the communication protocol of the terminal device into a unified data format, and outputs preprocessed data carrying a device type code.

[0016] The identifier injection unit receives the preprocessed data, indexes the privacy classification rule base based on the device type code, and generates matched privacy classification identifier embedded data headers.

[0017] The privacy classification rule base stores the mapping relationship between the device type and the privacy level, and dynamically responds to the index request.

[0018] The identifier injection unit outputs the standardized data integrating the privacy classification identifier to the data processing module;

[0019] The protocol adaptation layer processes the original data stream to generate pre-processed data output to the identifier injection unit;

[0020] The identifier injection unit extracts the device type code in the pre-processed data to retrieve the privacy classification rule library;

[0021] The identifier injection unit embeds the generated privacy classification identifier into the pre-processed data to generate standardized data output to the data processing module.

[0022] Further, the big model-based data fusion intelligent device linkage management and control system disclosed by the application further comprises:

[0023] The data processing module is provided with a hierarchical feature extraction engine;

[0024] The first processing layer performs illumination environment analysis: input the video frame in the standardized data, extract the brightness histogram, calculate the 95th percentile brightness value, and output the brightness scalar;

[0025] The second processing layer performs dynamic target density analysis: input the standardized data, adopt background modeling and morphological filtering, count the number of moving objects per square meter, and output the density scalar;

[0026] The feature fusion unit receives the brightness scalar and the density scalar, queries the dynamic weighting coefficient according to the privacy classification identifier, and calculates the composite feature vector;

[0027] The dynamic knowledge base matches the composite feature vector with the pre-stored scene strategy mapping relationship, and outputs the bound optimization algorithm identifier and the cooperative strategy identifier to the privacy algorithm fusion scheduling module;

[0028] The standardized data is input to the first processing layer and the second processing layer;

[0029] The first processing layer outputs the brightness scalar to the feature fusion unit;

[0030] The second processing layer outputs the density scalar to the feature fusion unit;

[0031] The feature fusion unit receives the privacy classification identifier to dynamically calculate the weighting value, fuses the brightness scalar and the density scalar to generate the composite feature vector.

[0032] Further, the big model-based data fusion intelligent device linkage management and control system disclosed by the application, the dynamic knowledge base comprises a strategy index layer, a weight storage layer and a strategy updating unit;

[0033] The strategy index layer stores the mapping relationship between the composite feature vector and the bound optimization algorithm identifier and the cooperative strategy identifier;

[0034] The weight storage layer records the historical performance indicators of each feature vector in the mapping relationship, including response delay and recognition accuracy;

[0035] The policy updating unit receives the execution delay indicators transmitted by the data feedback loop, and converts the delay indicators into weight correction coefficients;

[0036] The policy updating unit receives the confidence level classification signals transmitted by the control feedback loop, and dynamically adjusts the update frequency according to the high / medium / low three levels;

[0037] The policy updating unit fuses the weight correction coefficients and the update frequency parameters, and outputs the update instruction to the policy index layer to adjust the mapping relationship;

[0038] The policy updating unit synchronously outputs the update instruction to the weight storage layer to calibrate the historical performance indicators, and the weight storage layer outputs the latest historical performance indicators to the policy index layer to optimize the real-time matching accuracy. The data feedback loop transmits the execution delay indicators to the policy updating unit, the control feedback loop transmits the confidence level classification signals to the policy updating unit, the policy updating unit outputs the update instruction to the policy index layer, the policy updating unit outputs the update instruction to the weight storage layer, the weight storage layer transmits the latest historical performance indicators to the policy index layer, and the policy index layer outputs the optimized binding of the optimization algorithm identifier and the collaborative strategy identifier to the privacy algorithm fusion scheduling module.

[0039] Further, the data fusion intelligent device linkage management and control system based on a large model provided by the application is also used for:

[0040] Receiving the synchronously bound optimization algorithm identifier and the collaborative strategy identifier output by the data processing module;

[0041] Analyzing the optimization algorithm identifier to locate the feature extraction layer output node position in the pre-training model calculation graph;

[0042] Indexing the corresponding privacy rule library of the collaborative strategy identifier to obtain the noise amplitude parameter, inserting the noise injection node configured with the noise amplitude parameter at the located node position, dynamically linking the privacy encryption dynamic library corresponding to the collaborative strategy identifier to the noise injection node, compiling the calculation graph embedded with the noise node and the encryption library to generate the atomized calculation graph data, synchronously generating the strategy audit log recording the node positioning, noise parameter configuration, and encryption library linking operation details, outputting the atomized calculation graph data to the multiple optimization resource module, and outputting the strategy audit log to the security instruction generation module.

[0043] Further, the data fusion intelligent device linkage management and control system based on a large model provided by the application further comprises:

[0044] The atomized calculation graph data is packaged into an independent container;

[0045] The container internal integration strategy consistency engine performs the following cooperative operations:

[0046] Load the algorithm executable file of the fusion privacy operation;

[0047] Dynamically parse the configuration file corresponding to the cooperative strategy identifier, and generate noise amplitude and encryption library link parameters in real time;

[0048] Synchronously start the log recording service to continuously capture the noise injection node and the running state of the encryption library;

[0049] The container automatically checks the matching of the digital signature of the configuration file and the registration signature of the cooperative strategy identifier when starting;

[0050] The runtime synchronously changes the cooperative strategy identifier to the noise injection node and the encryption library link in real time.

[0051] Further, the large model-based data fusion intelligent device linkage management and control system disclosed by the application further comprises:

[0052] Load the algorithm executable file of the fusion noise node and the encryption library, parse the configuration file corresponding to the cooperative strategy identifier, and generate noise amplitude parameters and encryption library link instructions;

[0053] Start the log recording service to capture noise node output data and encryption library operation state;

[0054] In the container initialization stage, check the hash value deviation of the digital signature of the configuration file and the registration signature of the cooperative strategy identifier, and trigger the fuse alarm when the deviation value is > 5%;

[0055] In the container running stage, listen to the cooperative strategy identifier change event, and reconfigure the noise amplitude parameter and the encryption library link within ≤50 milliseconds;

[0056] Periodically verify the integrity of the configuration file, freeze the container and start a backup instance when tampering is detected, and the log recording service outputs noise indicators and encryption audit records to the security instruction generation module in real time.

[0057] Further, the large model-based data fusion intelligent device linkage management and control system disclosed by the application further comprises:

[0058] The arbitrator calculates the Euclidean distance between the simulation index and the actual execution index, activates the double closed-loop feedback module when the distance is detected for three times continuously exceeds the dynamic threshold, and sends a fuse trigger signal to the container;

[0059] After the container receives the fuse signal, the current running instance is frozen, the pure backup container is started to load the atomized calculation graph data, the container switching delay time is recorded, and the switching delay index is fed back to the double closed-loop feedback module.

[0060] The control feedback loop adjusts the dynamic threshold based on the simulated confidence level signal. The threshold for high-confidence scenarios is set to 80% of the dynamic threshold baseline value, and the threshold for low-confidence scenarios is set to 120% of the dynamic threshold baseline value.

[0061] The dual closed-loop feedback module calibrates the generation logic of the confidence grading signal based on the container switching delay index. When the switching delay is greater than 50 milliseconds, the simulation confidence grading signal is automatically downgraded by one level.

[0062] Furthermore, in the data fusion intelligent device linkage management and control system based on a large model described in the present invention, the security instruction generation module includes a reverse verification unit and an audit encapsulation unit;

[0063] The reverse verification unit receives the feature vector output by the atomic computation graph data, performs feature reconstruction to generate reconstructed data, and calculates the deviation between the feature vector and the reconstructed data;

[0064] When the deviation value is less than 5%, the irreversibility verification pass signal is output;

[0065] When the deviation value is ≥5%, a fuse trigger signal is output to the container instance of claim 7;

[0066] The audit encapsulation unit receives the irreversibility verification pass signal output by the reverse verification unit, receives the policy audit log output by the privacy algorithm fusion scheduling module, binds the irreversibility verification pass signal with the policy audit log to generate a digitally signed audit package, encapsulates the device control instruction and the digitally signed audit package to generate the final control instruction, and the final control instruction carries the irreversibility verification pass signal and the operation audit traceability chain.

[0067] Furthermore, the data fusion intelligent device linkage management and control system based on a large model of the present invention further includes:

[0068] The data feedback loop receives the delay indicator generated by the device executing the control instruction, converts the delay indicator into the dynamic knowledge base weight correction parameter, and transmits the weight correction parameter to the policy update unit of the dynamic knowledge base and the container policy driving engine;

[0069] A control feedback loop receives the confidence level signal output by the parallel simulation environment, dynamically adjusts the arbitrator sensitivity parameters according to the three levels of high confidence, medium confidence, and low confidence, and transmits the sensitivity parameters to the multiple optimization resource modules and the arbitrator;

[0070] In high-confidence scenarios, the weight correction parameter is tightened to the current value of the dynamic knowledge base weight parameter × 80%, and in low-confidence scenarios, the sensitivity parameter is relaxed to the current value of the arbitrator sensitivity parameter × 120%;

[0071] The strategy updating unit updates the knowledge base mapping relationship according to the weight correction parameter, the container strategy driving engine calibrates the noise injection amplitude according to the weight correction parameter, the multiple optimization resource module adjusts the resource allocation strategy according to the sensitivity parameter, and the arbitrator dynamically adjusts the deviation detection response speed according to the sensitivity parameter.

[0072] The present application has the following advantages:

[0073] The present application forms an atomic association of the algorithm scheduling instruction and the privacy policy parameter by binding the identifier transmission mechanism, eliminates the time sequence misalignment problem caused by independent decision in the traditional scheme, the atomic computing graph compilation technology deeply embeds the privacy operation node into the pre-training model computing graph structure to generate an indivisible execution unit, avoids the resource competition and additional overhead caused by independent scheduling of the privacy component, the double closed-loop feedback mechanism optimizes the knowledge base strategy weight through the data feedback loop, controls the feedback loop to dynamically adjust the resource allocation threshold according to the simulation confidence level signal, realizes the adaptive balance of the resource allocation efficiency and the privacy protection strength, the audit packaging unit binds the irreversible proof signal generated by the reverse verification and the full-link operation log to generate a digital signature audit package, so that the device control instruction carries a verifiable traceability chain output, meets the requirements of data desensitization irreversibility and legal compliance, and finally solves the core technical contradiction that the algorithm dynamic scheduling and the privacy security mechanism are difficult to efficiently cooperate in cross-scene data fusion. BRIEF DESCRIPTION OF DRAWINGS

[0074] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed in the embodiments. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor based on the drawings.

[0075] Figure 1 A system architecture diagram of a data fusion intelligent device linkage control system based on a large model is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0076] In order to make the purpose, technical solutions and advantages of the present application clearer, the following will combine the specific embodiments of the present application and the corresponding drawings to clearly and completely describe the technical solutions of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. The following will combine the drawings to describe the technical solutions provided by the embodiments of the present application in detail. In order to better understand the purpose of the present application, the following will further describe the present application in detail.

[0077] Please refer to Figure 1The application provides a large model-based data fusion intelligent device linkage management and control system, which comprises:

[0078] A data acquisition module acquires original data streams of terminal devices and outputs standardized data carrying privacy classification identifiers;

[0079] The data acquisition module comprises a protocol adaptation layer and an identifier injection unit, the protocol adaptation layer is provided with a multi-protocol analysis engine, an ONVIF protocol driver is loaded for a camera terminal to convert video streams into structured frame sequences, a Modbus-TCP parser is enabled for an access controller to generate state event records, and an MQTT converter is activated for an environmental sensor to output time sequence data packets. All converted data are packaged into preprocessed data in a unified JSON format, and device type coding metadata fields are embedded in the data header.

[0080] The identifier injection unit is integrated with a metadata processor and a rule indexer, the metadata processor separates the device type coding from the preprocessed data header, and the rule indexer queries a privacy classification rule library by taking the coding as a key. The privacy classification rule library stores the mapping relationship between device types and PII levels by using an in-memory database, and returns matched privacy classification identifiers through a hash index mechanism. A metadata extender writes identifiers into a new field in the data header to generate standardized data.

[0081] The privacy classification rule library constructs a distributed key-value storage architecture, the rule library allocates storage nodes through a consistent hash algorithm, and dynamically maintains a mapping table of device type coding and privacy levels. Index request triggers real-time node positioning logic, and returns corresponding privacy identifiers in milliseconds. The mapping table supports a hot update mechanism to respond to the policy expansion demand of newly added device types.

[0082] A data processing pipeline forms a logical closed loop, original data streams pass through a protocol identifier to classify device types, corresponding protocol parsers are called to convert data formats, and a device type coding is added to an output port to generate preprocessed data. The preprocessed data is transmitted to the identifier injection unit through a zero-copy channel, a metadata processor extracts device coding to trigger rule library queries, and a metadata extender embeds privacy identifiers to generate standardized data. Finally, structured data integrating device attributes and privacy levels are output to downstream modules.

[0083] A data processing module receives environmental parameters in the standardized data, queries a dynamic knowledge base to output synchronously bound optimization algorithm identifiers and cooperative strategy identifiers;

[0084] The data processing module comprises a hierarchical feature extraction engine and a dynamic knowledge base. The hierarchical feature extraction engine deploys parallel processing channels: the first processing layer performs brightness histogram analysis on the video frame sequence in the standardized data to extract high percentile brightness values to output an illumination intensity scalar; the second processing layer uses a Gaussian mixture model for background modeling, combined with a morphological filtering algorithm to count the number of moving targets per unit area to output a density scalar. The dual-channel processing realizes independent quantification of environmental parameters.

[0085] The feature fusion unit integrates a weighted calculation engine. The feature fusion unit receives the brightness scalar and the density scalar, and queries the dynamic weighting coefficient table according to the privacy classification identifier in the header of the standardized data. The weighting coefficient table stores weight configuration rules corresponding to different privacy levels, for example, the density scalar weight is increased in a high privacy level scenario. The weighted calculation engine performs a linear fusion algorithm to generate a composite feature vector, which comprehensively represents the correlation features of illumination intensity and target density.

[0086] The dynamic knowledge base constructs a strategy mapping matching mechanism. The strategy index layer of the dynamic knowledge base stores the mapping relationship between the composite feature vector and the binding identifier pair, which is fixedly composed of an optimization algorithm identifier and a collaborative strategy identifier. The matching engine uses the k-nearest neighbor algorithm to calculate the Euclidean distance between the input feature vector and the pre-stored vector, and returns the binding identifier pair corresponding to the closest mapping record. The mapping relationship optimizes the matching priority according to the historical performance index weight.

[0087] The strategy updating unit responds to the double feedback signals. The strategy updating unit receives the execution delay index of the data feedback loop, and generates a weight correction coefficient through a linear conversion model; simultaneously analyzes the confidence level classification signal of the control feedback loop, and dynamically adjusts the update frequency parameter. The fusion engine outputs an update instruction to the strategy index layer, triggering the feature vector matching range adjustment and the identifier priority reordering.

[0088] The data processing flow forms a technical closed loop. The standardized data is split into the dual processing layer to generate environmental feature scalars; the feature fusion unit dynamically weights the output composite feature vector according to the privacy identifier; and the dynamic knowledge base returns the binding identifier pair through similarity matching. The whole process realizes the conversion from environmental parameter analysis to strategy decision, and the binding output mechanism ensures the atomic synchronization of algorithm scheduling instructions and privacy control parameters.

[0089] The privacy algorithm fusion scheduling module analyzes the synchronization of the optimization algorithm identifier and the collaborative strategy identifier, compiles the privacy operation corresponding to the collaborative strategy identifier into a subgraph node of the pre-trained algorithm computation graph, generates atomic computation graph data of the fused privacy operation, and outputs the atomic computation graph data and the associated strategy audit log;

[0090] The privacy algorithm fusion scheduling module includes an identifier parsing unit, a node injection engine and a compiling engine. The identifier parsing unit receives a synchronization-bound optimization algorithm identifier and a collaborative strategy identifier, queries a pre-trained model registry for a calculation graph structure corresponding to the optimization algorithm identifier, and locates a feature extraction layer output node coordinate. The collaborative strategy identifier indexes a privacy rule library to obtain a noise amplitude parameter set in real time, and the rule library stores a mapping relationship between the strategy identifier and the perturbation intensity.

[0091] The node injection engine performs a calculation graph rewriting operation. After locating the feature layer output node coordinate, the node injection engine inserts a noise injection node, loads a noise amplitude parameter configuration node attribute, and associates the collaborative strategy identifier with a privacy encryption dynamic library to establish an encryption library runtime calling path to the noise injection node. The operation process preserves the original calculation graph logical structure and only extends a privacy operation node.

[0092] The compiling engine embeds the noise injection node and the associated encryption library calling path into an algorithm calculation graph to generate a calculation subgraph that fuses a privacy operation. The compiling outputs indivisible atomized calculation graph data to eliminate the resource overhead of independent scheduling privacy modules. A strategy audit log is created synchronously to structurally record node positioning coordinates, noise parameter values and encryption library link addresses. The log includes a timestamp and an operation hash value.

[0093] The atomized calculation graph data is transmitted to a multiple optimization resource module for execution, and the strategy audit log is output to a security instruction generation module. The calculation graph data includes complete algorithm instructions and a privacy operation node chain, and the audit log builds an operation process that is traceable. The entire process realizes deep integration of algorithm logic and privacy technology, binds identifier input to ensure scheduling consistency, and atomized compilation eliminates execution timing conflicts.

[0094] The multiple optimization resource module allocates computing resources to execute the atomized calculation graph data and simultaneously drives a parallel simulation environment to pre-play resource schemes. When an arbitrator detects that the deviation between simulation indicators and actual execution indicators exceeds a dynamic threshold for three consecutive times, a double closed-loop feedback module is activated.

[0095] The multiple optimization resource module includes a resource allocation unit, a parallel simulation unit and an arbitration analysis unit. The resource allocation unit deploys a multi-thread scheduling architecture to allocate GPU / CPU resources to execute the atomized calculation graph data based on the real-time load state of a computing resource pool. The resource allocation strategy uses a minimum waiting time first algorithm to dynamically adjust a computing node task queue. The parallel simulation unit copies actual scene input data to an isolated environment. A scheme generator iteratively generates multiple resource allocation schemes using a genetic algorithm, and a timing prediction model evaluates the execution efficiency of each scheme and outputs an optimal scheme number.

[0096] The arbitration analysis unit constructs a double-channel monitoring mechanism, and the arbitration analysis unit deploys an index collector to synchronously acquire simulation indexes output by a parallel simulation environment and running indexes of an actual execution pipeline. A distance calculation engine quantifies a difference value of the double-channel indexes by using a Euclidean distance algorithm, and a sliding window mechanism maintains calculation results of three continuous detection periods. A dynamic threshold comparator compares the distance value with a dynamically updated dynamic threshold, and generates a fuse trigger instruction when the distance value exceeds the threshold for three times in succession.

[0097] The fuse trigger instruction activates a double-closed-loop feedback module, and synchronously sends a fuse signal to a container management unit. The arbitration analysis unit outputs a dynamic threshold adjustment request to a control feedback loop, and triggers reconfiguration of a resource allocation strategy. The whole process realizes real-time verification of resource allocation and actual execution state, a pre-play mechanism improves decision reliability, and deviation detection guarantees system abnormal response capability.

[0098] The security instruction generation module receives an output result of the atomized computing graph data, receives a policy audit log output by the privacy algorithm fusion scheduling module, attaches the policy audit log to the output result to generate to-be-signed data, and executes digital signature on the to-be-signed data to generate a device control instruction;

[0099] The security instruction generation module includes a reverse verification unit and an audit packaging unit. The reverse verification unit receives a feature vector output by the atomized computing graph data, and deploys a lightweight variational autoencoder model to perform feature reconstruction. The reconstruction model generates a reconstructed sample simulating an original data structure through a decoding layer, a deviation analyzer calculates a statistical distribution difference value of the feature vector and the reconstructed data, and quantifies a degree of data irreversibility. A preset critical threshold is used to decide an irreversibility verification pass signal or a fuse trigger signal.

[0100] The audit packaging unit integrates a signal binding engine and a digital signature engine. The signal binding engine receives a verification signal output by the reverse verification unit and a policy audit log, and binds the irreversibility verification pass signal and the node operation record through a timestamp alignment algorithm to generate a structured data packet. The digital signature engine uses an asymmetric encryption algorithm to perform a signature operation on the data packet, and generates an audit packet carrying a digital fingerprint. An instruction packager attaches the device control instruction as a main data body to the digital signature audit packet to generate a final control instruction in a layered packaging structure.

[0101] The reverse verification unit provides a mathematical verification basis for the strength of privacy protection, and the audit packaging unit constructs a verifiable operation traceability chain. The audit attachment area of the final control instruction carries the digital signature audit packet, including an irreversibility verification proof and a whole-link operation node record.

[0102] The double-closed-loop feedback module receives a delay index generated by the device execution control instruction, updates a dynamic knowledge base weight parameter through a data feedback loop, and adjusts a resource allocation threshold according to simulation confidence level grading through a control feedback loop.

[0103] The double-loop feedback module includes a data feedback unit and a control feedback unit. The data feedback unit receives a delay indicator generated by the device executing the control instruction, and a delay data linear mapping is performed by a delay indicator converter to obtain a dynamic knowledge base weight correction parameter. The weight correction parameter represents the strategy execution efficiency, and is transmitted to the strategy updating unit of the dynamic knowledge base and the container strategy driving engine through a message queue. The parameter transmission process is additionally time-stamped to guarantee the time sequence integrity of the data.

[0104] The control feedback unit receives a confidence level signal output by the parallel simulation environment, and a signal classifier divides the signal into high, medium and low confidence levels according to historical accuracy data. A sensitivity adjuster dynamically calculates a sensitivity parameter of the arbitrator: the detection sensitivity is improved in a high confidence level scenario, and the fault tolerance is enhanced in a low confidence level scenario. The sensitivity parameter is synchronously transmitted to the multiple optimization resource module and the arbitrator through a control channel.

[0105] The parameter coupling mechanism realizes dynamic adjustment. The weight correction parameter output by the data feedback unit and the sensitivity parameter output by the control feedback unit form a synergy: the weight correction parameter is tightened to a reference value ratio in a high confidence level scenario, and the sensitivity parameter is relaxed to a reference value ratio in a low confidence level scenario. The parameter coupling mechanism balances the dynamic needs of the system between resource efficiency and privacy strength.

[0106] The execution unit responds to the logic. The strategy updating unit rearranges the mapping relationship priority of the dynamic knowledge base according to the weight correction parameter; the container strategy driving engine synchronously calibrates the disturbance intensity of the noise injection node; the multiple optimization resource module optimizes the calculation resource allocation strategy according to the sensitivity parameter; and the arbitrator dynamically adjusts the deviation detection response frequency. The execution unit response realizes the closed-loop control from the decision layer to the operation layer.

[0107] The data feedback quantifies the actual execution effect, and the control feedback predicts the system reliability state. The double-loop cooperative driving optimizes the strategy decision and adjusts the operation parameter, so that the resource allocation efficiency and the privacy protection strength continuously adapt to the dynamic scene changes, and the overall performance of the system is stable and balanced.

[0108] The data acquisition module outputs standardized data to the data processing module, the data processing module outputs a binding identifier to the privacy algorithm fusion scheduling module, the privacy algorithm fusion scheduling module outputs atomized calculation graph data and strategy audit logs to the multiple optimization resource module and the security instruction generation module, the multiple optimization resource module activates the double-loop feedback module after triggering the arbitrator threshold judgment, the double-loop feedback module outputs the updated dynamic knowledge base weight parameter to the data processing module, and outputs the resource allocation threshold to the multiple optimization resource module, the security instruction generation module outputs the device control instruction to the terminal device, and feeds back the execution delay indicator to the double-loop feedback module.

[0109] The protocol adaptation layer converts the original data stream to generate pre-processed data carrying device type encoding, and the identifier injection unit embeds a matched privacy classification identifier into the data header based on the device type encoding index privacy classification rule library to form standardized data. The standardized data is transmitted to the data processing module through a zero-copy channel to provide structured input containing device attributes and privacy levels.

[0110] The hierarchical feature extraction engine parses the environmental parameters in the standardized data: the first processing layer outputs the illumination intensity scalar, and the second processing layer outputs the target density scalar. The feature fusion unit dynamically weights to generate a composite feature vector according to the privacy classification identifier, and the dynamic knowledge base maps the output binding optimization algorithm identifier and collaborative strategy identifier through the k-nearest neighbor algorithm matching strategy table. The binding identifier pair is transmitted to the privacy algorithm fusion scheduling module through the message queue.

[0111] The identifier parsing unit locates the pre-trained model computation graph node coordinates, the node injection engine inserts disturbance nodes configured with noise parameters and links the encryption library. The compilation engine generates an atomized computation graph data output fused with privacy operations to the multiple optimization resource module; simultaneously creates a strategy audit log recording node parameter configuration, and outputs the log file to the security instruction generation module.

[0112] The resource allocation unit schedules computing resources to execute the atomized computation graph data, and the parallel simulation environment rehearses the resource scheme. The arbitration analysis unit calculates the Euclidean distance between the simulation index and the actual execution index, and when the detection result exceeds the dynamic threshold for three consecutive times, a fuse trigger instruction is generated to activate the double closed-loop feedback module. The arbitrator synchronously sends the fuse signal to the container management unit.

[0113] The data feedback unit converts the device execution delay index into a weight correction parameter, which is transmitted to the dynamic knowledge base strategy updating unit and the container strategy driving engine; the control feedback unit outputs the sensitivity parameter to the multiple optimization resource module and the arbitrator according to the confidence level classification signal. The weight parameter updates the knowledge base strategy priority and calibrates the noise amplitude, and the sensitivity parameter adjusts the resource allocation strategy and the deviation detection response.

[0114] The reverse verification unit quantifies the irreversibility of the feature vector, and the audit packaging unit binds the verification signal and the strategy audit log to generate a digital signature audit package. The final control instruction encapsulates the device operation command and the audit package and outputs it to the terminal device for execution, and the execution delay index is returned to the double closed-loop feedback module through the feedback channel to form a closed-loop control.

[0115] The data acquisition module realizes communication protocol conversion of multiple source heterogeneous devices through a protocol adaptation layer. The protocol adaptation layer has a multi-protocol analysis engine built-in, which converts the original data stream of the terminal device into preprocessed data in a unified data format. The identifier injection unit embeds a privacy classification identifier in the header of the preprocessed data based on the device type coding index privacy classification rule library through a metadata expansion mechanism, and generates standardized data output to the data processing module. This process realizes the conversion of raw data to structured data, providing a unified input carrying privacy attributes for subsequent modules.

[0116] The hierarchical feature extraction engine of the data processing module processes the environmental parameters in parallel: the first processing layer performs brightness histogram analysis on the video frame data and extracts the high percentile brightness value as the illumination intensity quantization indicator; the second processing layer calculates the unit area moving target density value through background modeling and morphological filtering technology. The feature fusion unit dynamically loads the weighting coefficient according to the privacy classification identifier, and fuses the illumination intensity and target density features into a composite feature vector. The dynamic knowledge base maps the preset strategy mapping table through feature vector matching, and outputs the synchronized binding of the optimization algorithm identifier and the collaborative strategy identifier to the privacy algorithm fusion scheduling module, forming the atomized association of algorithm selection and privacy strategy.

[0117] The privacy algorithm fusion scheduling module parses the optimization algorithm identifier to locate the node position of the pre-training model calculation graph, and simultaneously indexes the noise parameter of the privacy rule library associated with the collaborative strategy identifier. After outputting the feature extraction layer node of the calculation graph, a noise injection node configured with noise parameters is inserted, and a privacy encryption dynamic library is dynamically linked. The calculation graph rewriting engine compiles the privacy operation node into a subgraph structure of the algorithm calculation graph, generates atomized calculation graph data with fused privacy operations, and synchronously outputs strategy audit logs containing node positioning coordinates and parameter configuration details.

[0118] The multiple optimization resource module deploys optimization algorithm executors and parallel simulation environment double channels: the executors allocate computing resources to run atomized calculation graph data, while the simulation environment replicates actual scene input data to pre-play resource allocation schemes. The multiple detection arbitrator compares the Euclidean distance between simulation indicators and actual execution indicators in real time, and activates the double closed-loop feedback module when the distance exceeds the dynamic threshold for three consecutive times. This mechanism improves the reliability of resource decision-making through parallel pre-play.

[0119] The reverse verification unit of the security instruction generation module performs reconstruction inversion on the output feature vector, and verifies the data irreversibility through bias analysis. The audit packaging unit binds the irreversibility verification result with the strategy audit log to generate a digital signature audit package, which is attached to the device control instruction to form the final instruction. This process builds a full-link evidence chain of privacy operations.

[0120] The double closed-loop feedback module converts the device execution delay index into a weight correction parameter through a data feedback loop to update the mapping relationship weight of the dynamic knowledge base; and adjusts the resource allocation threshold parameter according to the simulation confidence level classification signal through a control feedback loop. The updated weight parameter returns to the data processing module to optimize strategy matching, and the adjusted threshold parameter is input into the multiple optimization resource module to realize dynamic balance of resource efficiency and privacy strength.

[0121] Specifically, the data fusion intelligent device linkage management and control system based on a large model comprises a protocol adaptation layer and an identifier injection unit in the data acquisition module;

[0122] The protocol adaptation layer converts the terminal device communication protocol into a unified data format and outputs pretreatment data carrying a device type code.

[0123] The identifier injection unit receives the pretreatment data, indexes the privacy classification rule base based on the device type code, and generates matched privacy classification identifier embedded data headers.

[0124] The privacy classification rule base stores the mapping relationship between the device type and the privacy level and dynamically responds to the index request.

[0125] The identifier injection unit outputs the standardized data integrated with the privacy classification identifier to the data processing module.

[0126] The protocol adaptation layer processes the original data stream to generate pretreatment data output to the identifier injection unit.

[0127] The identifier injection unit extracts the device type code in the pretreatment data to call the privacy classification rule base.

[0128] The identifier injection unit embeds the generated privacy classification identifier in the pretreatment data to generate standardized data output to the data processing module.

[0129] The protocol adaptation layer is built-in with a multi-protocol analysis engine, which loads corresponding communication driver components for different terminal device types. The camera device activates the video stream analysis protocol to convert the original video data into structured frame sequences, the environmental sensor enables a lightweight communication protocol conversion module to generate time sequence data packets, and the access controller calls a device state converter to generate structured state records. All processed data is packaged into pretreatment data in a unified format, and a unique device type code is embedded in the data header as metadata identification.

[0130] After the identifier injection unit receives the preprocessed data, the device type code is extracted as an index key to access the privacy classification rule base. The privacy classification rule base stores the mapping relationship between the device type and the privacy level in the memory database, and realizes millisecond-level response through the hash index mechanism. After the rule base returns the matched privacy classification identifier, the identifier injection unit inserts the privacy classification identifier field in the header of the preprocessed data through the metadata expansion mechanism.

[0131] The privacy classification rule base is deployed in a distributed architecture, and stores the mapping relationship table between the device type code and the privacy level. When receiving the index request of the identifier injection unit, the rule base locates the storage node through the consistent hash algorithm, and returns the corresponding privacy level identifier in real time. This process realizes the dynamic mapping of device type to privacy attribute.

[0132] The protocol adaptation layer adopts a pipeline architecture for processing the original data stream. The original data input protocol identifier determines the device type, calls the corresponding protocol parser to convert the data format, and finally adds the device type code to generate preprocessed data. The preprocessed data is transmitted to the identifier injection unit through the zero-copy memory channel, reducing the data transmission overhead.

[0133] The extraction operation of the device type code by the identifier injection unit is realized by the metadata parser. The metadata parser separates the header information of the preprocessed data, reads the device type code, and initiates an index request to the privacy classification rule base. After the rule base returns the privacy classification identifier, the metadata expander writes the identifier into the data header extension field, and generates standardized data output to the downstream module.

[0134] The entire data collection process forms a closed-loop processing chain: the protocol adaptation layer unifies the data format and adds device type metadata, and the identifier injection unit dynamically embeds the privacy attribute identifier based on the metadata. The structured output of the standardized data provides a complete data basis containing device type and privacy level for subsequent processing modules, supporting the execution of differentiated privacy policies.

[0135] Specifically, the data fusion intelligent device linkage control system based on a large model comprises:

[0136] The data processing module is provided with a hierarchical feature extraction engine;

[0137] The first processing layer performs illumination environment analysis: input the video frame in the standardized data, extract the brightness histogram, calculate the 95th percentile brightness value, and output the brightness scalar;

[0138] The second processing layer performs dynamic target density analysis: input the standardized data, use background modeling and morphological filtering, count the number of moving objects per square meter, and output the density scalar;

[0139] The feature fusion unit receives the luminance scalar and the density scalar, queries a dynamic weighting coefficient according to the privacy level identifier, and calculates a composite feature vector;

[0140] The dynamic knowledge base matches the composite feature vector with a pre-stored scene strategy mapping relationship, and outputs a bound optimization algorithm identifier and a cooperative strategy identifier to the privacy algorithm fusion scheduling module;

[0141] The standardized data is input into the first processing layer and the second processing layer;

[0142] The first processing layer outputs the luminance scalar to the feature fusion unit;

[0143] The second processing layer outputs the density scalar to the feature fusion unit;

[0144] The feature fusion unit receives the privacy level identifier to dynamically calculate the weighting value, fuses the luminance scalar and the density scalar to generate a composite feature vector.

[0145] The layered feature extraction engine adopts a parallel processing architecture. The first processing layer performs light analysis on the video frame sequence in the standardized data. This processing layer extracts the histogram distribution features of the luminance channel of the video frame, and quantifies the ambient light intensity by counting the high percentile luminance value. The light analysis result is output as a scalarized luminance indicator, representing the influence degree of the light condition of the current scene on the device linkage.

[0146] The second processing layer performs dynamic target density analysis on the standardized data, and uses a Gaussian mixture model to establish a scene background. By foreground target segmentation and morphological filtering operation, noise interference is eliminated, and moving target objects are accurately identified. The number of effective targets per unit area is counted to generate a density scalar, which reflects the distribution density features of the target objects in the scene.

[0147] The feature fusion unit receives the luminance scalar and the density scalar output by the double processing layer, and retrieves a dynamic weighting coefficient table according to the privacy level identifier in the header of the standardized data. The weighting coefficient table stores the feature weight configuration rules corresponding to different privacy levels, and the feature fusion unit executes a weighted fusion algorithm to generate a composite feature vector. This vector comprehensively represents the correlation features of the ambient light condition and the target distribution density.

[0148] The dynamic knowledge base stores a pre-stored scene strategy mapping table, and uses a feature vector similarity matching mechanism. The mapping table records the binding relationship between the composite feature vector range and the optimization algorithm identifier and the cooperative strategy identifier. The matching engine retrieves the closest vector range to the input feature vector, and outputs the corresponding bound identifier pair to the privacy algorithm fusion scheduling module.

[0149] The standardized data is implemented in the hierarchical feature extraction engine and parallel shunt processing: the video frame data is input into the first processing layer to generate a luminance scalar, and the device state and environment data are input into the second processing layer to generate a density scalar. The double scalar output is transmitted to the feature fusion unit through the data bus, and the unit dynamically loads the weight configuration rule according to the privacy classification identifier.

[0150] The feature fusion unit performs privacy weighted fusion of environmental features: according to the numerical level of the privacy classification identifier, the corresponding weighting coefficient is called to perform linear weighting calculation on the luminance scalar and the density scalar. The output result generates a dimension-compressed composite feature vector, which is input into the dynamic knowledge base as a comprehensive expression of the scene features for strategy matching.

[0151] The whole processing flow forms a closed loop from environmental feature extraction to strategy decision: hierarchical processing realizes environmental parameter quantization, privacy identifier drives feature fusion weighting, and composite feature vector matches the preset strategy to generate a binding identifier. This mechanism realizes the dynamic coupling of environmental features and privacy strategies, and provides a decision basis for subsequent privacy enhancement calculation.

[0152] Specifically, the data fusion intelligent device linkage control system based on a large model provided by the application comprises a dynamic knowledge base, a feature extraction unit, a feature fusion unit, a strategy decision unit and a control feedback loop.

[0153] The strategy index layer stores the mapping relationship between the composite feature vector and the binding optimization algorithm identifier and the cooperative strategy identifier;

[0154] The weight storage layer records the historical performance indicators of each feature vector in the mapping relationship, including response delay and recognition accuracy;

[0155] The strategy update unit receives the execution delay indicator transmitted by the data feedback loop, and converts the delay indicator into a weight correction coefficient;

[0156] The strategy update unit receives the confidence level classification signal transmitted by the control feedback loop, and dynamically adjusts the update frequency according to the high / medium / low three levels;

[0157] The strategy update unit fuses the weight correction coefficient and the update frequency parameter, and outputs an update instruction to the strategy index layer to adjust the mapping relationship;

[0158] The policy updating unit synchronously outputs an update instruction to the weight storage layer to calibrate a historical performance index, and the weight storage layer outputs a latest historical performance index to the policy index layer to optimize real-time matching accuracy. A data feedback loop transmits an execution delay index to the policy updating unit, and a control feedback loop transmits a confidence level classification signal to the policy updating unit. The policy updating unit outputs the update instruction to the policy index layer and the weight storage layer. The weight storage layer transmits the latest historical performance index to the policy index layer, and the policy index layer outputs an optimized binding of an optimization algorithm identifier and a collaborative policy identifier to the privacy algorithm fusion scheduling module.

[0159] The policy index layer adopts a distributed key-value storage architecture and stores a mapping relationship between a composite feature vector hash value and a binding identifier pair. The binding identifier pair includes an optimization algorithm identifier and a collaborative policy identifier, which maintain a fixed association relationship. An index engine retrieves the closest mapping record through a feature vector similarity matching algorithm and outputs the corresponding binding identifier pair to a downstream module.

[0160] The weight storage layer constructs a time series database and continuously records historical execution indexes of each feature vector associated policy. The index data includes response delay distribution statistics and recognition accuracy change curves, and a storage unit retains the latest historical data using a sliding window mechanism. A weight calculator dynamically calculates the comprehensive weight value of the feature vector based on a time decay model to provide priority reference for policy matching.

[0161] The policy updating unit includes a signal converter and an update scheduler. The signal converter linearly maps the execution delay index of the data feedback loop into a weight correction coefficient. The update scheduler analyzes the confidence level classification signal of the control feedback loop and generates an update frequency parameter according to the rule of high confidence level and high frequency update, and low confidence level and low frequency update. The fusion engine synthesizes the weight correction coefficient and the update frequency parameter into an update instruction.

[0162] The update instruction is transmitted to the policy index layer to trigger mapping relationship adjustment. The mapping relationship adjuster reorders the association priority of the feature vector and the identifier according to the weight correction coefficient and synchronously updates the feature vector matching range threshold. The weight storage layer receives the update instruction and starts the index calibrator to recalculate the historical performance index based on the latest weight parameter and outputs the calibrated index data to the policy index layer to optimize the real-time matching algorithm.

[0163] The data feedback loop inputs the device execution delay index to the signal converter of the policy updating unit, and the control feedback loop inputs the confidence level classification signal to the update scheduler. The policy updating unit synchronously sends the update instruction to the policy index layer and the weight storage layer. The weight storage layer outputs the calibrated historical performance index to the policy index layer to drive the index layer to output the optimized binding identifier pair.

[0164] The whole dynamic updating mechanism forms a closed-loop control: the actual performance of the execution delay feedback quantization strategy, and the confidence signal represents the reliability of the simulation rehearsal. The updating instruction drives the mapping relationship and the weight index to optimize the strategy output to adapt to the system running state. The optimized binding identifier ensures the real-time collaboration of algorithm scheduling and privacy policy.

[0165] Specifically, the data fusion intelligent device linkage management and control system based on a large model comprises a privacy algorithm fusion scheduling module.

[0166] The synchronization binding optimization algorithm identifier and the collaborative strategy identifier output by the data processing module are received;

[0167] The optimization algorithm identifier is parsed to locate the feature extraction layer output node position in the pre-trained model calculation graph;

[0168] The noise amplitude parameter corresponding to the collaborative strategy identifier is obtained from the privacy rule library, a noise injection node configured with the noise amplitude parameter is inserted after the located node position, the collaborative strategy identifier is dynamically linked to the privacy encryption dynamic library to the noise injection node, the calculation graph embedded with the noise node and the encryption library is compiled to generate an atomized calculation graph data, a strategy audit log recording the node positioning, noise parameter configuration, and encryption library linking operation details is synchronously generated, and the atomized calculation graph data is output to the multiple optimization resource module, and the strategy audit log is output to the security instruction generation module.

[0169] The privacy algorithm fusion scheduling module receives the synchronization binding identifier pair output by the data processing module, and the identifier pair comprises an optimization algorithm identifier and a collaborative strategy identifier.

[0170] The collaborative strategy identifier is used as an index key to access the privacy rule library, and the rule library returns a noise amplitude parameter configuration set. The node injection engine inserts a noise injection node after the located feature extraction layer output node, and loads the obtained noise amplitude parameter to the node attribute configuration. The dynamic linker associates the collaborative strategy identifier with the privacy encryption dynamic library, and establishes a runtime encryption library calling path to the noise injection node.

[0171] The calculation graph compilation engine performs an atomization rewriting operation: the noise injection node and the associated encryption library calling path are embedded into the original algorithm calculation graph to generate a calculation subgraph with fused privacy operations. The compilation process preserves the original calculation logic to form an indivisible execution unit. A strategy audit log is synchronously generated, and the log records the positioning coordinates of the noise injection node, the noise amplitude parameter value, and the encryption library dynamic linking address.

[0172] The atomized computing graph data is output to a multiple optimization resource module for execution, and the strategy audit log is transmitted to a security instruction generation module. The computing graph data contains complete algorithm instructions and privacy operation nodes, eliminating the overhead of independently scheduling privacy components. The audit log adopts a structured storage format, containing a timestamp, a node identifier, and a parameter hash value, providing a data basis for operation traceability.

[0173] The entire scheduling process realizes deep integration of algorithms and privacy technologies: binding identifiers ensure scheduling consistency, node positioning enables accurate operation insertion, and dynamic linking supports hot loading of privacy components. Atomic compilation ensures the integrity of computing logic, and audit logs build full-link traceability capabilities.

[0174] Specifically, the data fusion intelligent device linkage management and control system based on a large model comprises:

[0175] The atomized computing graph data is encapsulated in an independent container;

[0176] The container integrates a strategy consistency engine inside, which performs the following collaborative operations:

[0177] Load the algorithm executable file of the fusion privacy operation;

[0178] Dynamically analyze the configuration file corresponding to the collaborative strategy identifier to generate noise amplitude and encryption library link parameters in real time;

[0179] Synchronously start the log recording service to continuously capture the running state of the noise injection node and the encryption library;

[0180] The container automatically verifies the matching of the configuration file digital signature and the registration signature of the collaborative strategy identifier when starting;

[0181] In real time, the collaborative strategy identifier is changed to the noise injection node and the encryption library link during runtime.

[0182] The atomized computing graph data is deployed in an independent running environment through containerization encapsulation. The container image generator packages the algorithm executable file of the fusion privacy operation and its dependent libraries into a lightweight virtualization image, ensuring the security of the execution environment through resource isolation mechanism. The container loads the image to initialize the running instance when starting.

[0183] The strategy consistency engine is integrated inside the container, and the configuration file associated with the collaborative strategy identifier is dynamically analyzed during runtime. The configuration file parser extracts the noise amplitude parameter and the encryption library link address, and updates them to the memory parameter pool in real time. The parameter pool is synchronized to the noise injection node and the encryption library calling interface through a shared memory channel.

[0184] The log recording service is automatically activated with the container startup, and adopts a bypass monitoring mode to capture the output data distribution characteristics of the noise injection node. At the same time, the calling frequency and response state of the privacy encryption dynamic library are monitored, and the operation state record containing the timestamp is written into the ring log buffer.

[0185] The security verification process is performed in the container initialization stage: the digital signature verifier compares the digital signature of the to-be-loaded configuration file with the registered signature associated with the collaborative policy identifier. The signature verification module calculates the hash value deviation, and when the deviation exceeds a predetermined threshold, a fuse mechanism is triggered to block the container startup.

[0186] The container running stage responds to policy changes in real time: the event listener captures the collaborative policy identifier update message, and the configuration reload engine parses the parameter set corresponding to the new identifier. The parameter updater updates the noise amplitude parameter and encryption library link address to the running instance on the premise of maintaining calculation continuity.

[0187] The whole container management mechanism realizes the dynamic adaptation of the calculation unit: the policy consistency engine maintains the real-time nature of the operation parameters, the log service provides the visibility of the running state, and the security verification mechanism prevents configuration tampering risks. The hot update capability supports the dynamic adjustment of the privacy policy, and guarantees the robustness of the system in a changing environment.

[0188] Specifically, the data fusion intelligent device linkage management and control system based on a large model provided by the application has a policy driving engine deployed inside a container, and is further used for:

[0189] Loading the algorithm executable files of the fusion noise node and the encryption library, parsing the configuration file corresponding to the collaborative policy identifier, and generating the noise amplitude parameter and the encryption library link instruction;

[0190] Starting the log recording service to capture the noise node output data and the encryption library operation state;

[0191] In the container initialization stage, the hash value deviation of the digital signature of the configuration file and the registered signature of the collaborative policy identifier is verified, and when the deviation value is greater than 5%, a fuse alarm is triggered;

[0192] In the container running stage, the collaborative policy identifier change event is listened to, and the noise amplitude parameter and the encryption library link are reconfigured within 50 milliseconds.

[0193] Periodically verify the integrity of the configuration file, freeze the container and start a backup instance when tampering is detected, and the log recording service outputs the noise indicators and encryption audit records to the security instruction generation module in real time.

[0194] The policy-driven engine loads the algorithm executable file of the fusion noise node and the encryption library, and reads the configuration file associated with the cooperative policy identifier through the configuration parser. The configuration file contains noise amplitude configuration items and encryption library link addresses, and the parser generates a corresponding parameter instruction set and transmits it to the runtime parameter register. The parameter register distributes the instruction set to the noise injection node execution unit and the encryption library call interface through the memory mapping mechanism.

[0195] The log recording service adopts a non-intrusive monitoring architecture, captures the output data distribution characteristics of the noise injection node through a data probe. At the same time, the interface listening module collects the calling state data of the privacy encryption dynamic library, and generates audit records containing operation timestamps and state codes. The log transmitter transmits the structured logs in real time to the downstream module through a dedicated channel.

[0196] The container initialization phase performs a security verification process: the signature verification module calculates the digital signature hash value of the configuration file, and compares it with the registration signature reference value bound to the cooperative policy identifier. The hash deviation analyzer detects the degree of signature deviation, and when the deviation exceeds a predetermined threshold, the fuse signal generator is triggered to block the container startup process and activate the alarm notification.

[0197] The container running phase deploys an event-driven mechanism: the change listener captures cooperative policy identifier update events in real time, and the configuration reload engine parses the parameter set corresponding to the new identifier. The parameter updater replaces the noise amplitude value and encryption library link address of the memory parameter register through atomic operations, and completes the hot update operation while maintaining calculation continuity.

[0198] The integrity check module periodically scans the configuration file storage area, and compares the current hash value of the file with the initial reference value through the hash verifier. When a hash mismatch is detected, the container freezer suspends the current running instance, and the container orchestrator starts a pure backup container to load the original image. The log relay continuously transmits noise index statistical data and encryption operation audit records to the security instruction generation module.

[0199] Specifically, the big model-based data fusion intelligent device linkage management and control system provided by the application further comprises:

[0200] The arbitrator calculates the Euclidean distance between the simulation index and the actual execution index, and when the distance exceeds the dynamic threshold for three consecutive times, the double-loop feedback module is activated, and a fuse trigger signal is sent to the container;

[0201] After receiving the fuse signal, the container freezes the current running instance, starts a pure backup container to load the atomized computation graph data, records the container switching delay time, and feeds back the switching delay index to the double-loop feedback module;

[0202] The control feedback loop adjusts the dynamic threshold based on the simulated confidence level signal. The threshold for high-confidence scenarios is set to 80% of the dynamic threshold baseline value, and the threshold for low-confidence scenarios is set to 120% of the dynamic threshold baseline value.

[0203] The dual closed-loop feedback module calibrates the generation logic of the confidence grading signal based on the container switching delay index. When the switching delay is greater than 50 milliseconds, the simulation confidence grading signal is automatically downgraded by one level.

[0204] The arbitrator deploys a multi-dimensional metric collector to capture, in real time, simulation metrics output by the parallel simulation environment and operational metrics from the actual execution pipeline. The distance calculation engine uses a Euclidean distance algorithm to quantify the difference between the dual-channel metrics and maintains the calculation results for three consecutive detection cycles through a sliding window mechanism. If the distance value exceeds the dynamic threshold three times in a row, the arbitrator generates a circuit breaker trigger instruction, activating the dual closed-loop feedback module and simultaneously sending a circuit breaker signal to the container instance.

[0205] Upon receiving the circuit breaker signal, the container management unit initiates emergency response: the instance freezer immediately suspends the memory state and processes of the currently running container, preventing potential error propagation. The container orchestrator simultaneously launches a clean backup instance, loading the original image of the atomic computation graph data from persistent storage. The switchover process records the time difference between the initial container freeze and the backup container startup completion time, generating an accurate switchover latency metric that is transmitted via a feedback channel to the dual closed-loop feedback module.

[0206] The control feedback loop receives confidence-level signals from the parallel simulation environment. The signal classification logic is based on historical simulation accuracy. The threshold adjuster dynamically adjusts the arbitrator's dynamic threshold based on the confidence level: a tightening strategy reduces the threshold ratio in high-confidence scenarios, while a loosening strategy increases the threshold ratio in low-confidence scenarios. The adjusted threshold ratio is then applied to subsequent anomaly detection decisions.

[0207] The calibration engine of the dual closed-loop feedback module receives container switching delay metrics and uses a time delay analyzer to evaluate switching efficiency. When the delay exceeds a predetermined threshold, a confidence degrader automatically downgrades the simulation confidence level by one level. This downgraded confidence level is then re-inputted into the control feedback loop, forming a closed-loop control chain with adaptive detection accuracy.

[0208] The arbitrator triggers circuit breaking based on indicator distance detection. The container management unit performs hot switching and reports latency data. The confidence signal dynamically degrades based on switching efficiency, and the degradation signal readjusts the detection threshold. This entire process achieves a dynamic balance between abnormal response and detection accuracy, improving the system's fault tolerance under complex operating conditions.

[0209] Specifically, the large model-based data fusion intelligent device linkage management and control system comprises a security instruction generation module, an atomic computing graph data output feature vector is received by the security instruction generation module, a feature reconstruction is performed to generate reconstruction data, and a deviation value of the feature vector and the reconstruction data is calculated;

[0210] The reverse verification unit receives the feature vector output by the atomized computing graph data, performs feature reconstruction to generate reconstruction data, and calculates the deviation value of the feature vector and the reconstruction data;

[0211] When the deviation value is less than 5%, an irreversible verification pass signal is output;

[0212] When the deviation value is greater than or equal to 5%, a fuse trigger signal is output to the container instance of claim 7;

[0213] The audit packaging unit receives the irreversible verification pass signal output by the reverse verification unit and the strategy audit log output by the privacy algorithm fusion scheduling module, binds the irreversible verification pass signal and the strategy audit log to generate a digital signature audit package, packages the device control instruction and the digital signature audit package to generate a final control instruction, and the final control instruction carries the irreversible verification pass signal and the operation audit traceability chain.

[0214] The reverse verification unit receives the feature vector output by the atomized computing graph data, and performs feature reconstruction through a lightweight decoding model. The reconstruction process uses a generative inversion algorithm to simulate the original data generation path to generate reconstruction data samples consistent with the original data structure. A deviation analyzer calculates the statistical distribution difference value of the feature vector and the reconstruction data to quantify the privacy protection strength.

[0215] When the deviation analysis result is lower than the predetermined critical value, the signal generator outputs an irreversible verification pass signal. The signal contains a timestamp and a feature vector hash digest, proving that the data desensitization operation meets the privacy protection requirements. When the deviation value reaches or exceeds the critical threshold, the fuse signal generator outputs a fuse trigger signal, which is transmitted to the container instance of claim 7 through a special channel.

[0216] The audit packaging unit receives the irreversible verification pass signal and the strategy audit log. The signal binding engine aligns the two input data according to the timestamp to generate a binding data package containing the verification result and the operation record. The digital signature module performs asymmetric encryption signature on the binding data package to generate a digital signature audit package. The instruction packager appends the digital signature audit package to the device control instruction to generate a final control instruction.

[0217] The final control instruction adopts a hierarchical packaging structure: the instruction header stores the device control command code, the payload area embeds the device operation parameters, and the audit attachment area carries the digital signature audit package. The audit attachment contains the irreversible verification pass signal and the complete operation audit traceability chain, and the traceability chain records the full-process nodes from feature generation to instruction output.

[0218] The security instruction generation module realizes the verifiability of the privacy protection effect and the auditability of the operation process: the reverse verification quantifies the desensitization strength, the fuse mechanism blocks the privacy leakage path, the digital signature guarantees the integrity of the audit data, and the structured instruction meets the dual needs of device control and compliance supervision. Finally, the control instruction is output to the terminal device for execution, completing the system processing closed loop.

[0219] Specifically, the data fusion intelligent device linkage management and control system based on a large model comprises:

[0220] A data feedback loop receives a delay indicator generated by the execution of the control instruction by the device, converts the delay indicator into a dynamic knowledge base weight correction parameter, and transmits the weight correction parameter to a strategy updating unit and a container strategy driving engine of the dynamic knowledge base;

[0221] A control feedback loop receives a confidence level classification signal output by the parallel simulation environment, dynamically adjusts the sensitivity parameter of the arbitrator according to high, medium and low confidence levels, and transmits the sensitivity parameter to the multiple optimization resource module and the arbitrator;

[0222] In a high confidence level scenario, the weight correction parameter is tightened to 80% of the current value of the dynamic knowledge base weight parameter, and in a low confidence level scenario, the sensitivity parameter is relaxed to 120% of the current value of the arbitrator sensitivity parameter;

[0223] The strategy updating unit updates the knowledge base mapping relationship according to the weight correction parameter, the container strategy driving engine calibrates the noise injection amplitude according to the weight correction parameter, the multiple optimization resource module adjusts the resource allocation strategy according to the sensitivity parameter, and the arbitrator dynamically adjusts the deviation detection response speed according to the sensitivity parameter.

[0224] The data feedback loop receives a delay indicator generated by the execution of the control instruction by the terminal device, and linearly maps the delay data to a weight correction coefficient through an index converter. The coefficient represents the strategy execution efficiency and is transmitted to the strategy updating unit and the container strategy driving engine of the dynamic knowledge base through a data bus. The transmission process uses a message queue with a timestamp to guarantee the time sequence integrity of the data.

[0225] The control feedback loop receives a confidence level classification signal output by the parallel simulation environment, and a signal classifier divides the signal into high, medium and low confidence levels according to preset rules. The sensitivity adjuster dynamically calculates the arbitrator sensitivity parameter according to the confidence level: the detection sensitivity is improved in a high confidence level scenario, and the fault tolerance is enhanced in a low confidence level scenario. The sensitivity parameter is transmitted to the multiple optimization resource module and the arbitrator through the control channel.

[0226] Under the driving of the confidence signal, the weight correction parameter and the sensitivity parameter realize dynamic adaptation: high confidence scene adopts tightening strategy to reduce the relative proportion of the weight correction parameter, and low confidence scene adopts relaxation strategy to improve the relative proportion of the sensitivity parameter. The mechanism balances the accuracy and stability requirements of the system under complex working conditions.

[0227] The strategy updating unit updates the mapping relationship of the dynamic knowledge base according to the weight correction parameter, and adjusts the association weight of the feature vector and the identifier through the priority rearrangement algorithm. The container strategy driving engine synchronously receives the weight correction parameter, and dynamically adjusts the disturbance intensity of the noise injection node through the amplitude calibrator. The multiple optimization resource module optimizes the computing resource allocation strategy according to the sensitivity parameter, and the arbitrator dynamically adjusts the response frequency of the deviation detection according to the sensitivity parameter.

[0228] The double-loop forms a synergistic optimization network: the data feedback quantifies the actual execution effect, and the control feedback predicts the system reliability state. The parameter transmission realizes the real-time linkage of the decision unit and the execution unit, the weight correction optimizes the long-term strategy quality, and the sensitivity adjustment guarantees the short-term execution stability. The system realizes the dynamic balance of resource efficiency and privacy intensity through continuous parameter iteration.

[0229] The application solves the core contradiction in cross-scene data fusion by establishing an integrated synergistic mechanism of algorithm scheduling and privacy protection. The data processing module outputs the synchronous binding optimization algorithm identifier and synergistic strategy identifier in the environment feature analysis stage, forming the atomization association of algorithm selection and privacy strategy. After the binding identifier is transmitted to the privacy algorithm fusion scheduling module, the privacy operation is compiled as a subgraph node of the algorithm computation graph through the computation graph rewriting technology, generating an atomized computation graph data of the fused privacy operation. This process eliminates the scheduling delay caused by the independent operation of the algorithm module and the privacy module in the traditional architecture, realizing the deep integration of computing instructions.

[0230] At the resource scheduling level, the multiple optimization resource module deploys a parallel simulation environment to pre-play the resource scheme, and compares the simulation index and the actual execution index in real time through the arbitrator. When the deviation is continuously detected to be out of standard, the double closed-loop feedback module is activated to trigger the dynamic adjustment mechanism. The data feedback loop converts the device execution delay into the knowledge base weight parameter to optimize the strategy decision accuracy; the control feedback loop adjusts the resource allocation threshold according to the simulation confidence level classification signal to realize the dynamic balance of resource efficiency and privacy intensity. The synergistic effect of the double closed-loop makes the system adapt to scene changes and maintain the best balance of computing efficiency and security protection.

[0231] The privacy security is strengthened by a full-link audit mechanism.The reverse verification unit of the security instruction generation module quantifies the irreversibility of the feature vector, and triggers the container fuse mechanism to block the privacy leakage path when the deviation exceeds the standard.The audit packaging unit binds the irreversibility verification signal with the policy audit log to generate a digital signature audit package, and finally the control instruction carries the audit package and operation trace chain to the terminal device.The design meets the data desensitization irreversibility verification requirement and provides legal and compliant operation evidence.

[0232] Finally, the identifier binding transmission mechanism solves the policy coordination timing problem, the atomic computing graph eliminates the scheduling resource competition, the double closed-loop feedback realizes the dynamic environment self-adaptation, and the audit trace ensures the privacy operation compliance.Four-layer technology cooperation breaks through the limitation of traditional system architecture and achieves efficient unification of algorithm adaptive scheduling and privacy security in cross-scene data fusion.

[0233] The specific embodiments of the present application are as follows:

[0234] The video stream collected by the camera and the access controller state data constitute the original input.The protocol adaptation layer activates the ONVIF protocol parser to convert the video frame into a structured sequence, and simultaneously enables the Modbus-TCP converter to generate access event records.After adding the device type code to the preprocessed data, the identifier injection unit indexes the PII-C1 privacy identifier from the privacy hierarchical rule library, and generates standardized data by embedding the data header through the metadata extension mechanism.

[0235] The standardized data is input into the layered feature extraction engine, the first processing layer extracts the video frame brightness histogram, calculates the ninety-fifth percentile brightness value, and outputs the illumination intensity scalar; the second processing layer uses the Gaussian mixture model to segment the moving target, and counts the effective target number per unit area to output the density scalar.The feature fusion unit loads the dynamic weighting coefficient according to the PII-C1 identifier, fuses the two scalars to generate a composite feature vector, and outputs the bound face recognition algorithm identifier and the differential privacy policy identifier.

[0236] The privacy algorithm fusion scheduling module parses the algorithm identifier to locate the ResNet model feature layer node, indexes the differential privacy policy to obtain the noise parameter, inserts the disturbance node with the configured noise amplitude after the located node, and dynamically links the homomorphic encryption dynamic library.The computing graph rewriting engine generates an atomic computing graph data of fused privacy operation, and synchronously outputs the audit log of the record node coordinates and the encryption parameters.

[0237] The multiple optimization resource module allocates GPU resources to execute the computing graph data, and simultaneously drives the simulation environment to copy the scene input to pre-play the resource scheme.The arbitrator compares the Euclidean distance between the simulation result and the actual face recognition delay in real time, and activates the double closed-loop feedback module when the deviation exceeds the dynamic threshold for three consecutive times.At this time, the arbitrator sends a fuse signal to the container management unit.

[0238] The data feedback loop converts the actual time delay into a weight correction parameter, updating the mapping relationship priority of the dynamic knowledge base; the control feedback loop adjusts the resource allocation threshold according to the simulation confidence level classification signal: the high confidence level scene tightens the threshold to the reference value ratio, and the low confidence level scene relaxes the ratio. The updated parameters optimize the resource allocation strategy and privacy operation intensity in real time.

[0239] The reverse verification unit performs VAE reconstruction on the output feature vector, and generates an irreversibility verification signal when the deviation is below the critical value. The audit packaging unit binds this signal with the policy audit log, and after digital signature, it is attached to the access control instruction. When the deviation exceeds the standard or the arbitrator triggers the fuse, the container management unit freezes the current instance and starts a pure backup container, and the delay index is switched to the double closed-loop module to calibrate the confidence model.

[0240] In the intelligent community access control scene, the application realizes the time sequence dislocation of the binding identifier to eliminate the algorithm scheduling and the privacy policy, atomizes the human face recognition and differential privacy operation of the calculation graph depth, and dynamically balances the recognition efficiency and privacy intensity according to the time delay index and the confidence signal through the double closed-loop feedback. The digital signature audit package provides a full-link operation trace chain, and the implementation process effectively solves the core contradiction described in the background art: when video stream and access data are fused, the privacy protection lag problem caused by algorithm rescheduling. Through the cooperation of the four layers of technology, the face recognition accuracy is guaranteed while maintaining the differential privacy intensity, reducing the response delay to an acceptable range, and eliminating the risk of privacy leakage.

[0241] The protocol adaptation algorithm processes multi-source heterogeneous device data. The multi-protocol analysis engine identifies the terminal device type, and calls the corresponding communication driver to convert the original data stream. The camera video stream activates the ONVIF protocol analysis to generate a structured frame sequence, and the access controller state data triggers the Modbus-TCP converter to generate event records. The converted data is packaged into a unified format preprocessing data, and the device type coding metadata is embedded.

[0242] The hierarchical feature extraction algorithm performs environmental quantitative analysis. The first processing layer uses a histogram statistical algorithm to process the luminance channel of the video frame, calculates the high percentile luminance value, and outputs the illumination intensity scalar. The second processing layer runs a Gaussian mixture model for background modeling, combines a morphological filtering algorithm to segment moving targets, and counts the effective target number per unit area to output the density scalar. The feature fusion algorithm dynamically loads the weighting coefficient according to the privacy classification identifier, linearly weights the two scalars to generate a composite feature vector.

[0243] The dynamic policy matching algorithm realizes decision optimization. The policy index engine adopts a feature vector similarity matching algorithm to retrieve a preset mapping table to output a bound algorithm identifier and a privacy policy identifier. The weight storage algorithm records historical performance indicators and dynamically calculates feature vector weights based on a time decay model. The policy updating algorithm fuses execution delay indicators and confidence signals to generate mapping relationship adjustment instructions and weight calibration instructions.

[0244] The computation graph rewriting algorithm deeply integrates privacy operations. The node positioning algorithm analyzes optimization algorithm identifiers to determine pre-trained model feature layer output node coordinates. The noise injection algorithm inserts a perturbation node of a configuration parameter after positioning the node and dynamically links an encryption library associated with a cooperative policy identifier. The compilation engine embeds privacy nodes into the original computation graph to generate indivisible atomized execution units.

[0245] The double-loop control algorithm dynamically balances system parameters. The data feedback algorithm converts device execution delays into weight correction coefficients to update knowledge base policy priorities. The control feedback algorithm adjusts resource allocation thresholds based on confidence level classification signals: high confidence scenarios tighten threshold benchmark ratios, and low confidence scenarios relax ratios. The sensitivity adjustment algorithm dynamically controls the response frequency of the arbitrator.

[0246] The container hot update algorithm ensures running continuity. The configuration parsing algorithm converts cooperative policy identifiers into corresponding parameters in real time, and atomic operations update memory register values. The signature verification algorithm compares the digital signature of the configuration file with the registered signature hash value, and triggers a fuse when the deviation exceeds the standard. The instance switching algorithm freezes abnormal containers and loads a clean backup image, and records switching delay calibration confidence models.

[0247] The privacy verification algorithm constructs an auditable evidence chain. The feature reconstruction algorithm inverts the original data through a lightweight decoding model, and the statistical distribution difference analysis algorithm quantifies the degree of data irreversibility. The audit binding algorithm aligns verification signals and operation logs by timestamp, and the digital signature algorithm generates tamper-proof audit packages. The instruction packaging algorithm integrates control commands, operation parameters, and audit attachments into a hierarchical instruction structure.

[0248] The pre-trained model serves as the core computing unit. The model library stores multiple pre-trained algorithm models, and the ResNet architecture model is loaded for face recognition tasks, and a lightweight convolutional neural network is used for environment analysis tasks. The model computation graph structure is registered in the system registry, and the feature extraction layer output node position coordinates are pre-defined and stored. The privacy algorithm fusion scheduling module analyzes optimization algorithm identifiers to locate model nodes and provides anchor points for privacy operation insertion.

[0249] A Gaussian mixture model performs background modeling. The second processing layer deploys a multimodal Gaussian distribution model to process video frame sequences, learning the distribution of background features through an expectation-maximization algorithm. A foreground object segmentation algorithm identifies moving objects based on differences in pixel probability distributions, while a morphological filtering model eliminates noise interference and outputs an accurate target density scalar.

[0250] The feature reconstruction model verifies data irreversibility. The secure instruction generation module integrates a variational autoencoder model, taking as input the feature vectors output by the atomized computation graph. The decoder layer attempts to reconstruct the original data structure. The statistical distribution difference between the reconstructed data and the original feature vector quantifies the strength of privacy protection. When the deviation falls below a threshold, a signal indicating irreversibility verification is passed is generated.

[0251] A dynamic knowledge base model optimizes policy matching. The policy indexing layer constructs a composite feature vector space index tree and uses a k-nearest neighbor algorithm to match the most recent policy mapping record. The weight storage layer establishes a time series database model, using a sliding window mechanism to retain the latest historical performance indicators. The policy update model integrates execution delay and confidence signals to generate mapping relationship weight adjustment instructions.

[0252] The privacy rulebase model stores policy parameter configurations. A distributed key-value store records the mapping between collaborative policy identifiers and noise amplitudes. Differential privacy policies store Laplace noise parameters, and k-anonymity policies preserve generalization level configurations. The indexing service uses consistent hashing to quickly locate parameter sets and support real-time configuration of noise injection nodes.

[0253] The simulation environment model previews resource allocation scenarios. The parallel simulation module replicates real-world input data to build a virtualized computing resource pool model. A Monte Carlo simulation algorithm randomly generates multiple resource allocation scenarios. A time series prediction model evaluates the execution efficiency of each scenario and outputs the optimal scenario number to the multi-resource optimization module.

[0254] A confidence grading model quantifies simulation reliability. A control feedback loop integrates a historical data analysis model to calculate the correlation coefficient between simulation results and actual indicators. A hierarchical logic tree model categorizes confidence levels into high, medium, and low based on the correlation coefficient intervals, and outputs signals to adjust the arbitrator's detection sensitivity threshold.

[0255] This invention achieves deep synergy between algorithm scheduling and privacy security through a layered technical architecture. During the data collection phase, a protocol adaptation layer is deployed to convert multi-source device data into a unified format. An identifier injection unit indexes the privacy classification rule base based on device type encoding, generating a privacy classification identifier that is embedded in the data header. This process provides structured, standardized data for subsequent processing, and the privacy classification identifier drives the execution of differentiated protection strategies.

[0256] Environmental feature extraction utilizes a layered processing architecture. The first processing layer analyzes the brightness histogram of video frames to output a scalar light intensity value. The second processing layer uses background modeling to calculate a scalar moving object density. The feature fusion unit dynamically loads weighting coefficients based on the privacy classification identifier, fusing the two environmental features into a composite feature vector. The dynamic knowledge base matches the feature vectors to a pre-set strategy mapping table and outputs a bound optimization algorithm identifier and a collaborative strategy identifier. This binding mechanism atomically links algorithm selection with privacy policy, eliminating the timing misalignment caused by independent decision-making.

[0257] After receiving the binding identifier pair, the privacy algorithm fusion scheduling module parses the optimization algorithm identifier to locate the node position in the pre-trained model computation graph. It then uses the collaborative policy identifier to index the privacy rule base to obtain the noise parameters. It then inserts a noise injection node with the configured parameters after the located node and dynamically links to the privacy encryption dynamic library. The computation graph rewriting engine compiles the privacy operation nodes into algorithm subgraphs, generating atomic computation graph data for the fused privacy operations. A synchronously generated policy audit log records the node coordinates and parameter configurations, providing a foundation for operation traceability.

[0258] Multiple optimized resource modules deploy a dual-channel execution architecture: the optimization algorithm executor allocates resources to run the atomic computation graph, while the parallel simulation environment replicates input data to rehearse resource solutions. The arbitrator calculates the Euclidean distance between simulated and actual metrics, activating the dual closed-loop feedback module when dynamic thresholds are continuously exceeded. The data feedback loop converts device execution delays into weight correction parameters, updating the weights of knowledge base mapping relationships. The control feedback loop adjusts the resource allocation threshold based on simulation confidence level signals, achieving a dynamic balance between resource efficiency and privacy.

[0259] Containerized encapsulation ensures stable operation of computing units. The policy consistency engine dynamically resolves configuration parameters corresponding to collaborative policy identifiers and synchronizes them in real time to the noise injection node and encryption library. Configuration file digital signatures are verified during container initialization, and policy changes are updated dynamically during runtime. A circuit breaker mechanism freezes the container and launches a backup instance if signature deviations exceed the specified limit. The log service continuously outputs operational status to the security instruction generation module.

[0260] Security instruction generation builds a verifiable privacy protection chain. The reverse verification unit reconstructs quantified data irreversibility through feature reconstruction, triggering container circuit breakers when deviations exceed standards. The audit encapsulation unit binds the irreversibility verification signal to the policy audit log, digitally signs and generates an audit package that is attached to the device control instruction. The final instruction carries a full-link operation traceability chain, meeting legal compliance requirements.

[0261] Double-loop feedback forms a global optimization network. Weight correction parameter updates knowledge base strategy priority and calibrates noise injection amplitude, sensitivity parameter adjusts resource allocation strategy and deviation detection response. Parameter transfer mechanism realizes real-time linkage between decision unit and execution unit, making the system continuously adapt to environmental changes and maintaining the best balance between algorithm efficiency and privacy protection in cross-scene data fusion.

Claims

1. A data fusion intelligent equipment linkage management and control system based on a large model, characterized by: include: The data collection module collects the original data stream of the terminal device and outputs standardized data with privacy classification identifiers; The data processing module receives the environmental parameters in the standardized data, queries the dynamic knowledge base and outputs the synchronously bound optimization algorithm identifier and collaborative strategy identifier; The privacy algorithm fusion scheduling module parses the synchronously bound optimization algorithm identifier and collaborative strategy identifier, compiles the privacy operations corresponding to the collaborative strategy identifier into subgraph nodes of the pre-trained algorithm calculation graph, generates atomic calculation graph data of the fused privacy operations, and outputs the atomic calculation graph data and associated policy audit logs. Multiple resource optimization modules allocate computing resources to execute atomic computational graph data while driving a parallel simulation environment to preview resource plans. When the arbitrator detects that the deviation between simulation indicators and actual execution indicators exceeds a dynamic threshold three times in a row, the dual closed-loop feedback module is activated. The security instruction generation module receives the output results of the atomic computation graph data and the policy audit log output by the privacy algorithm fusion scheduling module, appends the policy audit log to the output results to generate data to be signed, and executes the digital signature on the signed data to generate device control instructions; The dual closed-loop feedback module receives the delay indicators generated by the device executing the control instructions, updates the dynamic knowledge base weight parameters through the data feedback loop, and adjusts the resource allocation threshold according to the simulation confidence level through the control feedback loop; The privacy algorithm fusion scheduling module is also used to: Receive the synchronization binding optimization algorithm identifier and the collaboration strategy identifier output by the data processing module; Parse the optimization algorithm identifier to locate the output node position of the feature extraction layer in the pre-trained model calculation graph; The privacy rule base corresponding to the collaborative strategy identifier is indexed to obtain the noise amplitude parameter. A noise injection node with the configured noise amplitude parameter is inserted after the located node position. The collaborative strategy identifier is associated to dynamically link the privacy encryption dynamic library to the noise injection node. The computational graph that embeds the noise node and the encryption library is compiled to generate atomic computational graph data. A policy audit log is simultaneously generated to record the node location, noise parameter configuration, and encryption library link operation details. The atomic computational graph data is output to the multiple optimization resource module, and the policy audit log is output to the security instruction generation module. Atomized computation graph data encapsulation in independent containers; The container integrates a policy consistency engine to perform the following collaborative operations: Load the algorithm executable file that integrates privacy operations; Dynamically parse the configuration file corresponding to the collaborative strategy identifier and generate the noise amplitude and encryption library link parameters in real time; Synchronously start the logging service to continuously capture the running status of the noise injection node and the encryption library; The policy-driven engine is deployed inside the container, loading the algorithm executable file that integrates the noise node and the encryption library, parsing the configuration file corresponding to the collaborative strategy identifier, and generating noise amplitude parameters and encryption library link instructions; Start the logging service to capture the noise node output data and cryptographic library operation status.

2. The data fusion intelligent device linkage management and control system based on large models according to claim 1 is characterized in that: The data acquisition module includes a protocol adaptation layer and an identifier injection unit; The protocol adaptation layer converts the terminal device communication protocol into a unified data format and outputs pre-processed data carrying the device type code; The identifier injection unit receives the pre-processed data, indexes the privacy classification rule base based on the device type encoding, and generates a matching privacy classification identifier to be embedded in the data header; The privacy classification rule base stores the mapping between device types and privacy levels and dynamically responds to indexing requests; The identifier injection unit outputs the standardized data integrating the privacy graded identifier to the data processing module; The protocol adaptation layer processes the original data stream to generate pre-processed data and outputs it to the identifier injection unit; The identifier injection unit extracts the device type code from the pre-processed data and calls the privacy classification rule library; The identifier injection unit embeds the generated privacy-graded identifier into the preprocessed data, generates standardized data and outputs it to the data processing module.

3. The data fusion intelligent device linkage management and control system based on large models according to claim 1 is characterized in that: Also includes: The data processing module is equipped with a hierarchical feature extraction engine; The first processing layer performs lighting environment analysis: it inputs a video frame in normalized data, extracts the brightness histogram, calculates the 95th percentile brightness value, and outputs a brightness scalar; The second processing layer performs dynamic target density analysis: it inputs normalized data, applies background modeling and morphological filtering, counts the number of moving objects per square meter, and outputs a density scalar. The feature fusion unit receives the brightness scalar and the density scalar, queries the dynamic weighting coefficient according to the privacy classification identifier, and calculates the composite feature vector; The dynamic knowledge base matches the composite feature vector with the pre-stored scenario strategy mapping relationship, and outputs the bound optimization algorithm identifier and collaborative strategy identifier to the privacy algorithm fusion scheduling module; Standardized data is input into the first processing layer and the second processing layer; The first processing layer outputs the brightness scalar to the feature fusion unit; The second processing layer outputs the density scalar to the feature fusion unit; The feature fusion unit receives the privacy classification identifier, dynamically calculates the weighted value, and fuses the brightness scalar and the density scalar to generate a composite feature vector.

4. The data fusion intelligent device linkage management and control system based on large models according to claim 3 is characterized in that: The dynamic knowledge base includes a policy index layer, a weight storage layer, and a policy update unit; The strategy index layer stores the mapping relationship between the composite feature vector and the bound optimization algorithm identifier and the collaborative strategy identifier; The weight storage layer records the historical performance indicators of each feature vector in the mapping relationship, including response delay and recognition accuracy; The strategy update unit receives the execution delay index transmitted by the data feedback loop and converts the delay index into a weight correction coefficient; The policy update unit receives the confidence level signal transmitted by the control feedback loop and dynamically adjusts the update frequency according to three levels: high / medium / low; The strategy update unit integrates the weight correction coefficient and the update frequency parameter, and outputs the update instruction to the strategy index layer to adjust the mapping relationship; The policy update unit synchronously outputs update instructions to the weight storage layer to calibrate historical performance indicators. The weight storage layer outputs the latest historical performance indicators to the policy index layer to optimize real-time matching accuracy. The data feedback loop transmits the execution delay indicator input to the policy update unit. The control feedback loop transmits the confidence grading signal input to the policy update unit. The policy update unit outputs update instructions to the policy index layer. The policy update unit outputs update instructions to the weight storage layer. The weight storage layer transmits the latest historical performance indicators to the policy index layer. The policy index layer outputs the optimized bound optimization algorithm identifier and collaborative strategy identifier to the privacy algorithm fusion scheduling module.

5. The data fusion intelligent device linkage management and control system based on large models according to claim 1 is characterized in that: Also includes: Automatically verify the matching between the configuration file digital signature and the registration signature of the collaborative policy identifier when the container is started; Synchronize the collaborative strategy identifier changes to the noise injection node and encryption library link in real time during runtime.

6. The data fusion intelligent device linkage management and control system based on large models according to claim 1 is characterized in that: The container internal deployment policy-driven engine is also used for: During the container initialization phase, the hash value deviation between the configuration file digital signature and the registration signature of the collaborative policy identifier is verified. If the deviation is greater than 5%, a circuit breaker alarm is triggered. During the container operation phase, monitor the collaborative policy identifier change event and reconfigure the noise amplitude parameters and the encryption library link within ≤50 milliseconds; The integrity of the configuration file is periodically verified. When tampering is detected, the container is frozen and a backup instance is started. The logging service outputs noise indicators and encrypted audit records to the security instruction generation module in real time.

7. The data fusion intelligent device linkage management and control system based on large models according to claim 6 is characterized in that: Also includes: The arbitrator calculates the Euclidean distance between the simulation indicator and the actual execution indicator. When the distance exceeds the dynamic threshold three times in a row, the dual closed-loop feedback module is activated and a fuse trigger signal is sent to the container. After receiving the circuit breaker signal, the container freezes the currently running instance, starts a clean backup container to load the atomic computation graph data, records the container switching delay time, and feeds the switching delay indicator back to the dual closed-loop feedback module. The control feedback loop adjusts the dynamic threshold based on the simulated confidence level signal. The threshold for high-confidence scenarios is set to 80% of the dynamic threshold baseline value, and the threshold for low-confidence scenarios is set to 120% of the dynamic threshold baseline value. The dual closed-loop feedback module calibrates the generation logic of the confidence grading signal based on the container switching delay index. When the switching delay is greater than 50 milliseconds, the simulation confidence grading signal is automatically downgraded by one level.

8. The data fusion intelligent device linkage management and control system based on large models according to claim 1 is characterized in that: The security instruction generation module includes a reverse verification unit and an audit encapsulation unit; The reverse verification unit receives the feature vector output by the atomic computation graph data, performs feature reconstruction to generate reconstructed data, and calculates the deviation between the feature vector and the reconstructed data; When the deviation value is less than 5%, the irreversibility verification pass signal is output; When the deviation value is ≥5%, a fuse trigger signal is output to the container of claim 1; The audit encapsulation unit receives the irreversibility verification pass signal output by the reverse verification unit, receives the policy audit log output by the privacy algorithm fusion scheduling module, binds the irreversibility verification pass signal with the policy audit log to generate a digitally signed audit package, encapsulates the device control instruction and the digitally signed audit package to generate the final control instruction, and the final control instruction carries the irreversibility verification pass signal and the operation audit traceability chain.

9. The data fusion intelligent device linkage management and control system based on large models according to claim 1 is characterized in that: Also includes: The data feedback loop receives the delay indicator generated by the device executing the control instruction, converts the delay indicator into the dynamic knowledge base weight correction parameter, and transmits the weight correction parameter to the policy update unit of the dynamic knowledge base and the container policy driving engine; A control feedback loop receives the confidence level signal output by the parallel simulation environment, dynamically adjusts the arbitrator sensitivity parameters according to the three levels of high confidence, medium confidence, and low confidence, and transmits the sensitivity parameters to the multiple optimization resource modules and the arbitrator; In high-confidence scenarios, the weight correction parameter is tightened to the current value of the dynamic knowledge base weight parameter × 80%, and in low-confidence scenarios, the sensitivity parameter is relaxed to the current value of the arbitrator sensitivity parameter × 120%; The policy update unit updates the knowledge base mapping relationship based on the weight correction parameters. The container policy driving engine calibrates the noise injection amplitude based on the weight correction parameters. The multiple optimization resource module adjusts the resource allocation strategy based on the sensitivity parameters. The arbitrator dynamically adjusts the deviation detection response speed based on the sensitivity parameters.

Images