An AI interaction terminal control method and system based on a cloud platform

Abstract

The application discloses an AI interaction terminal control method and system based on a cloud platform, belongs to the technical field of artificial intelligence, and collects state parameter sets of multiple AI interaction terminals and uploads the state parameter sets to a cloud platform; a modeling engine containing behavior mode recognition and state coupling analysis is constructed based on the state parameters, and a multi-terminal coupling perception model is generated; terminal priorities are calculated according to the model, an asymmetric task mapping matrix is constructed, and a control strategy is generated; heterogeneous control instructions are issued to the terminals and execution feedback is collected; the model and the strategy are self-learned and updated in combination with original state parameters and feedback data; if a terminal fault is detected, task migration is performed based on coupling weights and redundancy, and a control matrix is updated; finally, cross-terminal linkage and control strategy self-evolution are realized through S-F-C closed-loop iteration; and the application realizes multi-terminal collaborative perception, self-adaptive control and task robust migration, and improves the intelligence and stability of an AI control system.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and specifically to an AI interactive terminal control method and system based on a cloud platform. Background Technology

[0002] Existing AI interactive terminals typically rely on local hardware to perform functions such as speech recognition, image processing, and natural language understanding. Some systems use cloud platforms for unified instruction distribution and data feedback, but they mainly respond with static rules or preset models, lacking the ability to adapt to real-time interactive behavior and the dynamic state of multiple terminals.

[0003] In addition, most current cloud control technologies adopt a synchronous task distribution model, which makes it difficult to deeply correlate and model the operating environment, user behavior and task load of multiple terminals. This results in sluggish response and chaotic priorities in multi-task concurrent scenarios, making it impossible to form an efficient and personalized interactive experience.

[0004] More importantly, existing systems generally lack a closed-loop evolution mechanism between terminal status and cloud platform. In other words, the cloud cannot proactively update control strategies based on terminal execution results, resulting in simplistic control commands that are difficult to adapt to users' changing needs.

[0005] Therefore, there is an urgent need to construct an AI interactive terminal control method that integrates multi-dimensional real-time perception feedback, cross-terminal asymmetric control, and control parameter self-learning mechanism to achieve multi-terminal intelligent coupling and self-evolutionary control centered on user behavior, thereby breaking through the technical bottlenecks of static control logic, fragmented response, and rigid strategies in existing AI terminal control systems. Summary of the Invention

[0006] The purpose of this invention is to provide an AI interactive terminal control method and system based on a cloud platform to address the shortcomings in the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a cloud platform-based AI interactive terminal control method, comprising:

[0008] S100: Collects a multi-dimensional set of state parameters S from multiple AI interactive terminals and uploads S to the cloud platform;

[0009] S200. Based on the above S, establish an interactive modeling engine in the cloud platform that includes a behavior pattern recognition submodule and a state coupling submodule to generate a multi-terminal coupled perception model M.

[0010] S300. Calculate the adaptive priority P of each AI interactive terminal in the current task sequence through M, and construct the control task allocation matrix T based on the asymmetric task mapping strategy.

[0011] S400. A control strategy C is generated based on the matrix T and the priority P, and heterogeneous control instructions are distributed from the cloud platform to the corresponding AI terminal.

[0012] S500: During the execution of instructions at the terminal, feedback data F is collected in real time, and F and the original state parameter S are input into the self-learning module to perform adaptive updates of the interaction strategy, thereby obtaining the updated perception model M′ and control strategy C′.

[0013] S600. If an abnormal terminal status or task execution failure is detected, the task is migrated to the optimal candidate terminal based on the terminal association weight and redundancy evaluation in the coupling model M′, and the control matrix T is updated.

[0014] S700 continuously performs SFC closed-loop iteration until the execution of tasks on each terminal is stable, thus completing cross-terminal linkage and self-evolution control.

[0015] Preferably, in S300;

[0016] Extract the state coupling factor μ(i,j,t) between each terminal in model M and the user behavior feature vector B_i to form a multidimensional coupling relationship graph G; where μ(i, j, t) is a real number representing the state coupling strength or degree of cooperation between terminal i and terminal j at time t;

[0017] A region partitioning algorithm based on graph traversal heat gradient is performed on graph G to aggregate terminals with similar behavioral trajectories and response delays into response subdomains;

[0018] Within each subdomain, a sliding window asynchronous weighted integral function is used to integrally score the terminal's load change rate and task semantic matching degree, generating a local dynamic priority P_i.

[0019] Based on the global coupling degree and each P_i, a polynomial offset task mapping function is constructed to realize the asymmetric allocation of tasks among terminals, forming a control task matrix T.

[0020] Preferably, in S300;

[0021] Based on the task allocation matrix T and priority vector P, a task-terminal mapping graph is constructed, and the constraint graph coloring algorithm is used to apply execution exclusion rules to the same type of control tasks. The set of instruction scheduling paths with no conflict and optimal resource load among multiple terminals is calculated.

[0022] For each terminal, the instruction semantic skeleton that matches its device type, interaction interface and current context is called from the cloud instruction template library, and a complete heterogeneous control instruction set C_i is generated through a dynamic template splicing mechanism;

[0023] After C_i is generated, the local semantic rewriting engine under the edge node is executed to perform instruction semantic completion and multi-language translation on the control scene.

[0024] Ultimately, the control policy is broadcast to both the terminal and its alternative agent devices simultaneously via a dual-channel push mechanism.

[0025] Preferably, in S500;

[0026] The terminal continuously records feedback data F during the instruction execution cycle, including task response latency, user interaction behavior trajectory, actual execution results and local exception codes;

[0027] The feedback data F and the initial state parameters S are input into the cloud self-learning module, the analysis engine based on the causal dependency path reconstruction mechanism is called, the abnormal feedback source is identified, a multi-factor causal chain graph is constructed in chronological order, and the structural weights of the model M are adjusted through the coupled tensor transformation function with memory residual update, and the updated model M′ is output.

[0028] Based on M′, the policy disturbance analyzer is invoked to simulate the impact of different policy paths on system performance. The original control policy C is dynamically optimized through the policy response tensor mapping method to generate a new policy C′.

[0029] Preferably, in S600;

[0030] By using the instruction receipt monitoring mechanism and feedback event frame analysis, it can be determined whether the terminal is in a state of execution failure, abnormal interruption, or resource unavailability, and it is marked as a fault node.

[0031] Extract the state coupling weights and behavioral redundancy vectors between the faulty terminal and other terminals from the updated perception model M′, call the redundancy scoring function, and sort the candidate terminals that have semantic similarity, perception similarity, and load-bearing capacity.

[0032] Path optimization based on dynamic transition graph is performed among candidate terminals. The terminal with the lowest coupling cost is selected to take over the task. At the same time, the structure mapping update module is called to reconstruct the control task allocation matrix T.

[0033] Preferably, the evaluation of candidate terminal redundancy includes:

[0034] Extract the state coupling tensor μ′(i,j,t) between the faulty terminal and the normal terminal and the task behavior instruction history vectors H_i and H_j to construct a terminal redundancy candidate set R;

[0035] Semantic similarity is calculated using the vector semantic offset function ΔSem(H_i, H_j), perceptual coupling is calculated using the collaborative perceptual vector matching rate Sim_s(Sensor_i, Sensor_j), and resource carrying capacity is evaluated using the combined load trend fitting function P_Load(j,t).

[0036] The three-dimensional indicators are normalized and input into the redundancy scoring function. In the formula, w1 is the semantic adaptability weight; w2 is the perception collaboration weight; w3 is the resource carrying capacity weight; sort all terminals in R in descending order of RScore and output the optimal candidate takeover sequence.

[0037] The present invention also provides an AI interactive terminal control system based on a cloud platform, comprising:

[0038] The perception and acquisition module collects a multi-dimensional set of state parameters S from multiple AI interactive terminals and uploads S to the cloud platform.

[0039] The model building module, based on the aforementioned S, establishes an interactive modeling engine in the cloud platform, including a behavior pattern recognition submodule and a state coupling submodule, to generate a multi-terminal coupled perception model M.

[0040] The priority evaluation module calculates the adaptive priority P of each AI interactive terminal in the current task sequence through the M, and constructs a control task allocation matrix T based on an asymmetric task mapping strategy.

[0041] The control strategy generation module generates a control strategy C based on the matrix T and the priority P, and the cloud platform distributes heterogeneous control instructions to the corresponding AI terminal.

[0042] The self-learning optimization module collects feedback data F in real time during the execution of instructions at the terminal, and inputs F and the original state parameter S into the self-learning module to perform adaptive updates of the interaction strategy, thereby obtaining the updated perception model M′ and control strategy C′.

[0043] If the control structure adjustment module detects an abnormal terminal status or task execution failure, it will migrate the task to the optimal candidate terminal based on the terminal association weight and redundancy evaluation in the coupling model M′, and update the control matrix T.

[0044] The closed-loop control module continuously iterates through the SFC closed-loop until the task execution of each terminal is stable, thus completing cross-terminal linkage and self-evolution control.

[0045] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0046] 1. This invention constructs a cloud-based interactive modeling engine that integrates multi-dimensional terminal state perception, behavior pattern recognition, and state-coupled computation. This enables cross-terminal collaborative modeling and adaptive generation of control strategies, effectively solving the problems of rigid response logic, low task scheduling efficiency, and poor device collaboration capabilities in existing AI interactive terminals. The introduced asymmetric task mapping mechanism and dynamic priority evaluation method can achieve personalized control allocation based on real-time load, task semantics, and user behavior characteristics, improving the overall control accuracy and response efficiency of the system.

[0047] 2. This invention achieves dynamic optimization and updating of the perception model and control strategy through a feedback-driven self-learning module and a closed-loop SFC iterative control mechanism, exhibiting good adaptability and robustness. The fault detection and redundancy migration mechanism, based on terminal coupling weights and a redundancy scoring function, enables seamless task migration, enhancing the system's fault tolerance and continuous interactive experience. The overall solution is applicable to various types of heterogeneous AI terminal environments, possessing significant engineering practical value and system intelligent evolution capabilities. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0049] Figure 1 This is a mind map of the method of the present invention.

[0050] Figure 2 This is a mind map of the system modules of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Example 1, please refer to Figure 1 As shown in this embodiment, an AI interactive terminal control method based on a cloud platform includes:

[0053] S100: Collects a multi-dimensional set of state parameters S from multiple AI interactive terminals and uploads S to the cloud platform;

[0054] S200. Based on the above S, establish an interactive modeling engine in the cloud platform that includes a behavior pattern recognition submodule and a state coupling submodule to generate a multi-terminal coupled perception model M.

[0055] S300. Calculate the adaptive priority P of each AI interactive terminal in the current task sequence through the M, and construct a control task allocation matrix T based on an asymmetric task mapping strategy.

[0056] S400. A control strategy C is generated based on the matrix T and the priority P, and heterogeneous control instructions are distributed from the cloud platform to the corresponding AI terminal.

[0057] S500: During the execution of instructions at the terminal, feedback data F is collected in real time, and F and the original state parameter S are input into the self-learning module to perform adaptive updates of the interaction strategy, thereby obtaining the updated perception model M′ and control strategy C′.

[0058] S600. If an abnormal terminal status or task execution failure is detected, the task is migrated to the optimal candidate terminal based on the terminal association weight and redundancy evaluation in the coupling model M′, and the control matrix T is updated.

[0059] S700 continuously performs SFC closed-loop iteration until the execution of tasks on each terminal is stable, thus completing cross-terminal linkage and self-evolution control.

[0060] S100: Collect a multi-dimensional set of state parameters S from multiple AI interactive terminals, including but not limited to device type, task context, real-time load, sensor data, and user behavior trajectory. The "multiple AI interactive terminals" include hardware devices with artificial intelligence modules, sensing interfaces, and communication functions. The terminals can be voice assistants, intelligent robots, in-vehicle voice control systems, smart wearable devices, smart surveillance cameras, smart home control screens, industrial human-machine interaction devices, etc.

[0061] The device type is used to identify the hardware category, operating system type, processing power indicators, interaction interface form (voice, touch, image, sensor, etc.), power supply mode, etc. of each AI interaction terminal, and is marked as Type_i.

[0062] Task context is used to record the current task situation of the AI ​​terminal, including:

[0063] The currently running AI task identifier (such as speech recognition, image recognition, command issuance, etc.);

[0064] The natural language content entered by the user; the history of multi-turn interactions; the task collaboration dependencies with other terminals; this set of context parameters is denoted as Ctx_i.

[0065] Real-time load is used to measure the current system resource usage status of the AI ​​terminal, including: CPU utilization; memory utilization; network bandwidth utilization; thread running status; battery level (for mobile devices); load status is used for dynamic scheduling and is denoted as Load_i.

[0066] In addition, sensor data includes, but is not limited to, environmental sensor data accessed by the terminal, such as temperature, humidity, light intensity, sound intensity, acceleration, GPS location information, face orientation, microphone array directionality data, etc., tagged as Sensor_i.

[0067] User behavior trajectories are extracted based on historical interaction data between the terminal and the user, including: the user's operation frequency and dwell time on the current terminal; daily usage time preferences; command input language type; response preferences for various AI tasks (e.g., preference for voice broadcast or image presentation); switching paths and latency features with other terminals; these behavior pattern parameters are used to train the user feature profile model, denoted as Profile_i.

[0068] The five parameters {Type_i, Ctx_i, Load_i, Sensor_i, Profile_i} mentioned above together constitute the state vector S_i of the terminal. The set of states of all terminals forms a multi-dimensional state matrix: S = {S1, S2, ..., S_n}, where n is the number of connected AI interaction terminals. Each S_i is compressed and preprocessed through local middleware or edge computing nodes, and then uploaded to the cloud platform through an encrypted channel for subsequent coupled calculations and policy generation by the modeling engine.

[0069] S200. Based on the above S, establish an interactive modeling engine in the cloud platform that includes a behavior pattern recognition submodule and a state coupling submodule to generate a multi-terminal coupled perception model M.

[0070] The interactive modeling engine is deployed in a distributed resource pool on a cloud platform and has heterogeneous data processing, low-latency modeling, and scalable inference capabilities. It includes at least the following two core sub-modules:

[0071] The Behavioral Pattern Recognition Submodule (BPR) is used for cross-device behavioral consistency modeling and temporal recognition of behavioral patterns based on uploaded multi-terminal user behavior trajectories (Profile_i).

[0072] The model is built using a pseudo-graph temporal state map structure:

[0073] Each terminal's user behavior vector is represented as a node in the graph;

[0074] The edges between nodes are defined by "interaction path frequency + terminal switching latency + semantic proximity".

[0075] We construct a weighted graph structure and introduce a multi-window adaptive step mechanism to capture short-term anomalous behavior and long-term usage inertia.

[0076] Ultimately, a multi-layered nested user interaction state graph is formed, which is used to output the user-terminal behavior coupling feature vector B_i.

[0077] State Coupling Submodule (SCS):

[0078] Used to analyze the real-time interaction (coupling degree) between different terminal states, that is, to identify how a change in the state of a certain terminal affects the response or performance of other terminals;

[0079] Introducing State Relational Vector Flow Mapping (SVR) technology:

[0080] For each pair of terminals i and j, calculate their state interdependence weights in dimensions such as task context, alternating load peaks, and sensor collaborative input;

[0081] The dynamic migration factor μ(i,j,t) = f(ΔLoad_ij,ΔCtx_ij, CoSensor_ij,RtBehavior_ij) is used; where f is a multinomial regression model, CoSensor_ij represents the time window coordination rate of sensor data between the two terminals, RtBehavior_ij represents the probability of behavioral response synchronization; ΔLoad_ij represents the difference in terminal load change, and ΔCtx_ij represents the task context difference, expressed as: In the formula, Ctx_i represents the context vector of the current task of terminal i, which is typically encoded by the following elements:

[0082] Current task type (voice broadcast, image recognition, control commands, etc.); semantic vector of user input content; aggregated features of multi-turn interaction history; scene tags (such as "indoor security", "vehicle control", etc.).

[0083] cosine_similarity indicates that the angle between two context vectors in the semantic space is evaluated using the standard cosine similarity method. The closer the result is to 1, the more consistent the semantics of the tasks are.

[0084] The coupling relationship is reduced to a shared sensing space by using the sparse tensor field encoding method.

[0085] The model M is constructed by the user behavior coupling feature B_i output by BPR and the state coupling tensor C_ij output by SCS: M = f(B, C); M is a control graph structure represented by graph neural network, which supports the following during task scheduling: real-time reconstruction of node priorities; adaptive task allocation strategy changes; and prediction of the cost of task migration path (e.g., response delay, cognitive interruption risk).

[0086] S300. Calculate the adaptive priority P of each AI interactive terminal in the current task sequence using the multi-terminal coupled sensing model M, and construct a control task allocation matrix T based on an asymmetric task mapping strategy. The specific process includes the following operations:

[0087] Extract the state coupling factor μ(i,j,t) and behavioral feature vector B_i of each terminal node from model M; construct the coupling control graph G(V,E) between terminals, where: node V represents each AI interaction terminal; the weight of edge E is the corresponding μ(i,j,t), representing the state interaction strength; the graph is a weighted heterogeneous graph that supports dynamic adjustment of connection strength and node state labels.

[0088] Execute the graph traversal heat gradient algorithm in graph G to calculate the propagation energy of each node's behavior in the interaction chain;

[0089] The graph traversal heat gradient method is the core strategy, and its steps include:

[0090] The states of multiple terminals and task nodes are mapped to a heterogeneous graph, and the edge weights between nodes are determined by the degree of coupling between states and the similarity between instructions.

[0091] Simulate the task propagation trajectory using a heatmap algorithm (e.g., a PageRank variant);

[0092] High-frequency path clustering regions are formed in the propagation graph, and the regions are divided based on the node activity gradient.

[0093] Set termination conditions (such as node activity convergence or task transfer stabilization) to complete the partitioning process.

[0094] Nodes with similar behavioral response characteristics (such as response delay distribution and trigger path similarity) are divided into response subdomains R1, R2, ..., R_k; each response subdomain corresponds to a type of task response tendency, such as voice broadcast type, image presentation type, combined command chain type, etc.

[0095] For each terminal within a response subdomain, perform a sliding window asynchronous integral evaluation. The evaluation elements include:

[0096] ΔLoad(t): The load change trend over a past period of time;

[0097] Ctx_match: Similarity between the semantic vector of the current task and the historical tasks of the terminal;

[0098] μ_mean: The average coupling degree of a terminal to other terminals;

[0099] The local priority P_i is generated using the following function, expressed as: P_i = α1·(1 − ΔLoad_i) +α2·Ctx_match_i + α3·μ_mean_i; where α1, α2, and α3 are policy weights, which can be dynamically optimized by controlling historical data. The weights are manually set based on the complexity of the task scenario and user preferences during the initial stage of system operation, for example, initial values ​​α1 = 0.4, α2 = 0.3, and α3 = 0.3. If the terminal is the response center node within the subdomain, an additional coordination guidance factor β is assigned for correction.

[0100] Define a matrix T = [t_ij], where t_ij represents the adaptation weight assigned to terminal i by task j. Introduce a polynomial offset asymmetric mapping function: t_ij = P_i × (1 + γ·ΔCtx(i,j)) × (1 − κ·Load_penalty_i); where ΔCtx(i,j) is the contextual difference between task j and terminal i; Load_penalty_i is the current load penalty function of the terminal; γ and κ are policy parameters used to control the influence of semantic offset and resource regulation; for example, γ = 0.7, κ = 0.3; T is not a symmetric matrix, allowing the same task to be assigned to multiple terminals with different weights, or to be empty. The initial values ​​can be set based on historical task data simulation, and can be dynamically adjusted during operation using moving average or Bayesian optimization.

[0101] Example of dynamic optimization: If the failure rate of task migration is > 5% for three consecutive rounds, increase γ and decrease κ to prioritize task semantic adaptation.

[0102] Load penalty trigger conditions: When CPU utilization > 80% and lasts for more than 5 seconds, or memory utilization > 85%, Load_penalty_i is automatically set to 1 (high-risk state), and the terminal is downgraded to low priority.

[0103] S400. A control strategy C is generated based on the matrix T and the priority vector P, and heterogeneous control instructions are distributed from the cloud platform to the corresponding AI terminal. The specific steps include the following:

[0104] Based on the control task allocation matrix T (representing the adaptation weight between tasks and terminals) and the terminal priority vector P, construct the task-terminal bipartite graph G_task;

[0105] In this graph, the set of nodes includes all currently pending tasks and terminal devices, with edge weights of t_ij×P_i, reflecting the terminal's willingness to adapt to and respond to tasks;

[0106] An improved graph coloring constraint algorithm is used to map conflicting or resource-overlapping tasks to different scheduling intervals.

[0107] Constraints include: concurrent execution of the same resource, avoidance of device context conflicts, and penalties for frequent interruptions;

[0108] Finally, the priority of task instructions, parallel combination method and target terminal set are determined to form a preliminary control strategy C′.

[0109] Each task instruction in control strategy C′ needs to be structured and translated according to the terminal hardware platform, operating system, interface protocol, and user language preferences.

[0110] A Semantic Command Skeleton Library is introduced, which stores the control behavior metastructures (such as "wake-up → recognition → response → execution") corresponding to each terminal type.

[0111] The cloud platform invokes this skeleton and, in conjunction with the terminal's current context (such as the current task mode and scene label), generates complete control instructions C_i through a dynamic template assembly engine.

[0112] It includes control intent identifiers, interaction modality parameters (voice / image / touch), execution time window, language encoding format, etc.

[0113] For terminals that support multimodal operation, a mode switching trigger is also embedded.

[0114] After the instruction structure is assembled, in order to ensure context consistency and language localization adaptability during actual execution, the instruction will be passed to the edge semantic rewriting module.

[0115] Based on the local knowledge graph and end-user configuration, this module fine-tunes the control semantics, such as adjusting the order of instructions and inserting commonly used local statements.

[0116] It can also perform cross-language translation, such as converting "Turn off the screen" into Chinese instructions or natural speech stream;

[0117] Ultimately, the cloud platform issues commands through a dual-channel push mechanism:

[0118] Main channel: Directly sends execution commands to the target terminal;

[0119] Backup channel: Sends replica instructions to the local area network edge agent node or mirror terminal where it is located, supporting failover and hot migration.

[0120] During the execution of control commands by the AI ​​terminal, the S500 system constructs a closed-loop architecture of "state perception – feedback transmission – policy evolution" to complete the real-time adaptive update of the perception model M and the control policy C. This includes the following key processes:

[0121] Each AI interactive terminal embeds a lightweight instruction execution listening module (Listener Agent) that automatically records the execution behavior after each execution of control policy C_i;

[0122] The collected data includes, but is not limited to: instruction execution start and end times, execution delay; user behavior response indicators, such as voice interruption, secondary operation, and repeated triggering; actual action execution status codes (such as success / failure, abnormal interruption); local resource consumption status (CPU, memory, sensor usage); environmental change data (such as noise intensity, light changes, user position offset, etc.); the collected feedback events are organized into structured event frames, including timestamps, terminal identifiers, task IDs, status codes, and data vectors; event frames are pushed to the cloud platform in real time via MQTT or CoAP protocols, supporting breakpoint resumption and time window cached upload.

[0123] The cloud-based self-learning module receives feedback data F and original state parameters S, forming a joint input pair (S, F), where S is the terminal's perceived state before task execution and F is the feedback information after execution.

[0124] The system builds a cause-effect graph G_cause for task execution, tracing each intermediate node (such as network status, load changes, and response time) between command input and user response.

[0125] The "task causal tracing algorithm" is used to identify the minimum set of causal paths P_c that lead to abnormal results (such as failure / delay / user dissatisfaction);

[0126] By utilizing the changes in node weights along each path, the state weight tensor in the coupled model is updated to generate a new perceptual model M′;

[0127] The update method is based on the "memory residual perception mechanism", which assigns learning and memory weights to significant variation factors to suppress ineffective updates.

[0128] Based on the terminal state changes described by M′, several “policy disturbance candidate paths” are automatically generated. Each path includes: new terminal instruction priority; changed task-terminal mapping relationship; instruction parameter fine-tuning scheme (such as execution latency tolerance, retry threshold, etc.).

[0129] Simulation tests are conducted on candidate paths to evaluate their performance improvement (execution success rate, response time, user satisfaction) under the current M′; the policy response tensor mapping function is used for scoring, and the optimal policy path is selected as the output C′;

[0130] If a sudden change in user behavior pattern is detected (such as a significant deviation from the original interaction logic), the personalized template module is invoked to adjust the strategy structure and form a "custom enhanced control strategy".

[0131] S600. If an abnormal state or task execution failure is detected in a certain AI interactive terminal, the system will perform task migration and control strategy reconstruction based on the terminal association weights and redundancy evaluation in the perception model M′. The specific steps include the following:

[0132] Each terminal is configured with a local execution status monitor and command acknowledgment mechanism:

[0133] If the instruction execution times out or the feedback status code is abnormal (such as abnormal abort, resource unavailable, communication failure), then the "fault event frame" is uploaded to the cloud.

[0134] The cloud performs timing determination and status marking on the fault event frame to confirm that terminal i is a fault node, and at the same time extracts the task list T_failed(i) bound to it.

[0135] Extracting state coupling weights and behavioral redundancy vectors:

[0136] Extract the coupling factor μ′(i,j,t) between the faulty terminal i and the other terminals j from the updated perception model M′, and extract its task semantic vector Ctx_i and historical behavior trajectory vector H_i;

[0137] Construct a candidate terminal set R = {j |μ′(i,j,t) >ε}, and filter out nodes with high state associations with faulty terminals. ε is a lower bound threshold for the coupling weight used to determine whether the state coupling strength between two terminals meets the minimum requirement of "redundancy relationship". The value of ε ranges from [0.3, 0.6], indicating that μ(i,j,t) must be greater than ε to enter the candidate set;

[0138] Redundancy scoring function calculation: For each candidate terminal j∈R, the following metrics are evaluated:

[0139] semantic similarity Ctx_i and Ctx_j are the semantic representation vectors of the current tasks of terminals i and j, respectively, and are usually composed of task type, user command content, environment label, etc.

[0140] Perceptual Coupling Sensor_i, Sensor_j: represent the sensor data sets of terminals i and j within the same time window, such as light intensity, temperature, sound intensity, face orientation, etc.

[0141] Resource absorption capacity This represents the load risk score of terminal j within the prediction time window;

[0142] The acquisition method is as follows: Build a sliding window prediction model (such as linear fitting or LSTM) based on historical load curves (CPU, memory, threads); output the probability of resource bottlenecks that may be triggered after the task is migrated;

[0143] After normalizing semantic similarity, perceptual coupling, and resource connectivity, the results are input into the redundancy scoring function: In the formula, w1 is the semantic adaptability weight; w2 is the perception collaboration weight; w3 is the resource carrying capacity weight; the weights can be dynamically set according to the importance of the task and the terminal type, for example, w1 is 0.2; w2 is 0.4; w3 is 0.4; the candidate sorting list R_sorted is obtained, and the terminal j* with the largest RScore is selected as the optimal takeover node.

[0144] Map each task in the task list T_failed(i) to the takeover terminal j*;

[0145] If the original task instruction contains terminal-specific modal parameters, the heterogeneous conversion module is called to perform instruction adaptation (such as voice interface → touch interface reconstruction).

[0146] Set all columns representing the tasks undertaken by terminal i in the original matrix to zero;

[0147] Insert a new task assignment weight t_j*k in the row where j* is located, and renormalize it according to the policy priority;

[0148] If j* resources are nearing a bottleneck, the system triggers a distributed forwarding mechanism to dispatch some tasks to the next candidate node in R_sorted.

[0149] The S700 system constructs a dynamic "SFC closed-loop control iterative mechanism" based on state parameters (S), feedback data (F), and control strategy (C) to achieve continuous linkage, self-learning, and self-evolution of control strategies among multiple terminals. The specific process includes the following key steps:

[0150] Constructing the SFC closed-loop control path:

[0151] Status input (S): The system periodically or based on event triggers collects real-time status parameters of each terminal, including task context, system load, user behavior, environmental awareness data, etc.

[0152] Feedback Acquisition (F): The results are collected after each round of control command execution, including task completion status, user response, abnormal events, and terminal status feedback;

[0153] Strategy Generation (C): Input S and F into the cloud modeling engine, dynamically update the control strategy C′ based on the latest model M′, and distribute it to each terminal for execution.

[0154] The system continuously executes a closed-loop cycle, forming a multi-round iterative sequence.

[0155] Each iteration updates elements and optimizes strategies:

[0156] Model Update (M′): In each round, the perception model M is updated based on the feedback F, and the terminal coupling degree, task partition weight and priority vector are adjusted. If user behavior habits, device availability and scene mode change, the system can adjust the model structure.

[0157] Control matrix T adjustment: The task allocation matrix is ​​reconstructed based on the new priority and resource status; if the priority of a terminal changes, fails, or the migration is successful, the corresponding mapping relationship in T is dynamically updated.

[0158] Strategy Evolution (C′): The control strategy includes the execution method, triggering order, and anomaly tolerance range; the system can select the optimal execution path through strategy perturbation simulation tests to improve overall response efficiency and stability.

[0159] Throughout the closed loop, terminals achieve linkage judgment and resource sharing through state coupling matrices; if the behavior pattern of a terminal evolves (such as an increase in interaction frequency or a shift in usage time), multiple terminals can be triggered to adjust their strategies simultaneously; an adaptive "master-slave" role switching mechanism is supported: if the original master terminal is used less, its tasks are smoothly migrated to terminals with frequent interactions.

[0160] Iteration Termination and Stability Determination Mechanism: The system defines the following termination condition to determine whether the current iteration has reached a "stable execution state":

[0161] The task completion rate is greater than or equal to the threshold (e.g., 98%); the user interaction interruption rate drops to a low threshold; there are no major strategy adjustments in the feedback F for three consecutive rounds; the control matrix T has no significant structural changes (e.g., node additions or deletions or weight mutations) for two consecutive rounds; after the stability conditions are met, the system freezes the current M′ and C′ versions and enters a "low-frequency monitoring state", only reactivating the closed-loop mechanism when there is an external emergency or when the user actively triggers it.

[0162] Example 2, please refer to Figure 2 As shown in this embodiment, an AI interactive terminal control system based on a cloud platform includes:

[0163] The perception and acquisition module collects a multi-dimensional set of state parameters S from multiple AI interactive terminals and uploads S to the cloud platform.

[0164] The model building module, based on the aforementioned S, establishes an interactive modeling engine in the cloud platform, including a behavior pattern recognition submodule and a state coupling submodule, to generate a multi-terminal coupled perception model M.

[0165] The priority evaluation module calculates the adaptive priority P of each AI interactive terminal in the current task sequence through the M, and constructs a control task allocation matrix T based on an asymmetric task mapping strategy.

[0166] The control strategy generation module generates a control strategy C based on the matrix T and the priority P, and the cloud platform distributes heterogeneous control instructions to the corresponding AI terminal.

[0167] The self-learning optimization module collects feedback data F in real time during the execution of instructions at the terminal, and inputs F and the original state parameter S into the self-learning module to perform adaptive updates of the interaction strategy, thereby obtaining the updated perception model M′ and control strategy C′.

[0168] If the control structure adjustment module detects an abnormal terminal status or task execution failure, it will migrate the task to the optimal candidate terminal based on the terminal association weight and redundancy evaluation in the coupling model M′, and update the control matrix T.

[0169] The closed-loop control module continuously iterates through the SFC closed-loop until the task execution of each terminal is stable, thus completing cross-terminal linkage and self-evolution control.

[0170] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A cloud-based AI interactive terminal control method, characterized in that: Comprise: S100, collect a plurality of AI interaction terminal multi-dimensional state parameter set S, and upload S to the cloud platform; S200, based on the S, in the cloud platform, an interaction modeling engine containing behavior pattern recognition sub-module and state coupling sub-module is established, and a multi-terminal coupling perception model M is generated; the interaction modeling engine is deployed in the distributed resource pool of the cloud platform, has heterogeneous data processing, low delay modeling and scalable inference ability; the behavior pattern recognition sub-module is used for cross-device behavior consistency modeling and behavior pattern timing recognition based on the uploaded multi-terminal user behavior trajectory; The state coupling sub-module is used for analyzing the real-time mutual influence relationship between different terminal states, identifying how a certain terminal state change affects the response or performance of other terminals; the model M is jointly constructed by the user behavior coupling features B_i output by the BPR and the state coupling tensor C_ij output by the SCS: M=f(B,C); M is a control graph structure represented by a graph, which supports real-time reconstruction of node priority, adaptive task allocation strategy change and prediction of task migration path cost in the task scheduling process; S300, calculate the adaptive priority P of each AI interaction terminal under the current task sequence through M, and construct a control task allocation matrix T based on an asymmetric task mapping strategy; including: extracting the state coupling factor μ(i,j,t) and the user behavior feature vector B_i between each terminal in the model M, forming a multi-dimensional coupling correlation graph G; wherein μ(i,j,t) is a real number, representing the state coupling strength or cooperation degree between terminal i and terminal j at time t; Perform regional division algorithm based on graph traversal heat gradient on the graph G, and aggregate terminals with similar behavior trajectories and response delays into response subdomains; In each subdomain, a sliding window asynchronous weight integral function is used to integrate and score the load change rate and task semantic matching degree of the terminal, generating a local dynamic priority P_i; Based on the global coupling degree and each P_i, a polynomial offset type task mapping function is constructed to realize asymmetric allocation of tasks among terminals, forming a control task matrix T; define matrix T=[t_ij], where t_ij represents the adaptive weight of task j allocated to terminal i; introduce a polynomial offset type asymmetric mapping function: t_ij=P_i×(1+γ·ΔCtx(i,j))×(1-κ·Load_penalty_i); in the formula, ΔCtx(i,j) is the context difference between task j and terminal i; Load_penalty_i is the current load penalty function of the terminal; γ, κ are strategy parameters for controlling semantic offset and resource regulation influence; T is not a symmetric matrix, allowing the same task to be allocated to multiple terminals with different weights, or allocated to empty; S400, generate a control strategy C according to the matrix T and the priority P, and distribute heterogeneous control instructions to the corresponding AI terminal by the cloud platform; S500, in the process of executing instructions on the terminal, real-time feedback data F is collected, and F and the original state parameter S are input into the self-learning module to perform interaction strategy adaptive update, and obtain an updated perception model M' and a control strategy C'; S600. If an abnormal terminal status or task execution failure is detected, the task is migrated to the optimal candidate terminal based on the terminal association weight and redundancy evaluation in the coupling model M′, and the control matrix T is updated. Each terminal is configured with a local execution status monitor and instruction acknowledgment mechanism. If the instruction execution times out or the feedback status code is abnormal, a fault event frame is uploaded to the cloud. The cloud performs timing determination and status marking on the fault event frame, confirms that terminal i is a fault node, and extracts the task list T_failed(i) bound to it. The coupling factor μ′(i, j) between the faulty terminal i and the other terminals j is extracted from the updated perception model M′. , t), construct a candidate terminal set R={j|μ′(i,j,t)>ε}; for each candidate terminal j∈R, perform redundancy scoring to obtain a candidate sorting list R_sorted, and select the terminal j with the largest RScore as the optimal takeover node; map each task in the task list T_failed(i) to the takeover terminal j, set all task columns of terminal i in the original matrix to zero, insert new tasks with assigned weights in the row where j is located, and renormalize according to policy priority; if the resources of j are close to the bottleneck, the system triggers the distributed forwarding mechanism to dispatch some tasks to the next candidate node in R_sorted; S700 continuously performs SFC closed-loop iteration until the execution of tasks on each terminal is stable, thus completing cross-terminal linkage and self-evolution control. 2.The AI interaction terminal control method based on a cloud platform of claim 1, wherein: In S300, a task-terminal mapping graph is constructed based on the task allocation matrix T and the priority vector P. An execution exclusion rule is applied to the same type of control task using a constraint graph coloring algorithm. The set of instruction scheduling paths with no conflict and optimal resource load among multiple terminals is calculated. For each terminal, the instruction semantic skeleton matching its device type, interaction interface and current context is called from the cloud instruction template library. A complete heterogeneous control instruction set C_i is generated through a dynamic template splicing mechanism. After C_i is generated, the local semantic rewriting engine under the edge node is executed to perform instruction semantic completion and multi-language translation for the control scenario. Ultimately, the control policy is broadcast to both the terminal and its alternative agent devices simultaneously via a dual-channel push mechanism. 3.The AI interaction terminal control method based on a cloud platform of claim 1, wherein: In S500, the terminal continuously records feedback data F during the instruction execution cycle, including task response latency, user interaction behavior trajectory, actual execution results, and local exception codes. The feedback data F and the initial state parameters S are input into the cloud self-learning module, the analysis engine based on the causal dependency path reconstruction mechanism is called, the abnormal feedback source is identified, a multi-factor causal chain graph is constructed in chronological order, and the structural weights of the model M are adjusted through the coupled tensor transformation function with memory residual update, and the updated model M′ is output. Based on M′, the policy disturbance analyzer is invoked to simulate the impact of different policy paths on system performance. The original control policy C is dynamically optimized through the policy response tensor mapping method to generate a new policy C′. 4.The AI interaction terminal control method based on a cloud platform of claim 1, wherein: In S600, the terminal is identified as having an execution failure, abnormal interruption, or unavailable resources through the instruction receipt monitoring mechanism and feedback event frame analysis, and is marked as a fault node. Extract the state coupling weights and behavioral redundancy vectors between the faulty terminal and other terminals from the updated perception model M′, call the redundancy scoring function to evaluate the redundancy of candidate terminals, and sort the candidate terminals that have semantic similarity, perception similarity and load-bearing capacity. Path optimization based on dynamic transition graph is performed among candidate terminals. The terminal with the lowest coupling cost is selected to take over the task. At the same time, the structure mapping update module is called to reconstruct the control task allocation matrix T. 5.The AI interaction terminal control method based on a cloud platform of claim 4, wherein: The evaluation of candidate terminal redundancy includes: Extract the coupling factor μ′(i,j,t) between the faulty terminal i and the candidate terminal j, and the task semantic vectors Ctx_i and Ctx_j to construct a terminal redundancy candidate set R, where i represents the faulty terminal, j represents the candidate terminal, and t represents the time corresponding to the coupling factor. Semantic similarity is calculated using the vector semantic offset function △Sem(Ctx_i,Ctx_j), and perceptual coupling is calculated using the collaborative sensing vector matching rate Sim_s(Sensor_i,Sensor_j). Resource carrying capacity is evaluated using the joint load trend fitting function P_Load(j,t). Here, Sensor_i and Sensor_j represent the sensor data sets of terminal i and terminal j within the same time window, respectively, and P_Load(j,t) represents the load risk score of terminal j within the prediction time window.

6. Normalize the three-dimension indicators into redundancy score function , wherein, is a semantic adaptability weight; is a perception cooperativeness weight; is a resource hosting capacity weight; and sort all terminals in R according to RScore in descending order, and output the optimal candidate hosting sequence.

7. A cloud platform-based AI interaction terminal control system for implementing the cloud platform-based AI interaction terminal control method of any one of claims 1-5. include: The perception and acquisition module collects a multi-dimensional set of state parameters S from multiple AI interactive terminals and uploads S to the cloud platform. The model building module, based on the aforementioned S, establishes an interactive modeling engine in the cloud platform, comprising a behavior pattern recognition submodule and a state coupling submodule, to generate a multi-terminal coupled perception model M. The interactive modeling engine is deployed in the distributed resource pool of the cloud platform and possesses heterogeneous data processing, low-latency modeling, and scalable inference capabilities. The behavior pattern recognition submodule is used for cross-device behavior consistency modeling and behavior pattern temporal recognition based on uploaded multi-terminal user behavior trajectories. The state coupling submodule is used to analyze the real-time mutual influence between different terminal states and identify how changes in the state of a certain terminal affect the response or performance of other terminals. The model M is constructed by the user behavior coupling feature B_i output by BPR and the state coupling tensor C_ij output by SCS, M=f(B,C). The model M is a control graph structure represented by graph neural network. The priority evaluation module calculates the adaptive priority P of each AI interactive terminal in the current task sequence through the M, and constructs a control task allocation matrix T based on an asymmetric task mapping strategy; the priority evaluation module extracts the state coupling factor μ(i,j,t) between each terminal and the user behavior feature vector B_i from the model M to form a multidimensional coupling association graph G. A region partitioning algorithm based on graph traversal heat gradient is performed on graph G to aggregate terminals with similar behavioral trajectories and response delays into response subdomains. Within each response subdomain, a sliding window asynchronous weight integration function is used to integrally score the load change rate of the terminal and the semantic matching degree of the task, generating a local dynamic priority P_i. Based on the global coupling degree and each P_i, a polynomial offset type asymmetric task mapping function is constructed to form a control task allocation matrix T. Where, T=[t_ij], t_ij represents the adaptation weight of task j assigned to terminal i, and: t_ij=P_i×(1+γ·ΔCtx(i,j))×(1-κ·Load_penalty_i) The control strategy generation module generates a control strategy C based on the matrix T and the priority P, and distributes heterogeneous control instructions to the corresponding AI terminal by the cloud platform. The self-learning optimization module collects feedback data F in real time during the execution of instructions at the terminal, and inputs F and the original state parameter S into the self-learning module to perform adaptive updates of the interaction strategy, thereby obtaining the updated perception model M′ and control strategy C′. If the control structure adjustment module detects an abnormal terminal status or task execution failure, it migrates the task to the optimal candidate terminal based on the terminal association weights and redundancy evaluation in the coupling model M′, and updates the control matrix T. The control structure adjustment module determines instruction execution timeout or feedback status code abnormality through the local execution status monitor and instruction acknowledgment mechanism configured on each terminal, and uploads the fault event frame to the cloud. It performs timing judgment and status marking on the fault event frame, confirms terminal i as the fault node, and extracts the task list T_failed(i). It extracts the coupling factor μ′(i, j, t) between the fault terminal i and the other terminals j from the updated perception model M′, and constructs a candidate terminal set R. It performs redundancy scoring on the candidate terminals in the candidate terminal set R, obtains the candidate sorting list R_sorted, and selects the terminal j with the largest RScore as the optimal takeover node. It maps the tasks in the task list T_failed(i) to the takeover terminal j and reconstructs the control task allocation matrix T. When the resources of the takeover terminal j* are close to the bottleneck, some tasks are dispatched to the next candidate node in the candidate sorting list R_sorted. The closed-loop control module continuously iterates through the SFC closed-loop until the task execution of each terminal is stable, thus completing cross-terminal linkage and self-evolution control.

Images