Multi-dimensional demand-driven large-model multi-task dynamic priority scheduling and multi-model collaboration method, system and application

Abstract

The invention discloses a multi-dimensional demand-driven large-model multi-task dynamic priority scheduling and multi-model collaboration method. The method comprises the steps of multi-task input and environment state perception: receiving an input concurrent task set and real-time environment parameters; multi-dimensional demand feature extraction: semantically analyzing the task set, and extracting a security demand level, a timeliness threshold, a social association degree and a personal comfort influence value of each task; dynamic weight generation based on a style model: calling a pre-training style decision model to generate a dynamic weight distribution rule; task priority quantitative calculation: carrying out tensor product calculation on the demand features and the weight vectors to obtain a comprehensive priority score; and multi-model collaborative scheduling execution: according to priority ranking, distributing high-priority tasks to a real-time response type model, distributing low-priority tasks to a resource optimization type model, and monitoring in real time and dynamically adjusting a queue. The invention also discloses a scheduling cooperation system for realizing the method, and the system has a wide application scene.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of model task scheduling optimization, and relates to a multi-dimensional demand-driven large model multi-task dynamic priority scheduling and multi-model collaboration method, system and application. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, large models have become the core force driving the intelligent transformation of various industries. From the initial natural language processing field to computer vision, speech recognition, intelligent control and other fields, the application scenarios of large models are increasingly rich. In intelligent customer service, smart home, intelligent driving, medical diagnosis, financial risk prediction and other industries, large models play a key role, helping enterprises and institutions to improve efficiency, reduce costs and innovate service models.

[0003] In recent years, with the explosive growth of data volume and the continuous improvement of computing power, the size and performance of large models have been continuously improved. More and more enterprises and research institutions invest a lot of resources in the research and application exploration of large models, and the market competition is increasingly fierce. At the same time, emerging technology directions such as multi-modal large models and multi-task learning have gradually become research hotspots, aiming to further improve the generalization ability and application range of large models.

[0004] However, with the complication and diversification of large model application scenarios, the problems of task scheduling and multi-model collaboration have become increasingly prominent. How to efficiently manage and schedule multiple large models so that they can dynamically allocate resources and adjust priorities according to the needs of different tasks, and realize the collaborative work between multiple models, has become a key problem that needs to be solved in the current artificial intelligence field. This not only relates to the performance and user experience of a single application system, but also affects the intelligent development process of the entire industry.

[0005] The existing technology has many deficiencies when facing multi-task concurrency and complex demand scenarios.

[0006] In fixed priority scheduling, since the priority is pre-set and fixed, it is difficult to adapt to complex and variable task demands. When some urgent tasks that are not pre-set as high priority appear, it may cause delay in the execution of critical tasks. For example, in an intelligent home control system, a fire alarm is suddenly detected, but if the priority of the fire alarm task is not pre-set high enough, and at this time the system is executing some low-priority tasks such as light adjustment, music playing, etc., it may cause the fire alarm to be processed in time, and the fire-fighting equipment cannot be started in time, causing serious consequences. For example, Chinese patent application CN119127420A only relies on a two-dimensional rule base of urgency and importance, lacking a mechanism for generating personalized weights.

[0007] The single model processing mode is difficult to cope with diversified task requirements. Different types of tasks may require different capabilities and knowledge, and a single model cannot perform well in all aspects. For example, in a comprehensive intelligent assistant system, both natural language dialogue tasks and image recognition, data analysis, and other tasks need to be handled. A single language model may perform well for natural language dialogue, but it may not be able to handle image recognition tasks, and it may not be able to accurately identify objects and scenes in images, and may not be able to meet the needs of users in multiple fields.

[0008] Static scheduling based on task types takes into account the differences in task types, but lacks in-depth analysis of the specific needs of tasks. The same type of task may have different urgency and resource requirements in different scenarios, and static scheduling cannot be flexible. For example, in video processing tasks, for real-time video conferencing video processing, the timeliness requirement is very high, and the video needs to be quickly encoded, transmitted and decoded; while for ordinary video file editing tasks, the timeliness requirement is relatively low. But the static scheduling mode cannot dynamically adjust resource allocation and priority setting according to the differences in these specific requirements.

[0009] In terms of multi-model collaboration, the early serial collaboration method lacks flexibility. Models are executed in a fixed order, and cannot be dynamically adjusted according to real-time task conditions. If in an image generation and description system, the quality of the generated image does not meet the expectations and needs to be regenerated, but in the serial collaboration method, the fixed process must be followed, i.e. image generation is completed first, and then image description is performed, which cannot timely correct and optimize the generated image, affecting the overall performance of the system and user experience. Moreover, in this way, the collaboration mechanism between models is simple, and the advantages of multiple models cannot be fully utilized, and it is also difficult to cope with complex multi-modal data and highly dynamic task environments. For example, when processing multi-modal data containing text, images and audio, it is difficult to effectively extract and fuse feature information of different modalities, resulting in insufficient processing capability for multi-modal data.

[0010] Therefore, it is of great practical significance and market value to study an efficient multi-dimensional demand-driven large model multi-task dynamic priority scheduling and multi-model collaboration system. SUMMARY

[0011] In order to solve the problems existing in the prior art, the purpose of the present application is to provide a multi-dimensional demand-driven large model multi-task dynamic priority scheduling and multi-model collaboration method, system and application.

[0012] The application can solve the problem that the task scheduling mode in the prior art cannot flexibly adapt to complex and changeable task requirements, and the problem that multi-model collaboration lacks flexibility and efficiency. Through the method of the application, reasonable scheduling of multi-task dynamic priority can be realized, and task priority can be dynamically calculated according to multi-dimensional demand characteristics such as safety requirement level, timeliness threshold, social correlation degree and personal comfort influence value, so as to ensure that urgent and important tasks are processed in time. At the same time, the multi-model collaboration efficiency is improved, different priority tasks are allocated to appropriate models for execution, the advantages of each model are fully utilized, resource utilization is optimized, and the efficiency, flexibility and adaptability of complex task processing are improved, so as to provide users with better and more efficient service experience, and meet the actual needs of multi-task processing scenes such as intelligent assistants, smart home control and intelligent office systems.

[0013] The application utilizes a scheduling and collaboration system including a multi-task input and environment state perception module, a multi-dimensional demand characteristic extraction module, a dynamic weight generation module, a task priority quantitative calculation module, a multi-model collaborative scheduling and execution module and the like to realize the method in the application.

[0014] The application provides a multi-dimensional demand driven large model multi-task dynamic priority scheduling and multi-model collaboration method, including the following steps:

[0015] Step S1: multi-task input and environment state perception

[0016] Receiving a set of concurrent tasks input by a user and real-time environment parameters, the environment parameters including physical scene state, device operation data and external event trigger signals and the like;

[0017] Step S2: multi-dimensional demand characteristic extraction

[0018] Performing semantic analysis on the task set to extract safety requirement level, timeliness threshold, social correlation degree and personal comfort influence value of each task;

[0019] Step S3: dynamic weight generation based on style model

[0020] Calling a pre-trained style decision model to generate a dynamic weight distribution rule according to user portrait and historical behavior data, the dynamic weight distribution rule mapping demand characteristics into a four-tuple weight vector W=(w 安全 ,w 时效 ,w 社交 ,w 舒适 );

[0021] Step S4: task priority quantitative calculation

[0022] The requirement features extracted from each task in step S2 are subjected to a tensor product operation with the weight vector generated in step S3, and the comprehensive priority score P of each task is output task =∑(F dim ×W dim );

[0023] Step S5: Multi-model collaborative scheduling execution

[0024] According to the comprehensive priority score, the tasks are ranked, high-priority tasks are assigned to real-time response models, low-priority tasks are assigned to resource optimization models, and the task execution state is monitored to dynamically adjust the queue.

[0025] In the safety requirement level calculation of step S2, first, the BERT model is called to perform context semantic analysis on the natural language description of each input task, and identify potential risk-related entities. The identification result will be matched with a pre-defined three-level risk knowledge graph, which takes device failure, biological hazards, and environmental disasters as the main line, and constructs L1, L2, and L3 level risk entity sets respectively. If the BERT model identifies L3 high-risk entities in the task, the absolute priority of the task needs to be guaranteed, and the safety requirement impact value F 安全 is forcibly set to the maximum weight upper limit of 0.9, which is hard constrained and verified by synchronous calling of environmental sensors to improve the confidence of entity identification and ensure immediate response and risk control of high-risk tasks.

[0026] When the identified entities are L1 and L2 entities: the maximum weight upper limit of the safety requirement impact value is not forcibly set, but the comprehensive safety risk is calculated through a risk superposition algorithm. When a task contains multiple risk entities, the formula R total is used to quantify the overall risk level, which is used as the safety feature component (F 安全 ) of the task for subsequent product calculation with the dynamic weight vector, and participates in the determination of the task priority score to reasonably reflect its influence in scheduling.

[0027] Specifically, when a task contains multiple L1 and L2 risk entities, the comprehensive safety risk is calculated by the formula R total =1-Π(1-R i ), where R i is the basic risk probability value of each identified entity, and the overall risk level is quantified by superimposing the independent probabilities of multiple entities, and the result is used as the safety feature component F 安全 .

[0028] In the timeliness threshold calculation of step S2, the system establishes a close mapping relationship between the task description and the current physical world clock through the dynamic time anchoring strategy. First, the natural language parsing module is called to extract the time expressions in the task, and the time standardization tool is used to accurately convert them into UTC timestamps, denoted as T deadline . Subsequently, the system accesses the Internet of Things device management module to obtain the current state data of the devices involved in the task. If the task execution depends on device linkage, the system retrieves the historical response delay of the corresponding device as the delay compensation item ΔT, and simultaneously obtains the current system time T now . Based on the above information, the system calculates the standardized urgency of the task using the formula , where T base represents the average task response period related to the scene, such as 300 seconds in an automatic cooking task in the kitchen. This formula is used to map task requirements of different time scales to a unified evaluation interval, and a smaller value indicates that the task is closer to the deadline. If the calculation result T critical is less than 0.2, the system determines that the task has immediate processing needs, automatically raises its timeliness feature value F 时效 , and guides it to enter the real-time response model processing path first, thereby ensuring that the time sensitivity of task execution is fully perceived and responded to, and improving the real-time performance and accuracy of the overall system scheduling.

[0029] In the timeliness threshold calculation in step S2, to ensure that the time reference T base dependent on the task scene can adapt to individual differences and behavior changes, the system introduces a dynamic calibration mechanism to build a time consumption probability distribution model based on user historical behavior data. The system first collects the time consumed by each task execution from user historical logs to form a sample set {t1, t2,..., t n} for different types of task scenes such as cooking or driving, and performs distribution fitting to reflect the time usage pattern. The Weibull distribution is selected as the statistical modeling basis, and its density function is defined as where t is the time consumed by a single task execution, λ is the scale parameter representing the typical time consumption interval, and k is the shape parameter determining the extent of distribution tail expansion. The system estimates the initial fitting parameters through maximum likelihood estimation, and then triggers a baseline parameter update every 50 new task records, re-optimizes the estimate value of λ using the EM algorithm, and dynamically adjusts T base to improve the accuracy and stability of the timeliness feature T critical . At the same time, to ensure the consistency of the multi-terminal user scheduling model, the system uses a federated learning mechanism to fuse the model update gradients of each local node across devices without transmitting original task records, realizing distributed evolution of the global time reference. This method enables the use of T baseIt can evolve according to real operation habits, and enhance the generalization ability and adaptability of time dimension priority judgment in multi-task multi-model system.

[0030] For social relevance analysis and personal comfort influence value quantification process, multi-level quantification, dynamic calibration and fusion strategy are adopted in the application, which specifically includes the following:

[0031] In the social relevance analysis of step S2, the system constructs a multi-layer social relationship network of the user based on a heterogeneous graph, and gives different structured weights to each social edge from the identity, content and behavior dimensions. The role label in the user address book is used for initialization assignment in the identity layer, and the edge weight is preset as 0.9 for supervisor, 0.7 for family, and 0.4 for neighbor, as a reference for basic social influence. The content layer introduces the RoBERTa model to analyze the semantic content and emotional polarity of the text communication content of the user in social platforms, message applications and other channels. When the system identifies a positive emotional trend, the corresponding edge weight increases by 0.2 based on the original basis, reflecting the importance of emotional affinity to the task. The behavior layer focuses on the quantification of user historical interaction behavior, and the interaction intimacy is calculated by the formula , wherein N interact represents the number of interactions with a specific contact in the last period, N total is the total number of social behaviors of the user, and T recent is the time interval from the last interaction (in hours). This design strengthens the social weight contribution of high frequency and recent interaction. After normalization of the above three-layer edge weights E1, E2 and E3, the social relevance index F i is aggregated according to the weight vector w 社交 . as the social feature value F 安全 of the task, which is used for subsequent fusion calculation of the four-tuple weight vector W=(w 时效 ,w 社交 ,w 舒适 ) generated by the style model, quantifying the influence degree of the social dimension in priority scheduling. This mechanism can accurately reflect the importance and intimacy of the personnel involved in the task in the social network, realize the interpersonal weight analysis in personalized multi-task collaborative scheduling, and also quantify the influence degree of the social dimension feature in the task priority, realizing the accurate modeling of the task urgency under the influence of the user's multi-level social network.

[0032] In the personal comfort influence value quantification process of step S2, a multi-source wearable device matrix is deployed to monitor the user's physiological and psychological state in real time, and these perception data are structured and mapped into comfort feature values in task scheduling. First, in terms of body discomfort, the system monitors the change rate of skin electric response signal, and when its differential value satisfies When the body temperature sensor detects a temperature below 32 degrees Celsius and the external environment humidity is higher than 80%, the system considers it a potential hypothermia risk in a cold and humid environment, and calculates the hypothermia risk coefficient where T skin is the current body temperature, k is the sensitivity coefficient, used to control the response gradient of the risk value to the temperature deviation, which is mapped as indicator item M2. In terms of psychological anxiety, the system collects voice input signals through the earphone terminal, analyzes the time series variance of the fundamental frequency F0, and when Var(F0)>50Hz, it is marked as an anxiety state and generates indicator item M3. Finally, the system aggregates the three normalized measurement values M1, M2, M3 according to their importance weights ω i in the comfort evaluation of each feature, to obtain the personal comfort influence value C=∑ω i ·M i , as the feature factor F 舒适 of the task in the comfort dimension, multiplied by the weight vector W=(w 安全 ,w 时效 ,w 社交 ,w 舒适 ) to calculate the final task priority score, converting subjective feelings into structured comfort influence values, ensuring that the system scheduling meets efficiency and risk while taking into account the physiological and psychological comfort needs of users.

[0033] In step S3, the style decision model realizes dynamic style selection of the user's current state and task context through a multi-modal matching engine, improving the personalization and adaptability of the scheduling strategy. The system first encodes the input into a feature vector, extracts the user's current stress index as the variance of the voice fundamental frequency and the text sentiment polarity value, and quantifies the task scene complexity through the number of concurrent tasks and the number of recognized risk entities, generating a 128-dimensional input feature vector V input . Subsequently, the system retrieves style embedding vectors from the pre-trained style library. Each style represents a specific scheduling preference, such as efficiency-oriented, interpersonal-sensitive, or risk-averse. During the retrieval process, the system calculates the cosine similarity between the input vector and each style embedding The style matching degree is measured, and the top 3 similar candidate styles are selected. If the safety requirement level detected in the input features reaches L2 or above, the system will skip the similarity sorting and directly specify the use of the risk-averse decision-making style to ensure the rigor of the high-risk task scheduling strategy. When the social association value exceeds 0.8 and involves family identity tags, the system switches to the interpersonal-oriented style to enhance the response adaptability of the human dimension. Finally, the selected style is input into the style verification module, which compares and discriminates the candidate style with the user's past behavior trajectory based on the generative adversarial network model. When the similarity confidence is less than 90%, it is marked as needing manual intervention, thereby ensuring the controllability and behavior consistency of style switching. This mechanism ensures the context awareness and user preference matching ability of style selection in the task priority weight generation process, making the style-driven weight vector W = (w 安全 ,w 时效 ,w 社交 ,w 舒适 ) more suitable for current multi-dimensional task scheduling needs.

[0034] In one specific embodiment of the present application, the use of a multi-modal matching engine enables more comprehensive feature fusion: the multi-modal matching engine encodes user voice fundamental frequency variance, text sentiment polarity value, task scene complexity, etc., and integrates multi-modal information including text semantics, physiological signals, and psychological state. Collaborative filtering relies more on user behavior history data, and user clustering focuses on group behavior commonality, which is insufficient for integrating real-time dynamic physiological and psychological states and task context characteristics.

[0035] Dynamic style adaptation is more accurate: by comparing and selecting the adaptive style with the pre-trained style vector library, and combining task scene dynamic triggering of special rules, such as directly specifying the risk-averse priority style when the safety requirement level reaches L2 or above, and switching to the interpersonal-oriented style when the social association value exceeds 0.8 and involves family, the system can respond in real time to dynamic changes in task urgency, safety risk, etc. The traditional model's personalized strategy is mostly based on static generation of historical data, making it difficult to adapt to real-time situational fluctuations.

[0036] In step S3, the weight mapping rule dynamically deforms the four-tuple weight vector W = (w 安全 ,w 时效 ,w 社交 ,w 舒适 ) through a style-related nonlinear adjustment strategy to enhance its sensitivity and adaptability to user style and task state. After matching the decision-making style, the system emphasizes the impact of task timeliness, applying exponential amplification transformation w′ 时效 = w 时效 × e 0.5×UrgencyScore to the timeliness dimension weight, where T criticalthe closer to the cutoff, the higher the score, and the exponential function ensures that the time weight grows nonlinearly. For users with a reflective matching style, the system introduces a sigmoid function to constrain all dimensions, and the calculation formula is where the parameter k = 5.0 controls the slope of the mapping curve, and the transformation is used to suppress the excessive bias of the weight in the original distribution, so that the weight is concentrated in the neutral interval, and the extreme decision-making tendency is reduced. In the adaptive style scenario, the system introduces the risk trend of environmental disturbance as an adjustment factor into the weight adjustment process, and the safety weight is fine-tuned according to the formula where represents the time gradient of the current risk value, and a is the disturbance learning rate set to 0.2, aiming to capture the dynamic feedback ability of environmental mutations. After the above three style transformations, the system performs normalization processing on the updated weight vector W' to ensure that dim = 1, and sets a hard lower bound constraint so that each dimension weight is not less than 0.05, thereby avoiding any demand feature being ignored in task scheduling. Through this nonlinear weight adjustment mechanism, the system realizes the fine integration of user style, task urgency, and environmental disturbance, making the final generated weight vector have higher adaptability and response accuracy in multi-model and multi-task scheduling.

[0037] In the weight mapping rule of step S3, the adaptive method of user portrait realizes personalized adjustment of style weight by combining professional attributes, behavior memory, and style fusion mechanism. The system first constructs the prior style preference distribution according to the user's professional label, taking the professional type as an upper constraint signal for style selection. If the user is identified as an engineer, the system will increase the probability of selecting the analytical style by 40%, strengthening the weight tendency of logical reasoning and efficiency priority; if the user is a salesperson, the probability of interpersonal-oriented style will increase by 60%, highlighting the importance of social dimension in the scheduling process. Subsequently, the system enables the historical behavior memory library to count the user's task scheduling intervention behavior in the past 30 days, extracts the number of times the user actively modifies the default weight, and calculates the style deviation according to the formula where N override represents the number of active adjustment behaviors, N total is the total number of such tasks, and Sign(Δw) represents the deviation sign of the user's actual preference direction and the system's default style direction, used to determine whether to enhance or suppress the current style. Finally, the system generates the final weight vector W final = β·W style +(1-β)·W history , where W style is derived from the style decision result in the current task scenario, and W historyThis represents the average style preference constructed based on users' historical behavior. The fusion coefficient β is controlled by the style shift. When δ s When β > 0.3, it is set to 0.2, indicating an enhanced influence of historical behavior; otherwise, it is set to 0.7, focusing more on current style decisions. This mechanism achieves a dynamic balance between user style and long-term behavior and short-term tasks, making the generated weight vector more aligned with user personality and real-time needs, and ensuring the continuous adaptability of the scheduling strategy and human-machine consistency in a multi-model, multi-task scheduling system.

[0038] In the weight mapping rules of step S3, when a style policy conflict is triggered by the task requirement dimension, the system activates a conflict resolution mechanism to ensure the continuity and stability of priority scheduling. The system first constructs a conflict matrix M of four-dimensional requirement features. 4×4 , where each element m ij =Cov(F i ,F j ) represents the demand dimension F i With F j In historical tasks, covariance is used to measure the degree of cooperation or mutual exclusion; positive covariance represents cooperation, and negative covariance represents a tendency towards conflict. When style selection triggers conflict, the system searches for the Pareto optimal frontier in the weight vector space to find the weight combination that satisfies the maximum weighted utility requirement without triggering severe conflict. Its optimization objective is max(∑w i F i ), constrained by constraint m ij w i w j <θ, where θ is the conflict intensity threshold set to 0.05, ensuring that the joint weight product of any two highly conflicting dimensions does not exceed an acceptable level. If no weight solution satisfying the conditions can be found within this feasible region, the system initiates a degradation protocol to alleviate conflict pressure. The first step forces the personal comfort weight to zero to release weight space; the second step uses the current task risk level as an adjustment factor to proportionally reduce the social dimension weight, using rule w. 社交 :=w 社交 -0.1×RiskLevel, where RiskLevel is a quantified value of the security level; a higher RiskLevel indicates a higher priority for ensuring security requirements. The system continuously iterates and adjusts the weights of each dimension until all conflicting covariance conditions meet the constraint thresholds. The path of this round of downgrade adjustment is stored in the style decision feedback module as a reference for future weight mapping corrections in the user profile model. This conflict resolution mechanism achieves dynamic reconciliation of the weight structure under multi-style intervention, ensuring that the multi-task, multi-model scheduling maintains the robustness and convergence of the operating strategy under complex multi-dimensional constraints.

[0039] In the task priority quantification calculation of step S4, the system adopts a fuzzy entropy weight fusion algorithm to integrate the uncertainty of task requirements and the subjectivity of user style preferences. First, the input four-dimensional requirement characteristics F dim are fuzzified to construct their corresponding triangular fuzzy number representation wherein the median f mid is the original requirement characteristic value, f low =f mid *(1-sigma dim ), f high =f mid *(1+sigma dim ), and sigma is the fluctuation rate of the current dimension in historical tasks, used to represent the fuzzy boundary of the stability of the dimension weight. Subsequently, the system calculates the objective weight of each fuzzy feature according to its information entropy, and the entropy value is defined as wherein mu i represents the membership degree at the three points of the fuzzy number, used to reflect the degree of uncertainty of the information, and the higher the entropy value, the more dispersed the dimension, and the lower the weight. The system calculates the objective weight based on this, and fuses it with the subjective weight W dim output by the style decision module, and finally the priority calculation adopts a weighted hybrid strategy: wherein the fusion coefficient gamma is set to 0.3, representing that the subjective style weight has higher control force in task scheduling, and the objective entropy weight is mainly used to adjust the reliability of the feature information amount. After the fuzzy number operation, in order to obtain accurate priority values that can be directly sorted, the system uses the barycenter method for defuzzification, that is, wherein mu(x) represents the membership function corresponding to the fuzzy number, and the integral calculation of the expected value of the fuzzy number is used as the final task score. Through the fuzzy entropy weight fusion algorithm, the system realizes the two-way reconciliation of the dynamic change of multi-dimensional characteristics and the adaptation of style weight, so that the task priority determination is more stable and interpretable in dealing with complex and uncertain environments.

[0040] In the specific implementation process of the present application, the style decision module is a submodule of the dynamic weight generation module, responsible for integrating the output of the style decision model and executing the complete style decision process. Specifically, it includes: calling the style decision model for multi-modal matching, determining the final style according to the task characteristics, generating a dynamic weight vector through a nonlinear adjustment strategy, and triggering a conflict resolution mechanism to handle style strategy conflicts. It is the "execution carrier" of the style decision model and contains extension functions such as style verification and weight mapping.

[0041] Style decision feedback module: auxiliary subassembly of style decision module, main function is to record the weight adjustment path in conflict resolution mechanism, and feed these historical adjustment data to user portrait model and style decision model.

[0042] In a specific embodiment of the present application, the dynamic priority scheduling scheme can also be integrated with the rule base scheme or the expert system scheme and applied in a hybrid scene to adapt to system performance differences.

[0043] The present application also provides a scheduling coordination system for implementing the above method, which comprises: a multi-task input and environment state perception module, a multi-dimensional demand feature extraction module, a dynamic weight generation module, a task priority quantization calculation module, and a multi-model collaborative scheduling execution module.

[0044] The multi-task input and environment state perception module is used to obtain a task set and environment parameters.

[0045] The multi-dimensional demand feature extraction module is used to extract a safety demand level, a timeliness threshold, a social correlation degree, and a personal comfort influence value.

[0046] The dynamic weight generation module is used to generate a dynamic four-tuple weight vector according to a user portrait and historical behavior.

[0047] The task priority quantization calculation module is used to perform tensor operation on the demand features and the weight vector and output a task priority score.

[0048] The multi-model collaborative scheduling execution module is used to schedule tasks to a real-time response type model or a resource optimization type model according to the priority, and dynamically adjust an execution queue.

[0049] The present application also provides an application of the above method or the above system in intelligent home control, production line task scheduling, etc.

[0050] The beneficial effects of the present application include: the present application improves the accuracy of task response through dynamic priority scheduling: the prior art adopts fixed priority or static scheduling method, which cannot adapt to the differences in real-time task requirements. The present application solves the problem of "one-size-fits-all" priority setting in static scheduling by extracting multi-dimensional demand characteristics (safety, timeliness, sociality, comfort) and generating dynamic weights, combining user style and environmental state to adjust the priority in real time, dynamically adjusts according to the real-time safety risk, time urgency, social relevance, etc. of the task, ensures that high-priority tasks are processed first, and improves the response accuracy of the system to complex scenarios. The present application enhances the processing capacity and flexibility of the system through multi-model collaboration: the single model in the prior art cannot cope with diversified tasks, and the early serial collaboration method has fixed process and cannot be dynamically adjusted. The present application assigns tasks to different types of models based on priority, supports dynamic adjustment of the queue, and integrates the advantages of multiple models to process multi-modal data, solving the problems of single model capacity limitation and serial collaboration flexibility deficiency. Through multi-model collaboration, real-time tasks can be efficiently processed, non-urgent tasks can be optimized, and the process can be dynamically corrected to enhance the processing capacity of the system for diversified and multi-modal tasks. The present application improves user experience and system robustness through personalized and adaptive scheduling: the prior art lacks consideration of user's personalized needs and weak adaptability to environmental changes. The present application introduces user portrait, historical behavior data and style decision model, combines personal comfort data perceived by wearable devices, and dynamically generates a weight vector that meets the user's preferences. It solves the problem of "general scheduling ignoring individual differences" in the prior art, and through personalized weight allocation, such as engineers focusing on analytical style, sales focusing on interpersonal orientation style and environmental adaptability adjustment, and prioritizing related tasks when the user feels uncomfortable, it balances the user's physiological and psychological needs while ensuring efficiency and safety, improving user experience; at the same time, through the conflict resolution mechanism, the conflict between safety and social needs is handled to ensure stable operation of the system under complex constraints and enhance robustness. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0052] Figure 1 is the overall architecture diagram of the present application.

[0053] Figure 2 is the safety requirement level calculation flowchart of the present application.

[0054] Figure 3 is the social relevance analysis schematic diagram of the present application. DETAILED DESCRIPTION

[0055] The application will be further described in conjunction with the following specific examples and drawings. The process, conditions, experimental methods, etc. for implementing the application, except for the following specifically mentioned contents, are the general knowledge and common sense in the art, and the application has no special restrictions.

[0056] The application provides a multi-dimensional demand-driven large model multi-task dynamic priority scheduling and multi-model collaboration method, and the invention points involved include the following:

[0057] Multi-dimensional demand feature extraction mechanism: comprehensively analyze the task from four dimensions of safety demand level, timeliness threshold, social correlation degree and personal comfort influence value, and realize accurate quantification of each dimension by combining BERT model, risk knowledge graph, time standardization tool, heterogeneous graph and wearable device data, to provide multi-dimensional basis for priority calculation.

[0058] Dynamic weight generation and adjustment strategy: generate a four-tuple weight vector based on user portrait, historical behavior data and pre-trained style decision model, select the adaptive style through a multi-modal matching engine, and introduce a nonlinear adjustment mechanism (such as exponential amplification, Sigmoid function smoothing, risk trend fine-tuning), while realizing personalized adjustment combined with professional attributes and behavior memory, and having a conflict resolution mechanism to ensure that the weight vector fits the user style and task demand.

[0059] Fuzzy entropy weight fusion priority quantization calculation: fuzzy processing of four-dimensional demand features, combining information entropy to calculate objective weights, and fusing with subjective weights of the style decision module, using a weighted hybrid strategy and barycenter method to solve the fuzziness, realizing two-way reconciliation of dynamic changes of multi-dimensional features and style weight adaptation, and obtaining an accurate task priority score.

[0060] Some terms involved in the application are explained as follows:

[0061] Working style: the behavior, preference and habit exhibited by an individual when facing tasks and work.

[0062] Decision-making style: a working style that tends to quickly analyze information, make decisive decisions, and focus on efficiency and results.

[0063] Analytical style: a working style that likes to analyze in depth, collect a large amount of data and make detailed reasoning, and postpones decision-making until all information is mastered.

[0064] Intuitive style: a working style that relies on intuition and experience for quick decision-making, does not rely on excessive data analysis, and tends to be innovative.

[0065] Task-oriented style: Focuses on achieving goals, emphasizes efficient task execution and ensuring completion, and prioritizes efficiency and productivity in doing things.

[0066] People-oriented style: Emphasizes interaction and teamwork, focusing on interpersonal relationships and communication in doing things.

[0067] Time-oriented style: Values time management, is good at planning and scheduling, and ensures timely completion of tasks.

[0068] Creative style: Tends to innovative thinking, exploring new ideas, and not being bound by conventional solutions.

[0069] Reflective style: Thoroughly considers before making decisions, analyzes multiple options and results, and avoids hasty action.

[0070] Adaptive style: Quickly adapts to new environments and tasks, and flexibly adjusts strategies to cope with changes.

[0071] Multi-dimensional demand-driven large model multi-task dynamic priority scheduling and multi-model collaborative system: A system that extracts multi-dimensional demand features, generates dynamic weights, quantifies priorities, and collaboratively schedules multiple models to cope with complex task scenarios.

[0072] Security demand level: Identifies potential risk entities in tasks through BERT models and risk knowledge graphs, and divides risk levels (such as L1, L2, L3 levels).

[0073] Time-sensitive threshold: Based on task time expressions and device status, calculates an index reflecting the urgency of task time.

[0074] Social connection degree: Based on user multi-layer social relationship network, quantifies the degree of social influence from identity, content, and behavior dimensions.

[0075] Personal comfort influence value: Combining wearable device monitoring of physiological and psychological data, quantifying the influence of tasks on user comfort requirements.

[0076] Dynamic weight vector: According to user portrait, historical behavior and task situation, generates a weight combination reflecting the importance of safety, timeliness, sociality and comfort dimensions.

[0077] Real-time response model: Used to handle high-priority tasks, capable of rapid response.

[0078] Resource optimization model: Used to handle low-priority tasks, focusing on resource optimization.

[0079] Risk knowledge graph: taking equipment failure, biological hazards and environmental disasters as the main line, a graph containing L1, L2 and L3 risk entity set is constructed.

[0080] Fuzzy entropy weight fusion algorithm: an algorithm for integrating task requirement uncertainty and user style preference subjectivity to calculate task priority.

[0081] The application provides a multi-dimensional demand-driven large model multi-task dynamic priority scheduling and multi-model collaborative method, which comprises the following steps:

[0082] Step S1. Multi-task input and environment state perception: receiving the concurrent task set and real-time environment parameters;

[0083] Step S2. Multi-dimensional demand feature extraction: semantic analysis of the task set, and extraction of demand features such as safety requirement level, timeliness threshold, social correlation degree and personal comfort influence value of each task;

[0084] Step S3. Dynamic weight generation based on style model: calling a pre-trained style decision model to generate a dynamic weight distribution rule, and performing demand feature mapping to obtain a weight vector;

[0085] Step S4. Task priority quantitative calculation: performing tensor product calculation on the demand features of each task in step S2 and the weight vector generated in step S3 to obtain a task comprehensive priority score;

[0086] Step S5. Multi-model collaborative scheduling execution: according to the comprehensive priority, high-priority tasks are allocated to real-time response type models, low-priority tasks are allocated to resource optimization type models, and the queue is monitored and dynamically adjusted in real time.

[0087] In step S2, the risk entities in the safety requirement level are divided into L1, L2 and L3 levels.

[0088] For example, when the L3 level entity is "fire" and the smoke detector reading exceeds 500ppm, the system confirms the risk level is valid and locks the task as a high safety priority task. For task content containing multiple L1 and / or L2 level risk entities, the system uses a risk superposition algorithm to calculate the comprehensive safety risk, i.e. using the formula R total =1-∏(1-R i ), where R i represents the basic risk probability value corresponding to each identified entity, and the overall risk degree is quantified by superimposing the independent probabilities of multiple entities, which is finally used as the safety feature component F 安全 of the task for subsequent calculation with the dynamic weight vector W=(w 安全 ,w 时效 ,w 社交 ,w舒适 ) of the product, the priority score of the task is output. This mechanism ensures the immediate identification and response to high-risk tasks, improving the system's perception and processing ability of potential dangers.

[0089] In the security requirement level calculation of step S2, to improve the response accuracy after the identification of L3-level high-risk entities, the system introduces an environmental sensor verification expansion module and integrates a dynamic risk propagation model to realize real-time prediction of potential disaster propagation paths and response time. When detecting L3-level risk entities such as fire and collapse, the system immediately calls the physical simulation engine, initializes the temperature field and gas concentration distribution according to the current environmental sensor readings, and calculates the temperature change rate and fire spread speed based on the thermodynamic diffusion equation , where is the rate of change of temperature with respect to time, and a is the thermal diffusion coefficient representing the heat conduction efficiency, is the spatial second-order gradient of the temperature field reflecting the direction and intensity of heat diffusion, and Q gas is the gas release rate per unit volume used to describe the heat source term generated by continuous combustion. The simulation model maps the temperature diffusion path to the indoor three-dimensional point cloud structure data obtained by real-time scanning, evaluates the shortest time path for the high-temperature area to expand to the key locations (passage exit, gas source interface), and determines the safety disposal time window. If the system determines that this window is less than 60 seconds, it immediately triggers the emergency priority upgrade mechanism to increase the comprehensive priority value of the corresponding task to the top of the task queue, bypassing the regular scheduling weight calculation logic, and activates the device linkage protocol to instruct the gas valve to automatically close, the exhaust system to start, and other operations. This mechanism effectively improves the response timeliness and execution efficiency of the task scheduling system in high-risk situations through the closed-loop fusion of physical field perception, simulation deduction, and action linkage, ensuring dynamic safety redundancy in multi-task coordination processes.

[0090] Embodiment 1: Daily life scene application

[0091] Taking the smart home control scene as an example, the user simultaneously issues three task instructions to the smart home system: one is to turn on the air purifier in the kitchen to deal with the slight kitchen oil smoke (task A); two is to start playing soothing music after 15 minutes (task B); and three is to receive and display important messages sent by a friend from a distance (task C).

[0092] The system first receives these concurrent task sets through input devices such as smart speakers and smart control panels. At the same time, various environmental sensors start working, such as smoke sensors detecting kitchen oil smoke concentration data as physical scene state parameters, and smart speaker and smart control panel running state data as device running data being acquired by the system.

[0093] In the multi-dimensional demand feature extraction stage, for task A, through semantic analysis and risk knowledge graph matching, it is found that kitchen fume may cause air quality to decline, but it belongs to a relatively low level of risk, and the safety demand level is rated as L1 level; in terms of timeliness, since the user issues an instruction immediately, the time difference between the current time and the task start time is small in the timeliness threshold calculation, and the absolute timeliness value is relatively high; the social correlation degree is 0 because it does not involve social correlation; in terms of personal comfort influence value, through wearable device monitoring, it is found that the user is currently active in the kitchen, and the fume has a certain influence on comfort, and the comprehensive comfort influence value is 0.4.

[0094] For task B, the safety demand level is 0 because playing music does not have safety risk; in the timeliness threshold calculation, according to the time expression "15 minutes later start playing" converted into UTC timestamp, combined with the current time and the possible device response delay parameter ΔT, the absolute timeliness value is calculated; the social correlation degree is 0; in terms of personal comfort influence value, considering that soothing music helps to improve the user's relaxation level, it is set to 0.3.

[0095] For task C, the safety demand level is 0; in terms of timeliness, since the message is received immediately, the timeliness is high; according to the multi-dimensional social relationship graph analysis, the friend identity corresponds to a certain weight, and if the message content is positive emotion, the weight is further improved, assuming that the comprehensive social correlation degree is 0.6; the personal comfort influence value is 0.

[0096] The style decision model generates a dynamic weight vector according to the user's past preferences for task processing, such as historical behavior data of paying more attention to timeliness or social relationship

[0097] W=(0.2, 0.3, 0.4, 0.1).

[0098] Then the task priority quantization calculation is carried out, the comprehensive priority score of task A is

[0099] PA=0.3×0.2+0.7×0.3+0×0.4+0.4×0.1=0.31;

[0100] The comprehensive priority score of task B is

[0101] PB=0×0.2+0.5×0.3+0×0.4+0.3×0.1=0.18;

[0102] The comprehensive priority score of task C is

[0103] PC=0×0.2+0.8×0.3+0.6×0.4+0×0.1=0.48.

[0104] According to the priority score ranking, task C has the highest priority, task A is second, and task B is the lowest. The system assigns task C to a real-time response model, such as a GPT-based model, which quickly processes the important message from the friend and displays it on the smart device screen; task A is assigned to a resource optimization model, such as a Llama-based model, which optimizes resource utilization while completing the task, turning on the kitchen air purifier; task B waits in the task queue and is executed when the resource optimization model has idle resources, starting to play soothing music after 15 minutes.

[0105] Example 2 Industrial Production Scene Application

[0106] In the task scheduling scenario of an automated production line, multiple tasks exist simultaneously on the production line. Task D is the detection and early warning of potential faults due to abnormal vibration of a key device; task E is the material distribution of a batch of products according to the production plan in 2 hours; task F is the processing of historical data accumulated in the production line monitoring system for production efficiency analysis.

[0107] The system receives these tasks through the interface of the production line monitoring system, and simultaneously obtains device operation data such as vibration sensor data and temperature sensor data of key devices, as well as real-time production progress data of the production line as environmental parameters.

[0108] For task D, through semantic analysis and risk knowledge graph matching, the abnormal vibration of the device may cause device failure, and the safety requirement level is rated as L2; in terms of timeliness threshold, the device abnormality needs to be handled immediately, so the absolute timeliness value is very low; the social association degree in the industrial production scene mainly considers the association with the upstream and downstream of the production process, assuming an association degree of 0.3; the personal comfort influence value is not considered in the industrial scene, and is set to 0.

[0109] For task E, the safety requirement level is 0; the timeliness threshold is calculated according to "material distribution in 2 hours", combined with possible device preparation time and transportation time delay parameters to calculate the absolute timeliness value; the social association degree is associated with the upstream and downstream of the production process, and is set to 0.2; the personal comfort influence value is 0.

[0110] For task F, the safety requirement level is 0; the timeliness threshold is relatively low because it is processing historical data and does not require immediate completion; the social association degree is 0; the personal comfort influence value is 0.

[0111] The style decision model generates a dynamic weight vector based on the historical data of the production line's scheduling habits and the emphasis on production efficiency and safety

[0112] W = (0.4, 0.3, 0.2, 0.1).

[0113] Task priority quantification calculation is performed, and a comprehensive priority score of task D is obtained

[0114] PD = 0.8 * 0.4 + 0.1 * 0.3 + 0.3 * 0.2 + 0 * 0.1 = 0.41

[0115] A comprehensive priority score of task E is obtained

[0116] PE = 0 * 0.4 + 0.6 * 0.3 + 0.2 * 0.2 + 0 * 0.1 = 0.22

[0117] A comprehensive priority score of task F is obtained

[0118] PF = 0 * 0.4 + 0.2 * 0.3 + 0 * 0.2 + 0 * 0.1 = 0.06

[0119] According to the priority ranking, task D has the highest priority, task E has the second highest priority, and task F has the lowest priority. The system assigns task D to the real-time response model, which uses its powerful analysis capabilities to quickly diagnose equipment abnormalities, such as by comparing with historical fault data, using fault prediction models, etc., to timely issue warnings and provide possible causes and solutions; task E is assigned to the resource optimization model, which reasonably arranges material distribution preparation work, such as scheduling transportation vehicles and preparing materials, according to the current state and resource situation of the production line; task F is processed when the resource optimization model is idle, and historical data is analyzed to generate a production efficiency report to provide a basis for subsequent production decisions. Through such task scheduling and multi-model collaboration, the production line can timely handle key tasks, reasonably arrange resources, improve production efficiency and stability, avoid production stagnation caused by equipment failure, and fully demonstrate the adaptability and advantages of the system in industrial production scenarios.

[0120] Taking a traditional industrial production scheduling system as an example, the system uses a static rule scheduling mode, pre-sets a fixed priority sequence for production tasks, such as device maintenance tasks having higher priority than material transportation tasks, and material transportation tasks having higher priority than data statistics tasks, and only relies on a single PLC (Programmable Logic Controller) module to handle all task instructions. In terms of device connection, wired Ethernet is used to realize communication between the control hub and production devices, and dynamic sensing devices such as wireless sensors cannot be accessed. When multiple tasks are triggered at the same time, the system strictly executes according to the pre-set priority sequence, for example, even if a material transportation task has caused a risk of shutdown due to lack of materials on the production line, if a low-level device maintenance task is being executed at this time, the system still needs to wait for the completion of the maintenance task before responding to the transportation demand.

[0121] In the above two example scenarios, the performance parameter indicators between the technical solutions of the present application and the prior art are compared as shown in the following table:

[0122] Performance parameter index table of different technical solutions

[0123]

[0124]

[0125] Example 3

[0126] In the specific implementation process of the present embodiment, a rule-based static priority scheduling and multi-model collaboration scheme can also be used: a task priority rule base covering different scenarios (such as home and office) is constructed in advance (such as "fire-related tasks have the highest priority, followed by work tasks, and finally personal chores"), the tasks are matched with the rule base to determine the priority, and then distributed to the corresponding model. This scheme does not require dynamic weight generation and complex feature extraction, and achieves scheduling through fixed rules, and is suitable for environments with relatively simple scenarios and limited task types.

[0127] Suppose in a small office scenario, there are tasks such as file printing, urgent meeting notification, and daily email processing. The pre-constructed task priority rule base sets: "urgent meeting notification task has the highest priority, file printing task is divided into high, medium, and low priority according to the urgency of the file, and daily email processing task has the lowest priority". When the system receives a task, such as an employee A sending a normal file printing task and receiving an urgent meeting notification task. The system matches the tasks with the rule base and determines that the urgent meeting notification task has a higher priority, and first distributes it to the message push model responsible for fast information transmission, and immediately sends a notification to all participants; while the file printing task enters the task queue and waits to be distributed to the print management model for processing. In this scenario, since the task types are relatively fixed and the scenario is simple, through this rule-based static priority scheduling and multi-model collaboration scheme, the task scheduling can be completed more efficiently, and complex computing resources are not required to dynamically analyze task features.

[0128] Example 4

[0129] In the specific implementation process of the present embodiment, an expert system-based multi-model collaboration scheme can also be used: integrate domain expert experience to construct a decision tree (such as "if the task involves safety risks, it is handled first; if it is a social task, it is sorted according to the closeness of the relationship"), determine the priority of the task through decision tree reasoning, and then distribute the task according to the model performance parameters (such as response speed and accuracy). This scheme emphasizes the direct application of expert knowledge, reduces the dependence on complex algorithm models (such as BERT and RoBERTa), and is suitable for fields with high requirements for explainability.

[0130] Taking the medical diagnosis field as an example, a multi-model collaborative diagnosis system based on an expert system is constructed. The experience of field experts is integrated to construct a decision tree, such as the rule "if the body temperature of a patient is higher than 38 DEG C and accompanied by a cough symptom, then the diagnosis of a respiratory tract infection disease is preferred; if the patient has a chest pain symptom and an electrocardiogram is abnormal, then the diagnosis of a heart-related disease is preferred". When patient data is input, for example, patient Zhang San, the body temperature is 38.5 DEG C and the cough is frequent, the system reasons through the decision tree and determines that the task (diagnosis of Zhang San) involves a respiratory tract infection related disease and the priority is high. Then, according to different model performance parameters, the task is assigned to a medical image analysis model that is good at diagnosing respiratory tract diseases (the model has a recognition accuracy of up to 90% for lung images and a fast response speed, and the average diagnosis time is 2 minutes), and the model analyzes the images such as the lung X-ray film of Zhang San to assist the doctor to quickly and accurately make a diagnosis. This scheme is very suitable for fields with high requirements for medical diagnosis explainability because the decision tree is constructed by using expert experience and the reasoning process is clear.

[0131] The protection scope of the present application is not limited to the above embodiments. Changes and advantages that can be thought of by those skilled in the art without departing from the spirit and scope of the present application are included in the present application and are protected by the appended claims.

Claims

1. A method for dynamic priority scheduling and multi-model collaboration of large-scale models driven by multi-dimensional needs, characterized in that, The method includes: Step S1. Multi-task input and environmental state awareness: Receive the input of concurrent task sets and real-time environmental parameters; Step S2. Multi-dimensional requirement feature extraction: Semantic parsing of the task set to extract the requirement features for each task; Step S3. Dynamic weight generation based on style model: Call the pre-trained style decision model to generate dynamic weight allocation rules, and perform requirement feature mapping to obtain weight vectors; Step S4. Task Priority Quantification Calculation: Perform tensor product calculation on the requirement features of each task in Step S2 and the weight vector generated in Step S3 to obtain the comprehensive task priority score; Step S5. Multi-model collaborative scheduling and execution: Sort according to the comprehensive priority, assign different models to tasks according to the priority, and monitor and dynamically adjust the queue in real time.

2. The method as described in claim 1, characterized in that, In step S2, the demand characteristics include safety demand level, timeliness threshold, social relevance, and personal comfort impact value; And / or, When calculating the security requirement level, the natural language description of each input task is subjected to contextual semantic analysis to identify potential risk-related entities. The identification results are then matched with the risk knowledge graph to construct a multi-level risk entity set. And / or, Risk levels include L1, L2, and L3; And / or, The risk knowledge graph is predefined, and the main lines of the graph include equipment failure, biological hazards, and environmental disasters. And / or, When the risk level is L1 or L2, the comprehensive safety level is calculated using a risk superposition algorithm. When the risk level is L3, the impact value of the mission's security requirements is F. 安全 Forced to the maximum weight limit, and simultaneously invoked environmental sensors for verification; And / or, The maximum weight limit is 0.

9.

3. The method as described in claim 2, characterized in that, When calculating the timeliness threshold, a mapping relationship is established through a dynamic time anchoring strategy, including: parsing and extracting the time expression in the task and converting it into a UTC timestamp; combining the UTC timestamp with the device's historical response latency, the current system time, and the scene-related average task response cycle to calculate the standardized urgency of the task; if the standardized urgency is less than a preset threshold, the timeliness feature value is automatically increased to guide priority processing. And / or, The threshold for the standardized urgency is set to 0.2; And / or, A dynamic calibration mechanism is introduced, and a time consumption probability distribution model is constructed based on user historical behavior data to obtain the statistical distribution characteristics of task execution time, and to determine the time benchmark on which each type of task depends, including: The execution time samples of each task are collected from user logs, and the distribution form is selected for fitting and modeling. The shape parameters and scale parameters are continuously optimized through maximum likelihood estimation and EM algorithm, and the scene-related average task response cycle is dynamically adjusted. And / or, It supports automatic parameter updates upon the arrival of new tasks and enhances the model's generalization ability through federated learning and cross-user transfer mechanisms, enabling adaptive characterization of timeliness requirements.

4. The method as described in claim 2, characterized in that, When constructing social relevance, a multi-layered social relationship network of users is built based on a heterogeneous graph, and different weights are assigned to the social edges between users and others from three dimensions: identity, content, and behavior. Specifically, the identity layer initializes the social edge weights through role tags in the address book, the content layer uses the RoBERTa model to identify emotional relationships in text communication and dynamically adjusts the edge weights, and the behavior layer quantifies based on the user's historical interaction behavior. By normalizing the weights of the three layers and aggregating them with the layer dimension weight vectors, a social relevance index is obtained. This index is then combined with a four-tuple weight vector generated by a style model to weight the social relevance of different dimensions, ultimately outputting the social relevance index, which is used to calculate the final task priority score.

5. The method as described in claim 2, characterized in that, When quantifying the impact value of personal comfort, a matrix of multi-source wearable devices is deployed to monitor the user's physiological and psychological state in real time. Features are extracted from three dimensions: physical discomfort, temperature and humidity imbalance, and psychological anxiety to assess the degree of discomfort. The normalized measurement values ​​of the above three dimensions are then weighted and fused with the importance weights of each feature according to the task context to finally output the comfort impact value, which is used to calculate the final task priority score.

6. The method as described in claim 1, characterized in that, In step S3, a style decision model is introduced into the task scheduling process. Multimodal feature matching is used to determine the matching relationship between the user's current state and the task context, enabling dynamic style selection and personalized scheduling strategy matching. The input is vectorized to extract user speech fundamental frequency variance and text sentiment polarity values, and the task scenario complexity is quantified. The input vector is compared with a pre-trained style vector library for similarity, and the top three candidate styles are selected. The scheduling style is then determined while meeting safety threshold requirements. Subsequently, a style verification mechanism is used to further verify the accuracy and stability of the style selection, thereby determining the adjustment requirements for the current task dimension weights. And / or, Once the style is determined, a non-linear weight adjustment mechanism is used to map the demand weight vector of the current task based on the response preferences of different styles to the four dimensions of safety, timeliness, social interaction and comfort, so as to achieve a fine integration of user style, task urgency and environmental disturbance. And / or, The style selection is constrained based on the user's occupational tags, and the user's task scheduling intervention behavior is extracted to form an individualized style weight offset function, so as to maintain the consistency between the user's personalized strategy in the execution of long-term tasks and the short-term response. And / or, When different task requirements trigger style strategy conflicts, a covariance matrix of requirements features is constructed to evaluate the synergistic or antagonistic relationship between the four-dimensional requirements. For conflict scenarios that trigger conflict resolution mechanisms, the weight combination that satisfies the maximum weighted requirement utility without triggering serious conflicts is solved, and / or, the personal comfort weight is zeroed, and / or, the social dimension weight is proportionally reduced to ensure the stability and convergence of the overall scheduling behavior in the multi-dimensional task scheduling system.

7. The method as described in claim 1, characterized in that, In step S4, the uncertainty of task requirements and the subjectivity of user style preferences are integrated through a fuzzy entropy weight fusion algorithm, including: The four-dimensional requirement features of the input are fuzzified to construct the corresponding triangular fuzzy number representation. The objective weight of the fuzzy feature is calculated based on the information entropy of each fuzzy feature. The objective weight result is fused with the subjective weight, and the priority is calculated using a weighted hybrid strategy.

8. The method as described in claim 1, characterized in that, The dynamic priority scheduling scheme is integrated with the rule base scheme or expert system scheme and applied to hybrid scenarios to adapt to differences in system performance.

9. A scheduling and coordination system implementing the method as described in any one of claims 1-8, characterized in that, The system includes: a multi-task input and environmental state perception module, a multi-dimensional demand feature extraction module, a dynamic weight generation module, a task priority quantification calculation module, and a multi-model collaborative scheduling and execution module; The multi-task input and environmental state perception module is used to acquire task sets and environmental parameters; The multi-dimensional demand feature extraction module is used to extract the impact values ​​of security demand level, timeliness threshold, social relevance, and personal comfort. The dynamic weight generation module is used to generate dynamic quadruple weight vectors based on user profiles and historical behavior. The task priority quantization calculation module is used to perform tensor operations on the requirement features and weight vectors and output the task priority score. The multi-model collaborative scheduling and execution module is used to schedule tasks to real-time response models or resource optimization models according to priority, and dynamically adjust the execution queue.

10. The method as described in any one of claims 1-8, or the system as described in claim 9, in applications including smart home control and production line task scheduling.

Images