Intelligent autonomous decision-making method based on end-side AI large model

Abstract

The application discloses an intelligent autonomous decision-making method based on an end-side AI large model, and belongs to the technical field of computers. The internal operation state of a terminal device, the task logic structure of an AI large model, external environment perception information and user instructions are comprehensively considered, and on this basis, the importance of task logic elements is adaptively modulated and calibrated in terms of environment and resource sensitivity, and then differentiated elastic execution strategies are matched for the logic elements to guide the AI large model to execute decisions. Thus, the application can solve the problem of reduced behavior stability and decision reliability of the existing end-side AI large model under uncertain conditions such as resource limitation, dynamic environment and task complexity, and ensure that the key functions remain stable in long-term operation, thereby improving the system-level resilience of the end-side intelligent autonomous decision-making system.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to an intelligent autonomous decision-making method based on a large edge AI model. Background Technology

[0002] Deploying large-scale AI models on edge devices to enable real-time autonomous decision-making in complex scenarios is an important development direction in the field of edge intelligence. Such applications typically face multiple real-world constraints, including severely limited computing power and energy consumption at the terminal, highly dynamic external physical environments, and the complex and ever-changing logic of the decision-making tasks themselves.

[0003] When dealing with these constraints, related technologies often suffer from insufficient inter-module coordination. Resource management relies solely on static or simple dynamic adjustments based on the internal state of the equipment, failing to comprehensively assess external environmental risks and task criticality. Environmental perception information is typically used only as model input data, and its deeper semantic risks are not translated into direct basis for system resource scheduling. Furthermore, the lack of mechanisms for autonomous learning and evolution from long-term operational experience makes it unable to adapt to equipment performance degradation and the challenges of new scenarios.

[0004] The aforementioned shortcomings lead to a significant decrease in the behavioral stability and decision-making reliability of existing edge-side intelligent decision-making systems when dealing with resource fluctuations, sudden environmental changes, and complex tasks, making it difficult to achieve long-term robust operation while ensuring key functions. Therefore, how to achieve system-level resilient edge-side autonomous decision-making under multiple uncertainties has become an urgent technical challenge to be solved. Summary of the Invention

[0005] This application provides an intelligent autonomous decision-making method based on a large edge AI model, the technical solution of which is as follows: On the one hand, an intelligent autonomous decision-making method based on a large edge AI model is provided, the method comprising: The system acquires and processes the internal operating status data of the terminal device, the logical structure data of the decision-making tasks to be executed by the AI ​​big model, the external environment perception data, and the user instruction data, and obtains resource constraint features, task logic features, and environmental semantic features, respectively. Based on the environmental semantic features, the importance of each logical element in the task logical features is modulated to generate an environment-adaptive importance weight. Based on the resource constraint features, the resource sensitivity of the environment-adaptive importance weight is calibrated to generate a resilient scheduling weight and a resource allocation focus signal. Based on the resilient scheduling weights, the resource allocation focus signals, the task logic features, and the environmental semantic features, differentiated elastic execution strategies are matched for each logical element of the decision task. The AI ​​large model is controlled to make execution decisions based on the elastic execution strategy.

[0006] On the one hand, an intelligent autonomous decision-making system based on a large edge AI model is provided, the system comprising: The acquisition module is used to acquire and process the internal operating status data of the terminal device, the logical structure data of the decision-making tasks to be executed by the AI ​​big model, the external environment perception data, and the user instruction data, and obtain resource constraint features, task logic features, and environmental semantic features respectively. The modulation module is used to modulate the importance of each logical element in the task logic feature based on the environmental semantic features, generate environmentally adaptive importance weights, and perform resource sensitivity calibration on the environmentally adaptive importance weights based on the resource constraint features, to generate resilient scheduling weights and resource allocation focusing signals. The matching module is used to match differentiated elastic execution strategies for each logical element of the decision task based on the resilience scheduling weight, the resource allocation focus signal, the task logical features, and the environmental semantic features. The execution module is used to control the execution decisions of the AI ​​large model according to the elastic execution strategy.

[0007] On one hand, a computer device is provided, the computer device including one or more processors and one or more memories, the one or more memories storing at least one computer program, the computer program being loaded and executed by the one or more processors to implement the intelligent autonomous decision-making method based on edge AI large model.

[0008] On the one hand, a computer-readable storage medium is provided, wherein at least one computer program is stored in the computer-readable storage medium, the computer program being loaded and executed by a processor to implement the intelligent autonomous decision-making method based on the edge AI large model.

[0009] On the one hand, a computer program product or computer program is provided, which includes program code stored in a computer-readable storage medium. The processor of a computer device reads the program code from the computer-readable storage medium and executes the program code, causing the computer device to execute the aforementioned intelligent autonomous decision-making method based on a large edge AI model. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of the implementation environment of an intelligent autonomous decision-making method based on a large edge AI model provided in this application embodiment; Figure 2 This is a flowchart of an intelligent autonomous decision-making method based on a large edge AI model provided in an embodiment of this application; Figure 3 This is a flowchart of another intelligent autonomous decision-making method based on a large edge AI model provided in this application embodiment; Figure 4 This is a flowchart of another intelligent autonomous decision-making method based on a large edge AI model provided in this application embodiment; Figure 5 This is a flowchart of another intelligent autonomous decision-making method based on a large edge AI model provided in this application embodiment; Figure 6 This is a flowchart of another intelligent autonomous decision-making method based on a large edge AI model provided in this application embodiment; Figure 7 This is a flowchart of another intelligent autonomous decision-making method based on a large edge AI model provided in this application embodiment; Figure 8 This is a schematic diagram of the structure of an intelligent autonomous decision-making system based on a large edge AI model provided in an embodiment of this application. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0013] In this application, the terms "first," "second," etc., are used to distinguish identical or similar items with essentially the same function. It should be understood that there is no logical or temporal dependency between "first," "second," and "nth," nor are there any restrictions on quantity or execution order.

[0014] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0015] Figure 1 This is a schematic diagram illustrating the implementation environment of an intelligent autonomous decision-making method based on a large edge AI model provided in this application embodiment. See also... Figure 1 The implementation environment may include node 110 and system 140.

[0016] Node 110 is connected to system 140 via a wireless or wired network. Optionally, node 110 can be a smartphone, tablet, laptop, desktop computer, etc., but is not limited to these. Node 110 has an application installed and running that supports intelligent autonomous decision-making based on a large edge AI model.

[0017] System 140 is a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms. System 140 can provide background services for applications running on node 110.

[0018] Traditional edge AI large-scale model deployment solutions suffer from several problems under conditions such as limited terminal computing power and energy consumption, dynamic changes in the external physical environment, and complex decision-making task logic. These problems include insufficient inter-module coordination, inadequate resource management and adjustment, failure to translate environmental perception information into system resource scheduling basis, and a lack of autonomous learning and evolution mechanisms. As a result, existing systems exhibit reduced behavioral stability and decision-making reliability when dealing with resource fluctuations, sudden environmental changes, and complex tasks, making it difficult to achieve long-term operation while ensuring critical functions.

[0019] To address this, this application proposes an intelligent autonomous decision-making method based on an edge-side AI large-scale model. This method includes: acquiring and processing internal operating status data of the terminal device, logical structure data of the decision-making task to be executed by the AI ​​large-scale model, external environment perception data, and user instruction data, respectively obtaining resource constraint features, task logical features, and environmental semantic features. Based on the environmental semantic features, importance modulation is applied to each logical element in the task logical features to generate environment-adaptive importance weights; and based on the resource constraint features, resource sensitivity calibration is performed on the environment-adaptive importance weights to generate resilient scheduling weights and resource allocation focusing signals. Based on the resilient scheduling weights, resource allocation focusing signals, task logical features, and environmental semantic features, differentiated elastic execution strategies are matched to each logical element of the decision-making task. The AI ​​large-scale model is then controlled to execute decisions according to the elastic execution strategies.

[0020] For ease of understanding, the following explains some key terms in this embodiment: Resource constraint features are used to characterize the internal operating status of terminal devices, such as quantitative indicators like computing power, memory usage, power consumption, and device health. These features provide the foundational data for subsequent resource sensitivity calibration.

[0021] Task logic features are used to describe the internal structure and logical relationships of the decision-making tasks to be performed by the large AI model, such as the various logical elements contained in the task, the dependencies between these elements, and their initial importance assessment. This feature provides structured information at the task level for subsequent importance modulation and policy matching.

[0022] Environmental semantic features are used to reflect perceived information about the external environment and the intent of user instructions. These features can include semantic assessment results such as environmental risk level, task scenario criticality, and the urgency of user intent, providing context-awareness capabilities for the decision-making process.

[0023] The environment-adaptive importance weight is generated by dynamically adjusting the initial importance of task logical features in conjunction with environmental semantic features. This weight reflects the relative importance of each logical element in a decision-making task within a specific environmental context.

[0024] The resilience scheduling weight is generated by further calibrating the environmentally adaptive importance weights, taking into account the resource constraints of terminal devices. This weight aims to ensure that the system prioritizes the execution of critical logical elements when resources are limited or fluctuating, thereby improving decision-making resilience.

[0025] The resource allocation focus signal is generated based on the distribution of resilience scheduling weights and resource constraints. This signal is used during the resource allocation process to prioritize key logical elements with higher resilience scheduling weights, guiding resources towards these key elements.

[0026] A flexible execution strategy is an execution plan customized for each logical element in a decision-making task and dynamically adjusted according to actual conditions. This strategy can include parameters such as computational precision, computational unit allocation, and memory caching strategies, aiming to enable large AI models to complete decision-making tasks optimally under varying environmental and resource conditions.

[0027] For the intelligent autonomous decision-making method described in this application, please refer to [link / reference needed]. Figure 2 Specifically, this may include the following steps: 201. Acquire and process the internal operating status data of the terminal device, the logical structure data of the decision-making tasks to be executed by the AI ​​big model, the external environment perception data, and the user instruction data to obtain resource constraint features, task logic features, and environmental semantic features, respectively.

[0028] Specifically, the internal operating status data of terminal devices can be collected periodically, such as obtaining real-time CPU utilization, memory usage, and battery level through APIs provided by the operating system. This raw data can be directly used as a component of resource constraint features, or processed through simple statistical methods such as averaging and maximum values. For example, when CPU utilization exceeds a certain fixed threshold, it can be marked as "resource strain." The logical structure data of the decision-making tasks to be executed by the AI ​​large-scale model can be predefined as a series of sequentially executed steps, each assigned a fixed importance level. These steps and their preset importance levels together constitute the task's logical features. For example, an image recognition task can be decomposed into steps such as "image preprocessing," "feature extraction," and "classification judgment," and an initial importance value can be manually assigned to each step. External environment perception data can be directly mapped to preset environmental state labels; for example, when a light sensor detects low light, the environmental state is marked as "dim." User command data can be identified through keyword matching to identify a few preset command types, and each command type can be assigned a fixed urgency level. The combination of these labels and urgency levels constitutes the environmental semantic features. For example, when a user command contains the word "urgent," its urgency level is set to the highest.

[0029] 202. Based on environmental semantic features, the importance of each logical element in the task logical features is modulated to generate environmentally adaptive importance weights. Based on resource constraint features, the resource sensitivity of the environmentally adaptive importance weights is calibrated to generate resilient scheduling weights and resource allocation focusing signals.

[0030] Specifically, environmental semantic features can be used to linearly adjust the initial importance of each logical element in the task's logical features. For example, when the environmental risk level is high, the initial importance weights of all logical elements can be uniformly multiplied by a fixed gain coefficient to generate environmentally adaptive importance weights. Alternatively, a weight adjustment factor can be obtained from a pre-defined lookup table based on a single indicator in the environmental semantic features (e.g., environmental risk level) and applied to the initial importance weights of all logical elements. Building upon this, resource constraint features can be used to perform simple thresholding on the environmentally adaptive importance weights. For example, when resource availability is below a certain fixed threshold, all environmentally adaptive importance weights can be uniformly reduced by a fixed percentage, or weights below a certain secondary threshold can be set to zero to generate resilient scheduling weights. Thus, the resource allocation focus signal can be generated as a set of binary tags; for example, logical elements with resilient scheduling weights above a certain fixed threshold can be marked as "high priority," and the rest as "normal priority."

[0031] 203. Based on resilient scheduling weights, resource allocation focus signals, task logical characteristics, and environmental semantic characteristics, differentiated elastic execution strategies are matched for each logical element of the decision task.

[0032] Specifically, resilience scheduling weights, resource allocation focus signals, task logic features, and environmental semantic features can be input into a rule-based decision engine. This engine matches a predefined resilient execution strategy to each logical element based on a predefined set of rules, such as "if the resilience scheduling weight is high and the resource allocation focus signal is high priority, then a high-precision strategy is selected; if the resilience scheduling weight is low and resources are sufficient, then a low-power strategy is selected." For example, for the "feature extraction" logical element in an image processing task, when its resilience scheduling weight is high and the environmental semantic features indicate high risk, a high-precision strategy using floating-point calculations can be matched. Conversely, when its resilience scheduling weight is low and the environmental semantic features indicate low risk, a low-precision strategy using fixed-point calculations can be matched.

[0033] 204. Control the execution decisions of the AI ​​big model based on the flexible execution strategy.

[0034] Specifically, the flexible execution strategy can be resolved into a series of configuration instructions for the AI ​​large-scale model inference engine. These instructions may include setting a fixed computational precision for specific logical elements (e.g., 8-bit integers or 16-bit floating-point numbers), specifying the computing units to be used (e.g., CPU or NPU), and allocating a preset memory cache size. When executing decisions, the AI ​​large-scale model will strictly follow these preset configuration parameters. For example, for a logical element assigned a "low-power strategy," its computation will be forced to execute in low-power mode, even if more powerful computing resources are available at that time.

[0035] This application comprehensively considers the internal operating status of the terminal device, the task logic structure of the AI ​​big data model, external environmental perception information, and user instructions. Based on this, it performs environmental adaptive modulation and resource sensitivity calibration on the importance of task logic elements, and then matches differentiated elastic execution strategies to each logic element to guide the AI ​​big data model's execution decisions. Therefore, this application can solve the problem of reduced behavioral stability and decision reliability of existing edge-side AI big data models under uncertain conditions such as resource constraints, dynamic environments, and complex tasks, ensuring the robustness of key functions in long-term operation and improving the system-level resilience of edge-side intelligent autonomous decision-making systems.

[0036] In some of the solutions mentioned above in this application, it is proposed to acquire and process the internal operating status data of the terminal device, the logical structure data of the decision-making tasks to be executed by the AI ​​large model, the external environment perception data, and the user instruction data to generate resource constraint features, task logic features, and environmental semantic features to support subsequent weight modulation and decision execution. However, in this process, the existing methods fail to effectively integrate the comprehensive analysis of internal resource status, task logic structure, and external environmental risks, resulting in feature extraction being too isolated and unable to capture the dynamic changes of device performance degradation trends, task dependencies, and environmental risks, thereby affecting the accuracy and robustness of subsequent adaptive modulation and resource scheduling.

[0037] In response, this application proposes a more refined feature extraction method to overcome the above limitations. By performing in-depth analysis and semantic processing on multi-source heterogeneous data, it generates more insightful resource constraint features, task logic features, and environmental semantic features.

[0038] Specifically, see Figure 3 The method includes: 301. Perform predictive analysis and performance degradation analysis on the internal operating status data to generate quantitative indicators, including predicted resource availability values ​​and equipment health assessment values, to form the resource constraint characteristics.

[0039] This process aims to comprehensively assess the operational status of terminal devices from both time and health dimensions. For example, by deploying time-series prediction models based on recurrent neural networks (RNNs) or long short-term memory networks (LSTMs), real-time resource monitoring sequences such as CPU utilization, memory usage, and battery level can be learned and predicted to obtain resource availability forecasts for a future period. Simultaneously, machine learning models such as support vector machines (SVMs) or Gaussian process regression (GPRs) can be used, combined with historical operating parameters of the device (such as temperature, error logs, and sensor readings), to model performance degradation trends, assess the current health of the device, predict potential failure risks, and generate a device health assessment value. These quantitative indicators collectively constitute a dynamic and forward-looking description of the device's resource status.

[0040] 302. Perform causal logic modeling and importance analysis on the logical structure data to generate logical relationships and weight information, including the dependencies between each step in the decision-making task and the initial importance weights, so as to form the logical characteristics of the task.

[0041] This step aims to gain a deep understanding of the internal structure and key components of AI large-scale model decision-making tasks. For example, knowledge graph-based or ontological methods can be used to explicitly model the various logical elements of the decision-making task (such as data collection, feature extraction, model inference, and result output) and their causal and sequential relationships, constructing a directed acyclic graph (DAG). Based on this, an initial importance weight can be assigned to each logical element according to domain expert experience, historical task execution data, or graph structure analysis (such as the PageRank algorithm and betweenness centrality) to reflect its criticality in the entire decision-making task. This information collectively depicts the structured blueprint of the task.

[0042] 303. Perform risk semantic recognition, scenario criticality assessment and intent parsing on the external environment perception data and the user instruction data to generate semantic assessment results including environmental risk level, task scenario criticality and user intent urgency, so as to form the environmental semantic features.

[0043] This process aims to transform raw sensory data and commands into high-level semantic information, enabling the system to understand the complexity of the external world and user needs. For example, multimodal fusion technology can be used to integrate heterogeneous data from sensors such as cameras, radar, lidar, and microphones, and risk semantic recognition can be performed using deep learning models (such as Transformer networks) to determine whether there are potential dangers in the current environment (such as pedestrian intrusion, abnormal objects, or severe weather), thereby outputting the environmental risk level. Simultaneously, based on Geographic Information System (GIS) data, pre-set rule bases, or scene classification models, the criticality of the physical scene in which the current task is located can be assessed (such as highways, urban intersections, and factory workshops), generating the criticality of the task scene. For user command data, Natural Language Processing (NLP) techniques, such as intent recognition and sentiment analysis based on pre-trained models like BERT or GPT, can be used to analyze the deep intent and urgency of user commands. These semantic evaluation results collectively provide rich and understandable external contextual information for large AI models.

[0044] Through the above technical solutions, this application solves the problems of isolated feature extraction and inability to capture dynamic changes in existing methods. Specifically, by performing predictive analysis and performance degradation analysis on internal operating status data, the system can detect resource bottlenecks and equipment health risks in advance, thereby avoiding decision-making errors caused by information lag during resource scheduling and ensuring the stable operation of the AI ​​large model in resource-constrained environments. Causal logic modeling and importance analysis of logical structure data enable the system to clearly identify critical paths and core steps in decision-making tasks, providing a solid foundation for subsequent weight modulation and flexible execution strategy matching, and avoiding blind optimization of tasks. Furthermore, by performing risk semantic recognition, scenario criticality assessment, and intent parsing on external environment perception data and user command data, the system can understand the complexity of the external environment and the urgency of user needs in real time and accurately. This allows the AI ​​large model to dynamically adjust its behavior according to the actual situation. For example, when facing high-risk environments or urgent user commands, it prioritizes the execution of critical tasks, thereby improving the adaptability, reliability, and resilience of the edge AI large model's autonomous decision-making. This comprehensive feature extraction method provides comprehensive and high-quality input for subsequent environmentally adaptive importance weight generation, resource sensitivity calibration, and flexible execution strategy matching, ensuring the robustness and efficiency of the entire decision-making process.

[0045] To address this, this application further proposes predictive and performance degradation analysis of internal operating status data to form resource constraint characteristics. In some embodiments described above, this application proposes predictive and performance degradation analysis of internal operating status data to form resource constraint characteristics. However, in this process, existing methods may only perform simple static analysis or single-index evaluation, failing to effectively predict dynamic changes in resource availability and long-term degradation trends in equipment performance. This results in inaccurate and insufficiently robust resource constraint characteristics, unable to provide reliable constraint information when resource fluctuations are frequent or equipment aging intensifies, thus affecting the adaptability of subsequent decisions and system resilience.

[0046] To address this, this application further proposes predictive and performance degradation analysis of the internal operating status data to generate quantitative indicators, including predicted resource availability values ​​and equipment health assessment values, to form the resource constraint characteristics, specifically including: The real-time resource monitoring sequence in the internal operating status data is processed by time-series prediction to generate the resource availability prediction value.

[0047] The historical operating parameters of the equipment in the internal operating status data are modeled to reflect performance degradation trends, and a health assessment value for the equipment is generated.

[0048] The resource availability prediction value and the equipment health assessment value are weighted and fused to obtain the resource constraint characteristics.

[0049] This involves performing time-series prediction processing on the real-time resource monitoring sequence within the internal operational status data to generate a resource availability prediction value. The aim is to predict resource availability over a future period based on continuous records of resource usage such as CPU utilization, memory usage, network bandwidth, and battery level of terminal devices over a specific time period, thus providing forward-looking resource status information. For example, an ARIMA (Autoregressive Integral Moving Average) model or exponential smoothing can be used to model the real-time resource monitoring sequence, analyzing the sequence's trends, seasonality, and random fluctuations to predict future resource availability. Alternatively, deep learning methods, such as Long Short-Term Memory (LSTM) networks or Gated Recurrent Units (GRUs), can be used to train the resource monitoring sequence to learn complex temporal dependencies, thereby generating a high-precision resource availability prediction value.

[0050] This process models the performance degradation trends of historical operating parameters from the internal operational status data of the equipment, generating a health assessment value. The aim is to establish a mathematical model of equipment performance changes over time by analyzing the changing trends of historical operating parameters (such as temperature, power consumption, error rate, and sensor readings), thereby assessing the equipment's health. For example, statistical learning methods such as regression analysis (linear regression, multinomial regression) or support vector regression (SVR) can be used to model the relationship between historical operating parameters and equipment performance indicators, predict the downward trend of equipment performance, and generate a health assessment value accordingly. Alternatively, physical model-based or data-driven machine learning methods (such as random forests and gradient boosting trees) can be used, combined with equipment failure modes and historical maintenance records, to construct an equipment health index model. This model, by inputting current operating parameters, outputs a quantitative health assessment value.

[0051] The resource constraint feature is obtained by weighted fusion of the predicted resource availability value and the equipment health assessment value. This aims to assign different weights to two or more quantitative indicators from different sources based on their importance or reliability, and then combine them to obtain a more comprehensive and robust integrated indicator. Here, the goal is to comprehensively consider short-term resource fluctuations and long-term equipment health status to form a more accurate resource constraint feature. For example, linear weighted fusion can be used: Resource Constraint Feature = w1 × Predicted Resource Availability Value + w2 × Equipment Health Assessment Value, where w1 and w2 are preset weights, and w1 + w2 = 1. These weights can be determined based on experience, expert knowledge, or optimization using historical data. Alternatively, adaptive weighted fusion can be used, dynamically adjusting the weights according to the characteristics of the current environment or task. For example, when resources are extremely scarce or the task is urgent, the weight of the predicted resource availability value can be increased. In scenarios with severe equipment aging or long-term operation, the weight of the equipment health assessment value can be increased. This can be achieved through fuzzy logic, reinforcement learning, or rule-based systems.

[0052] Through the above technical solutions, this application addresses the limitations of existing methods in the formation of resource constraint characteristics. By performing time-series prediction processing on real-time resource monitoring sequences, the system can proactively observe the dynamic changes in resource availability, thereby providing early warnings of potential resource shortages or overload risks, enabling subsequent decision-making to be proactive rather than reactive. Simultaneously, by modeling the performance degradation trends of historical equipment operating parameters, the system can quantitatively assess the long-term health status and performance degradation trends of the equipment, avoiding unexpected performance declines due to equipment aging, thus providing a more stable basis for decision-making. By weightedly fusing the predicted resource availability value with the equipment health assessment value, the system comprehensively considers short-term resource fluctuations and long-term equipment health status, overcoming the one-sidedness of single-indicator assessments and providing a more comprehensive, dynamic, and robust resource constraint characteristic. This allows the edge AI model to make intelligent autonomous decisions based on more accurate and reliable resource information, improving its adaptability, stability, and system resilience in complex dynamic environments such as frequent resource fluctuations and accelerated equipment aging, ensuring that key functions can still operate robustly under multiple uncertainties.

[0053] In some of the solutions mentioned above in this application, causal logic modeling and importance analysis are proposed to generate task logic features. However, in the process of implementation, the identification of dependencies and importance weights may not be accurate enough, and the complex logical structure of the decision-making task cannot be effectively captured, resulting in inaccurate task logic features and affecting the stability of subsequent decisions.

[0054] To address this, this application further proposes causal logic modeling and importance analysis of the logical structure data to generate logical relationships and weight information, including dependencies and initial importance weights among the steps in the decision-making task, to form the logical features of the task. Specifically, this includes: performing causal deconstruction analysis on the logical structure data to identify multiple logical elements included in the decision-making task and the causal dependencies between these elements; constructing a directed acyclic graph (DAG) to represent the reasoning logic of the decision-making task based on these causal dependencies; assigning initial importance weights based on historical data or domain knowledge to each logical element in the DAG; and defining the DAG and the associated initial importance weights together as the logical features of the task.

[0055] Specifically, causal deconstruction analysis is performed on the logically structured data to identify the multiple logical elements contained in the decision-making task and the causal dependencies between these elements. The aim is to deeply analyze the internal structure of the decision-making task, decomposing it into manageable logical elements with clearly defined functional boundaries, and revealing the causal relationships between these elements. Identifying causal dependencies is fundamental to understanding the task execution process, predicting potential impacts, and optimizing decision paths. For example, Natural Language Processing (NLP) techniques can be used to perform semantic analysis on the task description text, combined with a predefined task pattern library or expert knowledge graph, to automatically identify verbs, noun phrases, etc., in the task as potential logical elements, and determine the causal relationships between them through dependency parsing or event extraction techniques. Alternatively, domain experts can analyze a large number of historical task execution logs through manual annotation or semi-supervised learning, manually or with assistance, identifying the atomic operations (logical elements) of the task and their preconditions and post-effects, thereby constructing a causal relationship model.

[0056] Based on this causal dependency, a directed acyclic graph (DAG) is constructed to represent the reasoning logic of the decision-making task. A DAG is a powerful data structure that can clearly and unambiguously represent the order, parallelism, and dependencies between logical elements in a task, while avoiding logical deadlocks or infinite loops caused by circular dependencies. Constructing a DAG helps visualize the task flow, facilitating subsequent scheduling, optimization, and execution. For example, after identifying all logical elements and their causal dependencies, each logical element can be treated as a node in the graph, and the causal dependency can be represented as a directed edge from the cause node to the result node. Algorithms such as topological sorting can verify the acyclicity of the graph and ensure the reasonable order of task execution. Furthermore, graph databases (such as Neo4j) can be used to store and manage these logical elements and dependencies. The native query language and algorithms of graph databases can efficiently construct and manipulate DAGs and support complex path analysis and pattern matching.

[0057] Furthermore, each logical element in the directed acyclic graph is assigned an initial importance weight based on historical data or domain knowledge. This initial importance weight reflects the inherent value or criticality of each logical element in the overall decision-making task. These weights form the basis for subsequent environmental adaptive modulation and resource sensitivity calibration, ensuring that the system prioritizes the execution of critical logical elements when resources are limited or the environment changes. For example, based on historical data, initial importance weights can be learned and assigned to each logical element using machine learning models (such as regression or classification models) by analyzing past task execution success rates, their impact on decision outcomes, resource consumption, or user feedback. Alternatively, combining domain knowledge, domain experts can set predefined importance levels or weight values ​​for different logical elements based on their experience and understanding of the task. For example, in an autonomous driving task, "obstacle recognition" might be assigned a higher initial importance weight than "playing music."

[0058] The directed acyclic graph (DAG) and its associated initial importance weights are collectively defined as the task's logical features. This step integrates the task's structured logical representation (DAG) with the intrinsic value of each logical element (initial importance weights) to form a comprehensive "task logical feature" suitable for subsequent processing. This integration ensures that task execution considers not only the order of steps but also the relative importance of each step, providing richer and more refined input for intelligent autonomous decision-making. For example, the DAG can be represented as an adjacency matrix or adjacency list, and the initial importance weights can be stored as attribute lists or vectors corresponding to graph nodes. These two data structures can be encapsulated in a unified data object or class as a complete representation of the task's logical features. Alternatively, graph embedding techniques can be used to encode the structural information of the DAG and the initial importance weights of the nodes into a low-dimensional vector representation. This vector representation captures the graph's topology and node attributes, facilitating subsequent processing and analysis by machine learning models.

[0059] Through the above technical solution, this application can accurately model and quantify the logical structure of decision-making tasks. By performing causal deconstruction analysis on the logical structure data, it can comprehensively identify the logical elements and their causal dependencies in the decision-making task, avoiding the omission or ambiguity of dependencies that may exist in traditional methods, thus laying a solid foundation for subsequent logical reasoning. A directed acyclic graph (DAG) is constructed based on the identified causal dependencies, intuitively and unambiguously representing the reasoning logic of the decision-making task, effectively preventing circular dependencies and ensuring the clarity and executability of the task flow. Each logical element in the DAG is assigned an initial importance weight based on historical data or domain knowledge, making the importance assessment of each logical element objective and avoiding subjective assumptions, providing a reliable starting point for subsequent weight adjustment and resource calibration. The DAG and the associated initial importance weights are jointly defined as task logical features. This integration not only provides a structured view of the task but also quantifies the criticality of each component, making the generated task logical features more accurate and complete. This improves the accuracy of task logic features, thus providing high-quality input for subsequent importance modulation based on environmental semantic features and resource sensitivity calibration based on resource constraint features, thereby enhancing the stability and reliability of the entire intelligent autonomous decision-making method.

[0060] In some of the solutions mentioned above in this application, risk semantic recognition, scene criticality assessment and intent parsing are proposed to generate environmental semantic features. However, in the process of implementation, the deep semantic risks of environmental perception data may not be fully extracted and utilized, resulting in environmental semantic features that are not accurate and comprehensive enough to effectively support subsequent decision-making.

[0061] To this end, this application further proposes to perform risk semantic recognition, scenario criticality assessment, and intent parsing on external environment perception data and user instruction data, generating semantic assessment results including environmental risk level, task scenario criticality, and user intent urgency, in order to form environmental semantic features, specifically including: Multimodal feature extraction and fusion processing are performed on the external environment perception data to generate a comprehensive environmental feature vector.

[0062] The comprehensive feature vector of the environment is input into the micro-risk identification model deployed on the edge, and the environmental risk level is output.

[0063] The comprehensive feature vector of the environment is matched and analyzed with the preset scenario safety rule base to output the criticality of the task scenario.

[0064] The user command data is subjected to real-time semantic analysis to determine the urgency of the user's intent.

[0065] The environmental risk level, the criticality of the task scenario, and the urgency of the user's intent together constitute the semantic features of the environment.

[0066] This process involves multimodal feature extraction and fusion of the external environment perception data to generate a comprehensive environmental feature vector. The aim is to extract and integrate a unified and comprehensive feature representation from heterogeneous external environment perception data. This avoids the information limitations of single-modal data and enhances the depth of environmental understanding. For example, deep learning-based fusion methods can be employed, such as using convolutional neural networks (CNNs) to process image data and recurrent neural networks (RNNs) to process time-series sensor data, fusing features from different modalities through attention mechanisms or feature splicing layers. Alternatively, methods based on traditional signal processing and feature engineering can be used, such as performing Fourier transforms on audio data to extract spectral features, performing target detection and tracking on radar data, and then statistically aggregating and vectorizing these features.

[0067] The comprehensive feature vector of the environment is input into a miniature risk identification model deployed on the edge, which outputs the risk level of the environment. The aim is to efficiently and accurately assess the risk level of the current environment on resource-constrained terminal devices. This miniature risk identification model can be implemented using various lightweight techniques. For example, large risk identification models can be compressed into smaller models suitable for edge deployment through model pruning, quantization (such as 8-bit or 4-bit quantization), or knowledge distillation. Alternatively, compact and computationally inexpensive shallow machine learning models, such as support vector machines (SVM) or decision trees, can be directly used. These models are optimized for specific risk patterns during training to ensure inference efficiency on the edge.

[0068] The comprehensive environmental feature vector is matched and analyzed against a pre-defined scenario security rule base to output the scenario criticality of the task. This aims to objectively assess the importance or sensitivity of the current task's scenario based on predefined business logic or security specifications. This matching analysis can be implemented using an expert system or rule engine. For example, by defining a series of "IF-THEN" rules, specific feature values ​​or combinations of features in the comprehensive environmental feature vector can be mapped to different scenario criticality levels. Alternatively, a fuzzy logic system can be used, employing fuzzy rules and membership functions to handle the fuzziness of environmental features, thereby outputting continuous task scenario criticality values.

[0069] Real-time semantic analysis of user command data determines the urgency of the user's intent. Its purpose is to quickly and accurately understand the intent and urgency implied in the user command, enabling the system to respond promptly. This real-time semantic analysis can employ lightweight natural language processing (NLP) techniques, such as keyword matching, syntactic analysis, or fine-tuning based on small-scale pre-trained language models (e.g., MobileBERT, TinyBERT), to identify core verbs, time adverbs, or sentiment tendencies in the command and map them to predefined levels of user intent urgency. Alternatively, rule-based regular expression matching can be used to quickly identify expressions of urgency in the command.

[0070] The environmental risk level, the criticality of the task scenario, and the urgency of the user's intent collectively constitute the environmental semantic features. The aim is to integrate multi-dimensional semantic assessment results into a unified and comprehensive feature representation, providing rich and accurate contextual information for subsequent decision-making tasks. This can be achieved through simple feature concatenation, directly linking the numerical values ​​or codes of each assessment result into a high-dimensional vector. Alternatively, a small neural network (such as a multilayer perceptron) can be used to non-linearly combine these assessment results to learn the complex relationships between them, thereby generating more expressive environmental semantic features.

[0071] Through the above technical solutions, this application can improve the accuracy and comprehensiveness of environmental semantic features. Specifically, by performing multimodal feature extraction and fusion processing on external environmental perception data, the system can integrate complementary information from different types of environmental data, extract richer semantic features, effectively avoid the limitations of a single data source, and ensure the comprehensiveness of the environmental comprehensive feature vector. Based on this, the environmental comprehensive feature vector is input into a micro-risk identification model deployed on the edge, enabling the system to efficiently identify dynamic risks on resource-constrained terminal devices, providing accurate risk quantification indicators, thereby enhancing the real-time performance and accuracy of environmental risk level assessment. Simultaneously, by matching and analyzing the environmental comprehensive feature vector with a preset scenario security rule base, the system can objectively assess the importance of scenarios according to predefined standards, enhancing the reliability of task scenario criticality assessment. Furthermore, real-time semantic parsing of the user command data ensures that the system can capture changes in user intent in real time and quickly respond to the urgency of user commands. The environmental risk level, the criticality of the task scenario, and the urgency of the user's intent are combined to form the semantic features of the environment, realizing the comprehensive integration of multi-dimensional evaluation results. This provides a solid, accurate, and comprehensive semantic input for subsequent decision-making tasks, thereby solving the problem of the underutilization of deep semantics in environmental perception data and providing a more reliable foundation for the intelligent autonomous decision-making of AI large models in complex and dynamic environments.

[0072] In some of the solutions mentioned above in this application, the importance of each logical element in the task logical features is modulated based on environmental semantic features to generate environmentally adaptive importance weights. However, in this process, the modulation process lacks a dynamic response mechanism for specific environmental risk levels and task scenario criticality, resulting in the generated weights failing to accurately reflect environmental changes and affecting the adaptability of decision-making and the efficiency of resource allocation.

[0073] To address this, this application further proposes an intelligent autonomous decision-making method based on a large-scale edge AI model. This method modulates the importance of each logical element in the task's logical features based on environmental semantic features, generating environment-adaptive importance weights. (See [link to relevant documentation]). Figure 4 The method includes: 401. Based on the environmental risk level and task scenario criticality in the environmental semantic features, perform dynamic gain calculation for semantic perception and generate semantic perception gain factor.

[0074] 402. Based on the semantically aware gain factor, a non-uniform weight modulation field is constructed, which defines a differentiated modulation intensity applicable to different initial importance weights.

[0075] 403. Use the weighted modulation field to perform field-driven modulation on the initial importance weights of each logical element in the logical features of the task, and generate the environment-adaptive importance weights.

[0076] Specifically, the environmental risk level refers to a quantitative indicator obtained after performing risk semantic recognition on external environment perception data and user instruction data. It is used to assess the potential harm or uncertainty that the current environment may cause to the decision-making task to be executed by the AI ​​large model. For example, this level can be a discrete classification value (such as low risk, medium risk, high risk) or a continuous risk index. The task scenario criticality refers to a quantitative indicator obtained after evaluating the scenario criticality of external environment perception data and user instruction data. It is used to measure the importance or urgency of the scenario in which the current decision-making task is located. For example, this criticality can be a numerical value representing the importance of the task, or a classification label (such as non-critical task, general critical task, core critical task). These two indicators together provide important semantic context information for subsequent weight modulation, ensuring that the modulation process can perceive and respond to dynamic changes in the external environment.

[0077] This semantically aware dynamic gain calculation aims to dynamically adjust or amplify the perception of the importance of task logic elements based on the current environment's risk level and the criticality of the task scenario. Its role is to transform abstract environmental semantic information into concrete numerical factors that can be used for weight modulation. One implementation method is to map the environmental risk level and task scenario criticality to a gain value using a pre-defined lookup table or rule-based inference engine. For example, a combination of high risk and high criticality might correspond to a larger gain value, while low risk and low criticality might correspond to a smaller gain value. Another implementation method is to use a machine learning model (such as a small neural network or regression model) with the environmental risk level and task scenario criticality as input, training the model to output a continuous gain factor. This model can be trained using historical data or expert knowledge to learn appropriate gain patterns under different environmental conditions. This semantically aware gain factor is the output of the semantically aware dynamic gain calculation; it is a numerical value used to quantify the strength or direction of the current environmental semantics' modulation of the importance of task logic elements. Its role is to serve as a key input for subsequently constructing a non-uniform weight modulation field, guiding how the modulation field differentially adjusts the initial importance weights of each logic element according to environmental changes. This factor can be a value greater than 1, indicating an amplification of importance; it can also be a value less than 1, indicating a suppression of importance; or, in specific cases, it can be 1, indicating no additional gain.

[0078] The non-uniform weighted modulation field is an abstract concept describing a mechanism that applies different modulation intensities to logical elements with different initial importance weights. Its "non-uniform" nature means that the modulation effect is not simply a proportional scaling of all weights, but rather differentiated based on the numerical value or other properties of the weights themselves. One implementation is to represent the modulation field as a piecewise function or lookup table, where different initial importance weight intervals correspond to different modulation intensities. For example, elements with higher initial weights may be further amplified, while elements with lower initial weights may be moderately suppressed or remain unchanged. Another implementation is to model the modulation field as a continuous non-linear function that takes the initial importance weights as input and outputs a modulated weight value, with the slope or curvature of the function varying across different input intervals, thus achieving non-uniform modulation. This differentiated modulation intensity refers to the different strengths of the modulation applied to logical elements with different initial importance weights within the non-uniform weighted modulation field. This differentiation is key to achieving fine-grained weight adjustment. For example, logic elements with higher initial importance weights may have stronger modulation strengths, giving them a more dominant role in the environment-adaptive weighting. Conversely, logic elements with lower initial importance weights may have weaker modulation strengths, or even remain unchanged, to avoid excessively interfering with their original importance. This differentiated modulation strength can be reflected in the gain coefficient, offset, or curvature of the nonlinear transform of the modulation function.

[0079] Field-driven modulation refers to using a pre-constructed weight modulation field to systematically and holistically adjust the initial importance weights of each logical element in the task's logical features. It emphasizes that the modulation process is driven by the rules or functions of the entire "field," rather than adjusting each logical element in isolation. One implementation is to use the initial importance weight of each logical element as input, substituting it into the function defined by the weight modulation field to directly calculate its modulated weight. Another implementation is to input the initial importance weights of all logical elements as a vector into a matrix transformation or neural network layer. This transformation or network layer internally encodes the rules of the weight modulation field, thus outputting the modulated weight vector of all logical elements. This environment-adaptive importance weight is obtained after modulation with environmental semantic features, reflecting the impact of current external environmental risks and the criticality of the task scenario on the importance of each logical element. Its role is to provide a more accurate and context-aware importance assessment for subsequent resource allocation and flexible execution strategy matching, ensuring that the AI ​​large model can prioritize critical and high-risk related logical elements in dynamic environments. This set of weights is dynamic and updates in real time as the semantic features of the environment change, thus enabling the decision-making process of the AI ​​large model to be environmentally adaptive.

[0080] Through the aforementioned technical solution, this application addresses the lack of dynamic response to specific environmental risk levels and task scenario criticality in related technologies by introducing a mechanism of semantically aware dynamic gain calculation and a non-uniform weight modulation field. Specifically, by performing semantically aware dynamic gain calculation based on the environmental risk level and task scenario criticality in environmental semantic features, a semantically aware gain factor is generated. This enables the system to capture and quantify the risk level of the external environment and the urgency of the task in real time, transforming this key semantic information into operable modulation parameters. Based on this semantically aware gain factor, a non-uniform weight modulation field is constructed. This field can apply differentiated modulation intensities according to logical elements with different initial importance weights, avoiding the insufficient accuracy caused by simple linear scaling and achieving refined and contextualized adjustment of the importance of task logical elements. Using this weight modulation field, the initial importance weights of each logical element in the task logical features are modulated in a field-driven manner to generate environmentally adaptive importance weights. This ensures that the weights accurately reflect the true importance in the current environment, thus providing a more resilient and adaptive scheduling basis for AI large models to perform decision-making tasks on resource-constrained and dynamically changing terminal devices. This mechanism enables large AI models to dynamically adjust the priority of their internal logical elements according to environmental changes, prioritizing the execution of high-risk or high-critical tasks, thereby improving the reliability of decision-making and the system's robust operation in complex and ever-changing environments.

[0081] In some of the embodiments described above in this application, a semantically perceptive dynamic gain calculation based on environmental semantic features is proposed to generate a semantically perceptive gain factor, which is used to construct a weight modulation field to modulate the importance weight. However, in its implementation, how to accurately combine the environmental risk level and the criticality of the task scenario to calculate the dynamic gain, so as to ensure that the gain factor can effectively reflect the environmental semantics and improve the accuracy of weight modulation, and avoid the deviation of subsequent modulation effect due to insufficient quantification of environmental semantics, is a problem that needs to be solved.

[0082] To address this, this application further proposes a method for dynamically calculating semantic-aware gain based on environmental risk level and task scenario criticality in environmental semantic features, generating a semantic-aware gain factor. This includes: obtaining predefined environmental risk gain coefficients and scenario criticality gain coefficients; performing a first modulation operation on the environmental risk level and the environmental risk gain coefficient to obtain an environmental risk modulation component; performing a second modulation operation on the task scenario criticality and the scenario criticality gain coefficient to obtain a scenario criticality modulation component; and fusing the environmental risk modulation component, the scenario criticality modulation component, and a preset gain benchmark value to obtain the semantic-aware gain factor.

[0083] Specifically, predefined environmental risk gain coefficients and scenario criticality gain coefficients are obtained. The environmental risk gain coefficient is a preset parameter used to quantify and adjust the influence of environmental risk levels on the semantic-aware gain factor. Its function is to assign different weights to different types of environmental risks to reflect their potential impact on the execution of decision-making tasks. For example, this coefficient can be a set of fixed values ​​obtained from historical data statistical analysis, corresponding to different risk categories (such as low risk, medium risk, and high risk), or it can be set through expert experience or domain knowledge and stored in a configuration database. The scenario criticality gain coefficient is a preset parameter used to quantify and adjust the influence of task scenario criticality on the semantic-aware gain factor. Its function is to ensure that the importance of the task is fully reflected in critical scenarios, thereby influencing subsequent weight modulation. For example, this coefficient can be set in a graded manner according to the task priority or the degree of impact on the overall system performance, or it can be obtained by training a machine learning model on historical task execution data.

[0084] The environmental risk level and the environmental risk gain coefficient are subjected to a first modulation operation to obtain the environmental risk modulation component. The first modulation operation aims to effectively combine the environmental risk level and the environmental risk gain coefficient to generate an environmental risk modulation component that accurately reflects the impact of environmental risk. The purpose of this operation is to transform the abstract risk level into a quantifiable impact factor. For example, the first modulation operation can use a multiplication operation: Environmental risk modulation component = Environmental risk level × Environmental risk gain coefficient. Alternatively, a nonlinear function (such as the Sigmoid function or ReLU function) can be used to transform the environmental risk level before multiplying it with the gain coefficient to achieve nonlinear amplification or suppression of the risk impact. The environmental risk modulation component is the output of the first modulation operation; it represents the specific contribution of the current environmental risk to the semantically perceived gain factor, providing a standardized measure of risk impact.

[0085] Simultaneously, a second modulation operation is performed on the task scenario criticality and the scenario criticality gain coefficient to obtain the scenario criticality modulation component. This second modulation operation aims to effectively combine the task scenario criticality with the scenario criticality gain coefficient to generate a scenario criticality modulation component that accurately reflects the criticality of the task scenario. The purpose of this operation is to transform the abstract scenario criticality into a quantifiable influence factor. For example, the second modulation operation can use a multiplication operation: Scenario criticality modulation component = Task scenario criticality × Scenario criticality gain coefficient. Alternatively, a piecewise linear function or exponential function can be used to transform the task scenario criticality before multiplying it with the gain coefficient to achieve differentiated influence within different criticality intervals. The scenario criticality modulation component is the output of the second modulation operation; it represents the specific contribution of the current task scenario criticality to the semantic perception gain factor, ensuring that the importance of the task under different critical scenarios is appropriately reflected.

[0086] The semantic-aware gain factor is obtained by fusing the environmental risk modulation component, the scene criticality modulation component, and a preset gain benchmark value. The preset gain benchmark value is an initial or fundamental value for the semantic-aware gain factor, providing a stable starting point for its calculation and ensuring that the semantic-aware gain factor remains at a reasonable level even when both environmental risk and scene criticality are low. For example, the preset gain benchmark value can be a fixed positive number, such as 1.0, or it can be dynamically adjusted according to the default operating mode of the terminal device. The fusion operation aims to organically combine the environmental risk modulation component, the scene criticality modulation component, and the preset gain benchmark value to generate the semantic-aware gain factor. For example, fusion can use a simple weighted summation method: semantic-aware gain factor = preset gain benchmark value + environmental risk modulation component + scene criticality modulation component. Alternatively, a more complex nonlinear fusion function can be used, such as learning and fusion through a small neural network layer. The semantic-aware gain factor is the output of the fusion operation; it is a quantitative indicator that comprehensively reflects the current environmental risk level, task scene criticality, and the urgency of the user's intent, serving as a key input for subsequently constructing a non-uniform weighted modulation field.

[0087] Through the above technical solution, this application can accurately calculate dynamic gain by combining environmental risk level and task scenario criticality, solving the problem of deviation in subsequent weight modulation effect caused by insufficient quantification of environmental semantics. Specifically, by obtaining predefined environmental risk gain coefficients and scenario criticality gain coefficients, a standardized quantitative basis is provided for environmental risk and scenario criticality, avoiding subjective arbitrariness. Through the first and second modulation operations, the environmental risk level and task scenario criticality are transformed into quantifiable environmental risk modulation components and scenario criticality modulation components, respectively, ensuring accurate capture of environmental semantics. These modulation components are fused with a preset gain benchmark value to generate a semantic-aware gain factor. This gain factor comprehensively reflects the risk level of the current environment and the criticality of the task, providing accurate and dynamic input for the subsequent construction of a non-uniform weight modulation field. This enables the AI ​​large model to adjust the importance weights of each logical element more finely and accurately according to changes in actual environmental semantics when performing decision-making tasks, thereby improving the adaptability, reliability, and resilience of decision-making. Especially in edge scenarios with limited resources and dynamically changing environments, it can effectively avoid decision-making errors or resource waste caused by insufficient understanding of environmental semantics.

[0088] In some embodiments described above in this application, a non-uniform weight modulation field is proposed to generate environment-adaptive importance weights. However, in its implementation, the differential modulation intensity is not accurately calculated based on the initial importance weights of the logical elements and the semantic-aware gain factor, resulting in an insufficiently refined and dynamic modulation process that cannot effectively adapt to changes in the importance of different logical elements, thereby affecting the accuracy and environmental adaptability of the weight modulation.

[0089] To address this, this application further proposes a method for constructing a non-uniform weighted modulation field based on a semantically aware gain factor. The specific implementation includes the following steps: Calculating the differentiated modulation intensity corresponding to each logical element based on the semantically aware gain factor and the initial importance weights of each logical element in the task's logical features. Associating and mapping the differentiated modulation intensity of each logical element with its corresponding initial importance weight to form a modulation intensity distribution map. Based on the modulation intensity distribution map, constructing a continuous or piecewise continuous weighted modulation function. This function takes the initial importance weights as input and outputs modulated weight values, thus constituting the non-uniform weighted modulation field.

[0090] The section on "calculating the differentiated modulation intensity corresponding to each logical element based on the semantically aware gain factor and the initial importance weights of each logical element in the task's logical features" aims to determine the degree to which the importance of each logical element should be adjusted based on environmental context information (semantically aware gain factor) and the inherent logical structure of the task (initial importance weights). This differentiated calculation ensures the precision and targeting of the modulation process. Specifically, this can be achieved in the following ways: One approach is to design a predefined nonlinear function or lookup table that takes the semantically aware gain factor and initial importance weights as input and outputs differentiated modulation intensity. For example, a function f(gain_factor, initial_weight) = gain_factor × g(initial_weight) can be designed, where g is a monotonically increasing function, ensuring that elements with higher initial importance receive greater modulation intensity under the same gain factor. Another approach is to train a lightweight neural network model that takes the semantically aware gain factor and the initial importance weights of each logical element as input features, and outputs differentiated modulation intensity for each logical element by learning the relationship between environmental changes and task performance in historical data. This model can capture complex nonlinear relationships and achieve more precise modulation.

[0091] The purpose of "associating the differentiated modulation intensity of each logical element with its corresponding initial importance weight to form a modulation intensity distribution map" is to establish a structured relationship between differentiated modulation intensity and initial importance weight. This distribution map provides the data foundation for the subsequent construction of the weighted modulation function, ensuring that the modulation function accurately reflects the desired modulation behavior. Specifically, this can be achieved in the following ways: One approach is to use the initial importance weight of each logical element as the x-axis and its corresponding differentiated modulation intensity as the y-axis, forming a series of data points (initial_weight_i, modulation_intensity_i). This set of data points constitutes the modulation intensity distribution map. Another approach is to divide the initial importance weight into several discrete intervals and define one or more representative differentiated modulation intensity values ​​for each interval. For example, for an element with an initial importance weight in [0, 0.2), its modulation intensity may be in the range [0.1, 0.3). For an element in [0.8, 1.0], its modulation intensity may be in the range [0.7, 1.0]. This mapping table can serve as a form of modulation intensity distribution map.

[0092] The key step in constructing a non-uniform weighted modulation field is "building a continuous or piecewise continuous weighted modulation function based on the modulation intensity distribution map. This function takes the initial importance weights as input and outputs modulated weight values, thus forming the non-uniform weighted modulation field." This modulation function, as the core mechanism, can receive any initial importance weights and output adjusted weight values ​​according to a preset modulation rule. Its continuous or piecewise continuous nature ensures the smoothness and predictability of the modulation process. Specifically, this can be achieved in the following ways: One approach is to use spline interpolation (such as cubic spline interpolation) or linear interpolation to construct a continuous weighted modulation function based on the data points in the modulation intensity distribution map. This function can smoothly connect the data points, ensuring that for any given initial importance weight, it can output a reasonable modulated weight value. Another approach is to divide the range of initial importance weights into multiple sub-intervals based on the characteristics of the modulation intensity distribution map, and define a simple function (such as a linear or polynomial function) within each sub-interval. These piecewise functions together constitute a piecewise continuous weighted modulation function. For example, one modulation formula is used when the initial importance weight is below a certain threshold, and another modulation formula is used when it is above the threshold.

[0093] By employing the aforementioned technical solution, this application overcomes the shortcomings of traditional modulation methods, namely, the inability to accurately calculate differentiated modulation intensity based on the initial importance weights and semantic-aware gain factors of logical elements. This results in a less refined and dynamic modulation process, failing to effectively adapt to changes in the importance of different logical elements, thus affecting the accuracy and environmental adaptability of weight modulation. This application calculates differentiated modulation intensity for each logical element and maps it to its initial importance weight to form a modulation intensity distribution map. This allows for the construction of continuous or piecewise continuous weight modulation functions, achieving refined and dynamic adjustment of the initial importance weights of each logical element. This non-uniform weight modulation field more accurately reflects the impact of environmental changes on the importance of each logical element, making the generated environmentally adaptive importance weights more aligned with actual needs. This not only improves the accuracy and environmental adaptability of weight modulation but also avoids abrupt weight changes through smooth function transitions, enhancing the stability and reliability of the AI ​​large-scale model decision-making process. Therefore, it effectively improves the resilience and performance of intelligent autonomous decision-making in resource-constrained and dynamically changing edge scenarios.

[0094] In some of the embodiments described above in this application, a weight modulation field is proposed to modulate the initial importance weights of each logical element in the task logic features to generate environment-adaptive importance weights, so that the importance weights can adapt to changes in the external environment and support subsequent resource scheduling. However, in its implementation, the modulation process may lack fine processing of the individual differences of each logical element, resulting in the weight adjustment being too uniform or static, failing to fully capture the specific impact of environmental semantics on each element, thereby reducing the accuracy and dynamic response capability of the environment-adaptive weights.

[0095] To address this, this application further proposes a method that utilizes a weighted modulation field to perform field-driven modulation on the initial importance weights of each logical element in the task's logical features, generating environment-adaptive importance weights. This method includes: inputting the initial importance weights of each logical element into the weighted modulation field and parsing out the field intensity modulation value acting on each logical element; performing a dynamic recalibration operation on the initial importance weights of each logical element based on this field intensity modulation value to generate the environment-adaptive weights corresponding to each logical element; and sequentially aggregating the environment-adaptive weights of all logical elements to obtain the environment-adaptive importance weights.

[0096] Specifically, when the initial importance weights of each logical element are input into the weight modulation field to parse the field intensity modulation value acting on each logical element, the weight modulation field can be implemented as a non-uniform mapping mechanism that can output the corresponding modulation intensity based on the input weight values ​​and context. For example, the weight modulation field can be a predefined, multi-dimensional lookup table, where the initial importance weight of each logical element serves as an index, and the lookup table returns a corresponding field intensity modulation value. Alternatively, the weight modulation field can be a function implemented by a neural network model, which takes the initial importance weights of each logical element as input, and outputs the field intensity modulation value for that logical element through its internal nonlinear transformation and learned weight distribution characteristics. In this way, the initial importance weight of each logical element can find its specific "position" in the weight modulation field and be assigned a modulation parameter that matches its initial value and field characteristics.

[0097] When performing dynamic recalibration on the initial importance weights of each logical element based on the field intensity modifier value to generate environment-adaptive weights for each logical element, this dynamic recalibration aims to personalize the initial importance weights according to the parsed field intensity modifier value. For example, this dynamic recalibration can use a multiplicative factor adjustment, where the initial importance weight is multiplied by a dynamic gain coefficient derived from the field intensity modifier value to obtain the environment-adaptive weights. Alternatively, this dynamic recalibration can use a nonlinear function mapping, transforming the initial importance weights through a nonlinear function parameterized by the field intensity modifier value to achieve more complex weight adjustment logic, such as amplifying high-importance weights, suppressing low-importance weights, or performing a smooth transition within a specific range.

[0098] When performing ordered aggregation of the environment-adaptive weights of all logical elements to obtain the environment-adaptive importance weights, this ordered aggregation aims to integrate the weights of all logical elements, after individual adjustments, into a unified whole for subsequent unified scheduling and decision-making. For example, this ordered aggregation could organize the environment-adaptive weights of all logical elements into a vector or list according to their original order or dependencies in the task's logical features. Alternatively, if the task's logical features are represented by a graph structure, this ordered aggregation could assign these environment-adaptive weights as attributes of graph nodes or edges, thereby forming a task logic graph with dynamic weight information. This aggregation method ensures the integrity and structure of the weight information, facilitating unified processing by subsequent modules.

[0099] Through the above technical solution, this application can perform refined and individualized dynamic adjustment of the initial importance weight of each logical element through a field-driven modulation mechanism. Specifically, the initial importance weight of each logical element is input into the weight modulation field. By inputting the initial weight, the weight modulation field can parse the field intensity modulation value for each element's specific position and initial value, reflecting individualized modulation rather than global uniform processing, avoiding insufficient accuracy caused by a one-size-fits-all adjustment. The field intensity modulation value acting on each logical element is parsed out. Based on the non-uniform modulation intensity distribution in the weight modulation field, the field intensity modulation value provides customized modulation parameters for each element, solving the problem that environmental semantic influences cannot be differentiated and applied to different elements. Dynamic recalibration is performed based on the field intensity modulation value. Dynamic recalibration adjusts the initial weight in real time according to the modulation value, ensuring that the weight responds to environmental changes and enhances adaptability, overcoming the limitations of static adjustment. Environmental adaptive weights are generated for each logical element. Each element generates weights independently, reflecting its specific environmental semantic influence, improving the accuracy and specificity of the weights. All environmental adaptive weights are aggregated in an ordered manner. This ordered aggregation maintains the overall dependency relationship and structural consistency of the task logic, ensuring the coherence of the generated environmental adaptive importance weights, thereby supporting the stability and effectiveness of subsequent resource scheduling.

[0100] To address this, this application further proposes a method that inputs the initial importance weights of each logical element into a weight modulation field and parses out the field intensity modulation values ​​acting on each logical element. When using the weight modulation field to perform field-driven modulation on the initial importance weights to generate environment-adaptive importance weights, the importance of task logical elements is adjusted according to environmental semantics. However, a key issue in its implementation is how to accurately map the initial importance weights of each logical element to the weight modulation field and parse out the field intensity modulation values ​​to ensure that the modulation intensity is positively correlated with the environmental risk level. If the mapping and parsing mechanism is inaccurate, the modulation values ​​may not accurately reflect environmental changes, thereby reducing the adaptability of weight modulation and the resilience of decision-making.

[0101] To address the aforementioned issues, this application proposes a method for inputting the initial importance weights of each logical element into a weight modulation field and resolving the field emphasis modulation value acting on each logical element. This method includes: The initial importance weight of each logical element is mapped to the corresponding spatial position in the weight modulation field.

[0102] In this weighted modulation field, field strength analysis is performed on the spatial location corresponding to each logic element. The field strength analysis operation analyzes the field strength modulation value based on the mapping region of the spatial location in the modulation intensity distribution map. The magnitude of the field strength modulation value is positively correlated with the modulation intensity represented by the mapping region.

[0103] Specifically, when mapping the initial importance weights of each logical element to their corresponding spatial locations in the weight modulation field, this step aims to transform the abstract, spatially attribute-less initial importance weights into representations with definite coordinates or regions within the weight modulation field. This mapping allows the weight modulation field to exert differentiated influences on different initial importance weights in a spatialized manner. For example, a linear mapping can be used, mapping the initial importance weights to one-dimensional or multi-dimensional spatial locations in the weight modulation field through a simple linear function, where the scaling factor and offset are preset to ensure that the mapped spatial locations are within the effective range of the weight modulation field. Alternatively, a nonlinear mapping can be used. Considering that different weight intervals may require different sensitivities, a nonlinear function (e.g., logarithmic, exponential, or sigmoid function) can be used for mapping, which helps to more finely distinguish logical elements of different importance levels when the weight distribution is uneven. A piecewise mapping can also be used, dividing the initial importance weights into several intervals (e.g., low, medium, and high importance), with each interval corresponding to a specific region or discrete point in the weight modulation field.

[0104] Based on this, field strength analysis is performed on the spatial positions corresponding to each logical element within the weighted modulation field. This step involves extracting the corresponding modulation intensity value, i.e., the field strength modulation value, from a pre-constructed modulation intensity distribution map (as described above) based on the spatial position of the logical element in the weighted modulation field. This distribution map reflects the modulation intensity that different initial importance weights should receive under the current environmental semantics. For example, if the modulation intensity distribution map is a discrete grid or lookup table, the nearest modulation intensity value can be found directly based on the spatial position, or the accurate field strength modulation value can be calculated based on the intensity values ​​of adjacent points using methods such as linear interpolation or bilinear interpolation. Another implementation is that if the modulation intensity distribution map is represented by a continuous mathematical function, the mapped spatial position can be directly substituted into the function for evaluation to obtain the field strength modulation value. This function can be a polynomial, Gaussian mixture model, or other fitting function, and its parameters are jointly determined by the environmental semantic features and the initial importance weights. It should be emphasized that the field strength analysis operation resolves the field strength modulation value based on the mapping region of the spatial location in the modulation intensity distribution map, and the magnitude of the field strength modulation value is positively correlated with the modulation intensity represented by the mapping region.

[0105] Through the above technical solution, this application ensures that the calculation of field intensity modulation values ​​is more accurate and environmentally sensitive, thereby enabling the generated environmentally adaptive importance weights to more accurately reflect the current complex operating environment and resource status. Specifically, mapping the initial importance weights of each logical element to their corresponding spatial locations in the weight modulation field transforms abstract weight values ​​into operable representations with spatial attributes, laying the foundation for subsequent field intensity analysis. This spatial mapping avoids direct fuzzification of weight values, improving processing accuracy. In the weight modulation field, field intensity analysis is performed on the spatial locations corresponding to each logical element. This operation resolves the specific field intensity modulation value based on the mapping area of ​​the spatial location in the modulation intensity distribution map. The modulation intensity distribution map is constructed based on environmental semantic features (such as environmental risk level and task scenario criticality) and initial importance weights; therefore, this analysis ensures that the obtained field intensity modulation values ​​accurately reflect the dynamic changes of the current environment and the actual needs of the task. Crucially, the magnitude of the field intensity modulation value is positively correlated with the modulation intensity represented by the mapped region. This means that when the environmental risk level or the criticality of the task scenario is high, the logical element will receive a larger field intensity modulation value, thus its initial importance weight will be adjusted more significantly in subsequent dynamic recalibration operations (as described above). This enhances the decision-making resilience and adaptability of the AI ​​large model in the face of resource fluctuations, environmental changes, and complex tasks, ensuring the stability and reliability of the system under multiple uncertainties.

[0106] In some of the embodiments described above in this application, resource sensitivity calibration is proposed to calibrate the importance weights of environmental adaptation based on resource constraint characteristics, and generate resilient scheduling weights and resource allocation focus signals. However, in its implementation, the calibration mechanism may lack dynamism and adaptability, and cannot effectively optimize the weight distribution when resource constraints are enhanced, resulting in unintelligent weight adjustment and difficulty in prioritizing key logical elements in resource allocation, thereby affecting the resilience and efficiency of decision-making.

[0107] To address this, this application further proposes a method for resource sensitivity calibration based on the importance weights of environmental adaptation using resource constraint characteristics. This method aims to generate resilient scheduling weights and resource allocation focusing signals. Specifically, see [link to relevant documentation]. Figure 5 The method includes the following steps: 501. Based on the resource constraint characteristics and the distribution characteristics of the importance weights that are adaptive to the environment, dynamic focusing control parameters are generated.

[0108] The focus control parameter is a core dynamic variable used to guide subsequent weight sharpening and resource allocation signal generation. It comprehensively reflects the current resource scarcity level of the system and the distribution characteristics of the importance weights of task logic elements, aiming to make the resource sensitivity calibration process more adaptive. For example, it can be implemented through an adaptive fusion module. This module receives the statistical characteristics of resource constraint features (e.g., CPU utilization, memory availability, battery power) and environmentally adaptive importance weights (e.g., variance, skewness, kurtosis, etc.), and dynamically calculates a scalar or vector-based focus control parameter according to preset fusion rules or machine learning models. When resources are scarce, this parameter value increases, indicating a stronger focus effect is needed. Alternatively, a rule-based expert system can be used. The system presets a series of thresholds and conditional statements regarding resource constraint features (e.g., resource availability prediction values, equipment health assessment values) and weight distribution characteristics (e.g., the number of high-weight elements, the concentration of weight distribution). When a specific combination of conditions is met, the system outputs the corresponding focus control parameter; for example, when resources are extremely limited and the number of high-weight elements is small, a high-intensity focus parameter is generated.

[0109] 502. Based on the focus control parameters, perform nonlinear sharpening processing based on the Pareto principle on the importance weights of the environment to generate the resilience scheduling weights. The nonlinear sharpening processing is configured to make the distribution of the resilience scheduling weights satisfy the Pareto distribution when resource constraints are enhanced.

[0110] The Pareto principle-based nonlinear sharpening process is a weight adjustment mechanism. Its core idea is to amplify the importance of a few critical logical elements while relatively suppressing the importance of the majority of non-critical logical elements, making the weight distribution more consistent with the Pareto principle. The resilient scheduling weights, obtained after this processing, are a more discriminative set of weights that more clearly indicate which logical elements should be prioritized in resource-constrained environments. For example, this can be implemented using a nonlinear function that takes environment-adaptive importance weights as input and outputs sharpened resilient scheduling weights. This nonlinear function can be an exponential function, a sigmoid function, or a tanh function, with its parameters dynamically adjusted by the focus control parameters. When the focus control parameters indicate resource scarcity, the nonlinearity of the function increases, resulting in higher-weighted inputs receiving greater output gains and lower-weighted inputs receiving smaller output gains, thus bringing the weight distribution closer to a Pareto distribution. Alternatively, piecewise linear or piecewise nonlinear functions can be used. Based on the focus control parameters, the environment-adaptive importance weights are divided into several intervals, and different gain factors or nonlinear transformations are applied to the weights of each interval. For example, a larger gain factor is applied to the top 20% of weight values. A smaller gain factor is applied to the remaining 80% of weight values ​​to simulate the characteristics of a Pareto distribution.

[0111] 503. Based on the distribution of the resilience scheduling weight and the focus control parameter, generate the resource allocation focus signal, which is used to prioritize high-weight logical elements during resource allocation.

[0112] The resource allocation focus signal is a specific instruction or marker used in subsequent resource allocation phases to guide the system to prioritize and protect logical elements identified as highly important. Its generation is based on the distribution of resilience scheduling weights and focus control parameters, ensuring targeted and efficient resource allocation. For example, this can be achieved by setting a dynamic threshold. This threshold is dynamically determined based on the distribution of resilience scheduling weights (e.g., calculating their mean, median, or a certain percentile) and focus control parameters (e.g., resource stress). All logical elements with resilience scheduling weights higher than this threshold are marked as "critical," and the identifiers of these critical logical elements and their corresponding priority information are encapsulated into structured data as the resource allocation focus signal. Alternatively, a sorting and proportion-based method can be used. All logical elements are sorted in descending order according to resilience scheduling weights, and then a proportion (e.g., the top N% of elements) is determined based on the focus control parameters. These top N% of logical elements are identified as critical elements, and a list containing their IDs and relative priorities is generated as the resource allocation focus signal.

[0113] Through the above technical solution, this application can solve the problem of the lack of dynamism and adaptability in existing calibration mechanisms when resource constraints are enhanced. Specifically, by generating dynamic focusing control parameters based on the distribution characteristics of importance weights adapted to the environment and resource constraints, the system can perceive subtle changes in resource status and weight distribution in real time, thereby avoiding the limitations of static calibration and enabling subsequent weight adjustments to respond more accurately to the current system state. On this basis, nonlinear sharpening processing based on the Pareto principle is performed on the importance weights adapted to the environment based on the focusing control parameters. This can selectively amplify the importance of a few key logical elements when resource constraints are enhanced, causing their weight distribution to tend towards a Pareto distribution. This ensures that core task logical elements are given priority under limited resources, improving the resilience of decision-making. In addition, a resource allocation focusing signal is generated based on the distribution of resilience scheduling weights and the focusing control parameters, providing a clear priority marker for subsequent resource allocation. This allows resources to be efficiently focused on high-weight logical elements, avoiding resource dispersion and waste, further ensuring the stable execution of key tasks, and improving the efficiency and reliability of autonomous decision-making of the edge AI large model in complex and changing environments.

[0114] In some of the embodiments described above in this application, dynamic focus control parameters are proposed for resource sensitivity calibration. However, in the implementation process, the resource stress level and weight concentration level are not effectively combined, resulting in insufficient calibration accuracy when system resources are under stress, and the inability to adaptively adjust the fusion rules, thereby affecting the generation of resilient scheduling weights.

[0115] To address this, this application further proposes a dynamic focusing control parameter based on the distribution characteristics of importance weights derived from resource constraints and environmental adaptation, specifically including: Based on resource constraint characteristics, a first scaling factor is determined to characterize the degree of resource stress in the system.

[0116] Based on the distribution of importance weights in environmental adaptation, a second scaling factor is determined to characterize the degree of weight concentration.

[0117] Based on the first scaling factor and the second scaling factor, focusing control parameters are generated through an adaptive fusion rule, wherein the adaptive fusion rule is configured to increase the contribution weight of the second scaling factor in the fusion when the first scaling factor increases.

[0118] The first scaling factor, used to characterize the resource stress of the system, quantifies the resource pressure currently faced by terminal devices. This first scaling factor can be based on comparing the predicted resource availability value in the resource constraint characteristics with a preset threshold; for example, when the predicted resource availability value falls below a certain threshold, the first scaling factor increases accordingly. Alternatively, the first scaling factor can be obtained by calculating a combined score of the predicted resource availability value and the device health assessment value, and normalizing it to a specific range to reflect a more comprehensive resource stress situation. Furthermore, by analyzing the historical trends and current instantaneous values ​​of various indicators in the resource constraint characteristics (such as CPU utilization, memory usage, network bandwidth, etc.), and combining this with machine learning models to predict future resource stress, and mapping this prediction to the first scaling factor, a more forward-looking assessment of resource stress can be provided.

[0119] The second scaling factor, used to characterize the degree of weight concentration, quantifies whether the importance weights for environment adaptation are concentrated or dispersed among the task's logical elements. This second scaling factor can measure the concentration of the distribution by calculating the Gini coefficient or entropy value of the importance weights for environment adaptation and mapping it to the second scaling factor. A higher Gini coefficient or lower entropy value indicates a more concentrated weight distribution. Alternatively, the second scaling factor can also characterize the degree of weight concentration by analyzing the kurtosis or skewness of the weight distribution, or by statistically analyzing the proportion of logical elements falling within a specific high-weight interval, and determining the second scaling factor accordingly. For example, the larger the proportion of high-weight elements, the larger the second scaling factor.

[0120] The adaptive fusion rule is configured to increase the contribution weight of the second scaling factor in the fusion process when the first scaling factor increases. Its function is to dynamically adjust the generation strategy of the focus control parameters, enabling it to more intelligently focus on the importance distribution of task logic elements when resources are limited. This adaptive fusion rule can be a weighted average function, where the weight of the second scaling factor is an increasing function of the first scaling factor. For example, the focus control parameters can be expressed as: Focus control parameters = w1 × first scaling factor + w2 (first scaling factor) × second scaling factor, where w2 (first scaling factor) is a function that increases with the first scaling factor. Alternatively, the rule can be a nonlinear mapping function based on a lookup table or neural network. This function takes the first and second scaling factors as input and outputs the focus control parameters. During training or design, it ensures that when the first scaling factor is high, the second scaling factor has a greater impact on the output, thus achieving dynamic weight concentration.

[0121] Through the above technical solution, this application can solve the problems of insufficient accuracy and adaptability in resource sensitivity calibration. Specifically, by determining a first scaling factor based on resource constraint characteristics to characterize the resource stress of the system, it ensures that the generation of focused control parameters closely depends on the actual resource status of the terminal equipment, thus accurately reflecting the constraint strength faced by the system when resources are limited. Simultaneously, a second scaling factor is determined based on the distribution of importance weights adaptive to the environment to characterize the degree of weight concentration. This ensures that the generation of focused control parameters fully considers the distribution characteristics of the importance of each logical element in the decision-making task, avoiding the potential impact of weight distribution on decision resilience during resource allocation. Furthermore, by configuring adaptive fusion rules, when the first scaling factor increases, i.e., when system resource constraints intensify, the contribution weight of the second scaling factor in the fusion can be dynamically increased, thereby optimizing the generation process of focused control parameters. This dynamic adjustment mechanism enables the generated focused control parameters to more accurately reflect which task logical elements are truly critical and require priority protection when resources are scarce, thus supporting the effective generation of subsequent resilience scheduling weights. This enables large AI models to more intelligently focus limited resources on a few high-importance task elements in complex scenarios with limited resources and dynamically changing environments, thereby improving the resilience, reliability, and efficiency of decision-making.

[0122] In some of the above-mentioned schemes in this application, nonlinear sharpening processing is performed on the importance weights of environmental adaptation based on the focus control parameters to generate resilient scheduling weights. However, in this process, there is a lack of specific mechanisms to accurately determine the sharpening intensity parameters and configure the transformation curve to ensure that the high importance weights obtain higher output gain, thereby effectively optimizing resource allocation when resource constraints are enhanced.

[0123] To address this, this application further proposes a method based on focused control parameters to perform Pareto-based nonlinear sharpening on the environmentally adaptive importance weights, generating resilient scheduling weights. Specifically, this involves: determining a sharpening intensity parameter for modulating the nonlinear transform function based on the focused control parameters; configuring the transform curve of the nonlinear transform function based on the sharpening intensity parameter, wherein the transform curve is configured such that the output gain for high-importance weights is higher than the output gain for low-importance weights; and using the configured nonlinear transform function, performing a transform calculation on the environmentally adaptive importance weights to obtain the resilient scheduling weights.

[0124] Among them, the focus control parameter, as a dynamic indicator, comprehensively reflects the current resource stress level of the system and the distribution characteristics of the importance weights in environmental adaptation. Its role is to guide subsequent nonlinear sharpening processing, ensuring that the sharpening intensity can be appropriately adjusted according to the real-time operating context. One implementation is that this parameter can be a scalar value, calculated by weighting multiple resource availability indicators (e.g., CPU utilization, memory usage, network bandwidth) and combining them with statistics on the distribution of importance weights in environmental adaptation (e.g., entropy or Gini coefficient). Another implementation is that this parameter can be a vector containing multiple independent components, each representing a specific aspect of resource constraints (e.g., resource scarcity index) and a specific characteristic of the weight distribution (e.g., concentration index), which are used to derive the sharpening intensity.

[0125] The sharpening intensity parameter quantifies the degree to which the nonlinear transformation function amplifies the difference between high and low importance weights. Higher sharpening intensity means greater differentiation of key elements; high weights will be further boosted, while low weights may be suppressed, thus focusing resources more intently on key elements. One implementation is that the sharpening intensity parameter can be directly mapped from the focus control parameters using a predefined monotonic function (e.g., the sigmoid function or an exponential function) to ensure a nonlinear response to increasing resource stress. Another implementation utilizes a lookup table or a small neural network, taking the focus control parameters as input and outputting an optimized sharpening intensity parameter, which can be fine-tuned using historical performance data.

[0126] Nonlinear transformation functions aim to remap environment-adaptive importance weights in a nonlinear manner, with the core objective of enhancing the relative importance of higher-weighted elements. The nonlinear nature ensures that the transformation does not treat all weights equally, but rather applies differentiated gains—that is, higher input weights receive higher gains. One implementation involves using a power function (e.g., y = x^p, where p > 1 is used for sharpening) or an exponential function (e.g., y = exp(kx)) as the nonlinear transformation function, where the exponent p or coefficient k is modulated by the sharpening intensity parameter. Another implementation can employ a piecewise linear function or spline interpolation function, with the piecewise points and slope dynamically adjusted according to the sharpening intensity parameter to achieve the desired differentiated gain effect.

[0127] A transform curve is a graphical representation of a nonlinear transform function, visually demonstrating how input weights map to output weights. A key characteristic of this curve, configured by the sharpening intensity parameter, is its steeper slope (or derivative) with respect to higher input weight values, ensuring that the output gain for high-importance weights is significantly higher than that for low-importance weights. One implementation involves configuring the transform curve by adjusting the parameters of a mathematical function (e.g., the exponent in a power-law function, the base in an exponential function, or the polynomial coefficients), directly based on the sharpening intensity parameter. Another approach is to define a set of control points for a Bézier curve or B-spline curve, the positions of which are dynamically calculated based on the sharpening intensity parameter to form the desired curve shape that prioritizes high-importance weights.

[0128] Resilient scheduling weights are weights assigned to each logical element after nonlinear sharpening. They represent a refined measure of importance that not only adapts to dynamic environmental changes but is also calibrated for resource scarcity, ensuring that critical elements receive priority in resource allocation and execution opportunities, thereby enhancing the overall system's resilience. Resilient scheduling weights can be obtained directly by applying a configured nonlinear transformation function to environment-adaptive importance weights. After the transformation, an optional normalization step (e.g., scaling to a sum of 1 or a specific range) can be performed to ensure consistency in subsequent scheduling or resource allocation algorithms.

[0129] Through the above technical solution, this application can dynamically adjust the degree of weight sharpening. The focused control parameters comprehensively reflect the current resource scarcity of the system and the distribution characteristics of the importance weights in environmental adaptation. Therefore, the determined sharpening intensity parameter can accurately adapt to the real-time system state, avoiding resource allocation imbalance caused by improper sharpening intensity settings when resources are limited, and ensuring the effectiveness and targeting of the sharpening process. Based on this, a transformation curve of a nonlinear transformation function is configured based on the sharpening intensity parameter, enabling the curve to apply higher output gain to high-importance weights and relatively lower gain to low-importance weights. This nonlinear characteristic effectively amplifies the importance of key logical elements, giving them significant priority in resource allocation, thus allowing limited resources to be more concentratedly allocated to parts crucial to task completion and system resilience when resources are scarce. The configured nonlinear transformation function is used to transform and calculate the environmentally adaptive importance weights, generating resilient scheduling weights. These resilient scheduling weights not only consider the dynamic changes in the environment but are also optimized under resource constraints, enabling each logical element of the decision-making task to obtain differentiated and more resilient scheduling priorities. This improves the decision-making reliability and system stability of edge AI large models in the face of resource fluctuations and environmental changes, ensuring the continuous operation of key functions and maintaining the execution of core tasks even under extreme conditions, thereby achieving a more accurate, efficient and resilient resource allocation strategy.

[0130] In some of the above-mentioned schemes of this application, a sharpening intensity parameter is proposed to generate resilient scheduling weights by modulating the nonlinear transformation function. However, in this process, the sharpening intensity parameter fails to accurately reflect the nonlinear growth characteristics of resource stress and ignores the influence of cognitive distribution characteristics, resulting in inaccurate sharpening processing and affecting the optimization effect of resource allocation.

[0131] To address this, this application further proposes a method for determining a sharpening intensity parameter for modulating a nonlinear transform function based on the focusing control parameter. The method includes: determining a quantified value of resource stress in the current system based on the focusing control parameter; inputting this quantified value of resource stress into a predefined intensity mapping function configured to map the input resource stress to a nonlinearly increasing sharpening intensity reference value; and applying a fine-tuning compensation based on the cognitive distribution characteristics in the focusing control parameter to the sharpening intensity reference value to obtain the sharpening intensity parameter.

[0132] Specifically, in determining the quantified value of resource stress in the current system, this step aims to extract or calculate a quantitative representation of the current system resource stress level from existing focus control parameters. The quantified value of resource stress reflects the pressure level faced by critical resources such as computing, storage, and network of terminal devices under their current operating state. Its role is to provide a basic input reflecting the system resource status for subsequent determination of sharpening intensity parameters. For example, the preset resource stress component can be directly extracted as a quantified value by analyzing the internal structure of the focus control parameter. If the focus control parameter is a vector, where a certain dimension specifically encodes the resource stress level, then the value of that dimension can be directly read. Alternatively, it can be calculated by performing a weighted summation or nonlinear combination operation on multiple sub-parameters in the focus control parameter. For example, by combining the first scaling factor representing resource availability in the focus control parameter with the device health assessment value, a comprehensive quantified value of resource stress can be calculated using a preset function model.

[0133] The resource stress quantification is input into a predefined intensity mapping function, which is configured to map the input resource stress to a non-linearly growing sharpening intensity benchmark value. This step aims to convert the aforementioned resource stress quantification into an initial sharpening intensity benchmark value through a predefined function. The core characteristic of this intensity mapping function is its non-linear growth configuration, meaning that as resource stress increases, the sharpening intensity benchmark value grows at a faster rate, thus responding more sensitively to changes in resource pressure. Its function is to establish a non-linear relationship between resource stress and sharpening intensity, ensuring a stronger sharpening effect can be applied when resources are under high stress. For example, the intensity mapping function can be a polynomial function, such as f(x) = ax^2 + bx + c (where a > 0), or an exponential function, such as f(x) = A × e^(Bx) (where B > 0), with the non-linear growth characteristic achieved by adjusting the parameters a, b, c or A, B. Furthermore, the intensity mapping function can also be a piecewise function, using different nonlinear curves for mapping in different resource stress intervals. For example, a gentle growth curve can be used in the low stress interval, while a steep growth curve can be used in the high stress interval.

[0134] Based on this, a fine-tuning compensation based on the cognitive distribution characteristics in the focus control parameters is applied to the sharpening intensity benchmark value to obtain the sharpening intensity parameter. This step aims to refine the initially obtained sharpening intensity benchmark value to better adapt to the importance weight distribution characteristics of the current task logic elements. The cognitive distribution characteristics refer to the information contained in the focus control parameters regarding the environmentally adaptive importance weight distribution, such as the concentration and dispersion of the weights. Through fine-tuning compensation, the sharpening intensity parameter can not only reflect resource scarcity but also take into account the distribution characteristics of the weights themselves, thereby achieving a more accurate sharpening effect. For example, fine-tuning compensation can be implemented through an additional compensation function that takes the second scaling factor representing the concentration of weights in the focus control parameters as input, outputs a compensation amount, and then adds or multiplies this compensation amount with the sharpening intensity benchmark value. When the weight distribution is highly concentrated, the compensation amount may be positive to further enhance the sharpening effect. Alternatively, it can be implemented through a small neural network model that receives the sharpening intensity benchmark value and the cognitive distribution characteristics encoded in the focus control parameters as input, and learns to obtain a better sharpening intensity parameter. The model can be trained offline by simulating different resource and weight distribution scenarios.

[0135] Through the above technical solution, this application can more accurately determine the sharpening intensity parameter used to modulate the nonlinear transformation function. Directly determining the quantified value of resource stress based on the focus control parameter avoids redundant calculations, improving efficiency and data consistency. Configuring this quantified value as a nonlinearly growing intensity mapping function allows the sharpening intensity benchmark value to respond more sensitively and reasonably to the escalation of resource stress, especially in highly resource-constrained scenarios, providing a stronger sharpening effect and ensuring that the importance of key task logic elements is fully highlighted. By applying fine-tuning compensation based on the cognitive distribution characteristics in the focus control parameter to the sharpening intensity benchmark value, the sharpening intensity parameter not only considers resource stress but also takes into account the specific distribution of adaptive importance weights in the current environment, avoiding rigid parameter settings. This refined parameter determination method ensures that the subsequently generated resilient scheduling weights can more accurately reflect the system resource status and the importance distribution of task logic elements, thus providing a more reliable basis for matching differentiated elastic execution strategies to each logic element of the decision-making task, improving the decision resilience and reliability of the edge AI large model under resource constraints and dynamic environmental changes.

[0136] In some of the solutions mentioned above in this application, a resource allocation focus signal is proposed to prioritize high-weight logical elements during resource allocation. However, in its implementation, how to dynamically determine the threshold to filter key logical elements to adapt to changes in resource constraints and weight distribution, ensure that key functions are prioritized when resources are scarce, and avoid inaccurate resource allocation or response delays due to fixed thresholds is a technical problem.

[0137] To address this, this application further proposes a method for generating a resource allocation focus signal based on the distribution of resilience scheduling weights and focus control parameters. The method includes: determining a dynamic focus threshold based on the focus control parameters, which is used to filter key logical elements from the resilience scheduling weights; identifying the logical elements corresponding to weights greater than or equal to the dynamic focus threshold as a set of key logical elements; and encapsulating the set of key logical elements and their corresponding priority information into a structured instruction to obtain the resource allocation focus signal.

[0138] The focus control parameter is a comprehensive quantitative indicator. Its generation is based on the distribution characteristics of importance weights derived from resource constraints and environmental adaptation. It dynamically reflects the current resource stress of the system and the concentration of importance weights among task logic elements. This parameter can be a single numerical value, such as a normalized value between 0 and 1, representing the overall resource pressure and task criticality focus requirements of the system. Alternatively, it can be a vector containing multiple components, such as one component representing resource availability and another representing the entropy of the weight distribution.

[0139] This dynamic focus threshold is a value calculated in real time based on the focus control parameter. Its value is not fixed but adaptively adjusts according to the system's current resource situation and task importance distribution. For example, when the focus control parameter indicates high resource scarcity, the dynamic focus threshold will increase accordingly, ensuring that only a very small number of the most critical logical elements pass the screening. Conversely, when resources are relatively abundant, the dynamic focus threshold may decrease, allowing more logical elements to be identified as critical. This threshold can be generated through a preset lookup table, a non-linear function based on the focus control parameter, or a machine learning model that predicts based on historical data and the current state.

[0140] The resilience scheduling weight is a resource-sensitivity-calibrated importance weight assigned to each logical element. It reflects the contribution of that logical element to the overall resilience of the decision-making task under the current resource constraints. This weight is usually a numerical value; the larger the value, the more critical the logical element.

[0141] The process of selecting key logical elements specifically involves comparing the resilience scheduling weight of each logical element with the dynamic focusing threshold. If the resilience scheduling weight of a logical element is greater than or equal to the dynamic focusing threshold, then that logical element is determined to be a key logical element. For example, all logical elements can be traversed and compared one by one. Alternatively, parallel computing can be used to select multiple logical elements simultaneously.

[0142] The set of key logical elements is a collection of all logical elements that have been selected as key after filtering. This set can be a list, an array, or a hash table, and contains unique identifiers for these key logical elements.

[0143] This priority information is associated with the set of critical logical elements and indicates the relative importance or execution order of these critical logical elements. This information can originate from the resilience scheduling weight value of the logical element itself; for example, a logical element with a higher weight value has a higher priority. It can also be a preset priority level, such as uniformly marking all critical logical elements as "high priority," or dividing them into different priority levels based on their weight range.

[0144] This encapsulation as a structured instruction refers to organizing and packaging the set of key logical elements and their corresponding priority information according to a predetermined data format. For example, it can be encapsulated as a JSON object, an XML document, a ProtocolBuffer message, or a specific binary data structure. This structured instruction ensures the clarity, integrity, and parsability of information when it is transmitted between different modules within the system (such as the resource scheduler and the execution engine).

[0145] The resource allocation focus signal is a generated instruction or data stream used to guide the subsequent resource allocation process. This signal clearly indicates which logical elements are currently the most critical and what priority they should receive, thus guiding the resource allocator to prioritize limited resources for these critical logical elements.

[0146] Through the above technical solution, this application can dynamically determine thresholds to filter key logical elements, thereby solving the problem of dynamically adapting to changes in resource allocation. Specifically, the focus threshold is dynamically determined based on focus control parameters, enabling the threshold to reflect the system's resource scarcity and weight concentration characteristics in real time, avoiding the rigidity problems that may occur with fixed thresholds in dynamic environments. This dynamic adjustment mechanism ensures that the system can accurately identify the truly high-weight logical elements under current conditions when resource constraints and weight distribution fluctuate, thereby optimizing resource allocation priorities. Identifying the logical elements corresponding to weights greater than or equal to the dynamic focus threshold in resilient scheduling weights as the set of key logical elements ensures that resources are focused only on the most critical logical elements, reducing resource consumption by unnecessary elements and improving the efficiency and accuracy of resource allocation. Furthermore, encapsulating the set of key logical elements and their corresponding priority information into structured instructions allows the identification results to be converted into directly executable instructions, facilitating rapid system response and ensuring that priority information is efficiently transmitted and applied in resource scheduling. Overall, this method enables edge AI large models to focus resources and schedule tasks more intelligently and flexibly when facing challenges such as limited resources and dynamic environmental changes, thereby ensuring the stability and reliability of core decision-making functions and improving the overall resilience of the system.

[0147] In some of the embodiments described above in this application, a dynamic focusing threshold is proposed to filter key logical elements. However, in this process, the threshold setting may not be dynamic and adaptive enough, and may not be able to effectively combine the changes in resource constraint strength and weight distribution characteristics. This results in insufficient accuracy in filtering key logical elements when resources are scarce or weights are highly concentrated, affecting the resilience and efficiency of resource allocation.

[0148] To address this, this application further proposes a method for determining a dynamic focusing threshold based on focusing control parameters. This method includes: parsing a first quantization component characterizing the strength of resource constraints and a second quantization component characterizing the weight distribution characteristics from the focusing control parameters; determining a basic focusing threshold based on the first quantization component; and adaptively correcting the basic focusing threshold according to the second quantization component to generate a dynamic focusing threshold.

[0149] Specifically, the first quantized component characterizing the strength of resource constraints and the second quantized component characterizing the weight distribution characteristics are extracted from the focused control parameters. This aims to extract two independent, quantified indicators from the comprehensive focused control parameters, reflecting the current resource stress level of the system and the distribution characteristics of the importance weights of each logical element in the decision-making task, respectively. This allows subsequent threshold determination and correction processes to be based on more refined and targeted information. For example, the focused control parameters can be a multi-dimensional vector, where specific dimensions or fields are predefined to correspond to resource stress (such as CPU utilization, memory usage, etc.) and the concentration of weight distribution (such as entropy, Gini coefficient, etc.). The parsing process directly reads the values ​​of these predefined fields. Alternatively, the focused control parameters may be a single numerical value or a low-dimensional vector that has undergone fusion processing. In this case, a pre-trained parsing module (such as a small neural network or a rule-based parser) can perform feature decomposition to separate the first and second quantized components.

[0150] Determining the basic focus threshold based on the first quantization component refers to establishing a preliminary, benchmark focus threshold using the parsed resource constraint intensity index. This basic focus threshold directly reflects the current system resource availability, ensuring that the threshold increases accordingly when resources are limited, thereby more rigorously filtering out the most critical logical elements. For example, a piecewise function or lookup table can be preset to map different value ranges of the first quantization component to different basic focus thresholds. When the resource constraint intensity is low, the basic focus threshold can be set lower, allowing more logical elements to be considered critical. When the resource constraint intensity is high, the basic focus threshold is increased accordingly to focus on fewer, more critical elements. Another approach is to design a continuous mathematical function (such as an exponential function or a sigmoid function) that directly calculates the basic focus threshold using the first quantization component as input. This function should ensure that the basic focus threshold increases accordingly with the increase of the first quantization component, and the rate of increase can be non-linear to better adapt to different levels of resource stress.

[0151] The adaptive adjustment of the basic focusing threshold based on the second quantization component to generate a dynamic focusing threshold refers to further refining the adjustment based on the characteristics of the importance weight distribution, building upon the basic focusing threshold. This adjustment mechanism avoids overly rigid threshold settings when the weight distribution is extremely concentrated or dispersed, thereby improving the accuracy of key logical element screening. For example, the second quantization component can represent the degree of concentration of the weight distribution (e.g., the reciprocal of entropy or the Gini coefficient). When the weight distribution is more concentrated (i.e., a few elements have high weights, and most elements have low weights), the second quantization component is larger. In this case, a positive adjustment can be made to the basic focusing threshold (e.g., multiplied by a correction factor greater than 1), slightly increasing the threshold to more rigorously screen out truly "peak" elements. Conversely, when the weight distribution is more dispersed, the second quantization component is smaller, allowing for a negative adjustment (e.g., multiplied by a correction factor less than 1) to avoid missing some relatively important elements that have low weights. Furthermore, the second quantization component can also be used to indicate the statistical characteristics of the weight distribution, such as variance, skewness, or kurtosis. When the variance of the weight distribution is small (weight values ​​are generally similar), the basic focusing threshold can be appropriately lowered to capture more relatively important elements. When the variance is large (weight values ​​differ significantly), the basic focusing threshold can be appropriately increased to focus more on a few high-weight elements. The correction function can be a non-linear function based on these statistical properties to ensure the flexibility and adaptability of the correction.

[0152] Through the aforementioned technical solution, this application can accurately separate two key pieces of information from the focusing control parameters: resource constraint strength and weight distribution characteristics. Based on this, a dynamic focusing threshold is determined step-by-step and adaptively. This mechanism ensures that threshold setting is no longer a static or simple linear adjustment, but rather deeply perceives the current system's resource scarcity and the inherent distribution patterns of the importance weights of task logic elements. When resources are scarce, the threshold can automatically increase, prompting the system to focus more on a few core logic elements. When the weight distribution is highly concentrated, the threshold can also be adjusted accordingly, avoiding misjudgments of truly critical elements due to a "one-size-fits-all" threshold. Overall, this dynamic and adaptive threshold determination method improves the accuracy and efficiency of selecting key logic elements under resource fluctuations and weight changes, thereby enhancing the resilience and reliability of the edge AI large-scale model's autonomous decision-making. This ensures that in complex and ever-changing environments, the system can always invest limited resources in the most critical decision-making stages, effectively coping with performance degradation and external emergencies.

[0153] In some of the embodiments described above in this application, a differentiated flexible execution strategy is proposed to match each logical element of the decision-making task in order to achieve adaptive decision-making. However, in the process of its implementation, due to the lack of dynamic division of logical element priorities, deep integration of multi-dimensional features and explicit binding of resource constraints, the strategy matching may be inaccurate, the resource utilization efficiency may be low, and the priority execution of key logical elements may not be guaranteed in complex environments.

[0154] To address this, this application proposes a method that matches differentiated elastic execution strategies to each logical element of a decision-making task based on resilient scheduling weights, resource allocation focus signals, task logical features, and environmental semantic features. (See [link to relevant documentation]). Figure 6 The method includes: 601. Based on the resource allocation focus signal, determine the priority partitioning of the logical elements to be executed in the decision task.

[0155] 602. For any logical element among the logical elements to be executed, the resilience scheduling weight corresponding to the logical element, the environmental semantic features related to the logical element, and the dependency relationship of the logical element in the task logical features are fused to generate a multi-dimensional policy feature vector of the logical element.

[0156] 603. Input the multi-dimensional strategy feature vector of each logical element into the strategy synthesis model, synthesize and output a flexible execution strategy customized for each logical element, wherein the strategy synthesis model constrains the upper limit of resource consumption of the synthesized strategy based on priority partitioning.

[0157] Specifically, when determining the priority partitioning of logical elements to be executed in a decision-making task, this step aims to classify each logical element in the task into different priority levels based on resource allocation guidance signals. The purpose of this is to ensure that critical or high-priority logical elements receive priority resource guarantees and execution opportunities in resource-constrained or complex environments, thereby addressing the problems of inefficient resource utilization and the inability of critical logical elements to be executed first. For example, this can be achieved through a pre-defined rule engine that assigns logical elements to predefined priority levels (such as "core," "important," "minor," etc.) based on explicit priority labels or thresholds contained in the resource allocation focus signal. Another approach is to use a machine learning classifier, such as a support vector machine or decision tree, which is trained to map logical elements to discrete priority levels based on features indicated by the resource allocation focus signal (such as resource scarcity, element criticality score, etc.).

[0158] When fusing the resilience scheduling weights corresponding to logical elements, the environmental semantic features related to the logical elements, and the dependencies of the logical elements in the task's logical features to generate a multi-dimensional policy feature vector for the logical elements, this step aims to integrate various attributes such as the inherent importance of the logical elements, their external environmental context, and their structural role in the entire task process into a unified and information-rich vector representation. Its purpose is to provide a comprehensive and refined input for subsequent policy synthesis, ensuring that the generated resilient execution policy fully considers all relevant factors of the logical elements, thereby achieving accurate policy matching. For example, this can be achieved through simple vector concatenation, directly connecting the numerical representations of resilience scheduling weights, environmental semantic features, and dependencies to form a longer comprehensive vector. Alternatively, weighted summation or averaging methods can be used, assigning different weights based on the importance or relevance of each feature type, and then combining them. Furthermore, a small neural network (such as a multilayer perceptron) can be used to non-linearly process and fuse these different feature sets to learn and generate an optimal fused representation.

[0159] In a policy synthesis model, the multidimensional policy feature vector of each logical element is input into the model. This model synthesizes and outputs a customized elastic execution policy for each logical element. The core step in generating customized execution policies is when the policy synthesis model constrains the resource consumption limit of the synthesized policy based on priority partitioning. This step explicitly combines resource constraints with priority information. The goal is to generate an execution policy for each logical element that is both adaptable to current conditions and resource-feasible, thereby ensuring the effectiveness of the policy and the rationality of resource use, avoiding resource overruns. For example, a rule-based policy generator can be used, which contains a series of predefined rules that map a specific range or combination of multidimensional policy feature vectors to a set of pre-configured elastic execution policies. In this process, priority partitioning acts as a filter or modifier, adjusting the resource parameters of the selected policy (e.g., reducing the computational precision of low-priority elements). Another implementation method is to utilize a reinforcement learning agent, which is trained to select or generate the optimal elastic execution policy. Its reward function considers both task performance metrics and penalties for exceeding resource limits, while priority partitioning influences the resource constraint in the reward function. In addition, generative models, such as variational autoencoders or Transformer-based models, can be used. These models take multidimensional policy feature vectors as input to generate parameters for resilient execution policies and use priority partitioning as additional conditional input to guide the generation process to produce policies that meet resource constraints.

[0160] Through the above technical solutions, this application can solve the problems of inaccurate strategy matching and low resource utilization efficiency in related technologies. By dynamically determining the priority partitioning of the logical elements to be executed based on resource allocation focusing signals, this application can ensure that key logical elements are given priority processing under conditions of limited resources or complex environments, avoiding efficiency losses caused by resource dispersion, thereby improving the resilience of the system in uncertain environments. By integrating the dependencies in resilience scheduling weights, environmental semantic features, and task logical features, this application generates multi-dimensional strategy feature vectors for logical elements. This provides comprehensive and refined input for strategy synthesis, enabling the generated elastic execution strategies to more accurately adapt to the importance of logical elements, environmental context, and their role in the task, thus improving the accuracy of strategy matching. When generating customized elastic execution strategies, the strategy synthesis model explicitly constrains the resource consumption limit of the synthesized strategies based on priority partitioning. This ensures that all generated strategies are resource-feasible, effectively preventing resource overruns, thereby optimizing resource utilization efficiency and ensuring the stable and reliable operation of the edge AI large model in complex and changing environments. Overall, this application improves the adaptability, reliability, and resource efficiency of autonomous decision-making in edge AI large models through a collaborative mechanism of dynamic priority allocation, deep feature fusion, and resource constraint binding.

[0161] In some of the embodiments described above in this application, priority partitioning of the logical elements to be executed in the decision task is proposed based on the resource allocation focus signal to assist in matching the flexible execution strategy. However, in its implementation, the existing priority partitioning method may lack sufficient dynamism and adaptability, and cannot efficiently allocate resources when resource constraints change, affecting the stability and reliability of the decision.

[0162] To address this, this application further proposes a method for determining the priority partitioning of logical elements to be executed in the decision-making task. This method includes: parsing the resource allocation focus signal and extracting a dynamic priority spectrum, which defines multiple priority levels with different guarantee levels; determining an adaptive classification boundary that maps logical elements to different priority levels based on the distribution of the dynamic priority spectrum and the resilience scheduling weights; and assigning each logical element to be executed to its corresponding priority level based on the adaptive classification boundary, thus forming the priority partitioning.

[0163] Specifically, in parsing the resource allocation focus signal and extracting the dynamic priority spectrum, which defines multiple priority levels with different guarantee levels, the resource allocation focus signal is generated by the system after comprehensively considering the importance weights of resource constraint characteristics and environmental adaptation. It aims to guide the prioritization of high-weight logical elements during resource allocation. This dynamic priority spectrum is a non-fixed set of priority levels that can be adjusted in real time according to the current system state and task requirements. It defines the resource guarantee level, execution latency requirements, or fault tolerance capabilities corresponding to different priority levels. One implementation is to directly construct the dynamic priority spectrum by parsing the priority instructions or metadata contained in the resource allocation focus signal. For example, the signal may explicitly indicate that three levels should be enabled: "real-time high priority," "normal priority," and "low guarantee priority," with their respective resource quotas and scheduling strategies. Another implementation is to input the resource allocation focus signal into a pre-trained priority spectrum generation model, such as a reinforcement learning or deep learning-based model. This model dynamically generates or selects the most suitable priority spectrum for the current scenario based on the signal's characteristics, including the number of priority levels and their specific guarantee levels.

[0164] Based on this, in the process of determining the adaptive classification boundary for mapping logical elements to different priority levels based on the distribution of the dynamic priority spectrum and the resilience scheduling weights, the adaptive classification boundary refers to the dynamic threshold or rule set used to classify logical elements into different priority levels according to their resilience scheduling weights. These boundaries are not fixed but can be adjusted in real time according to the requirements of the dynamic priority spectrum and the current distribution of resilience scheduling weights of all logical elements. One implementation is to use statistical methods to determine the adaptive classification boundary. For example, based on the number of priority levels defined in the dynamic priority spectrum, combined with the cumulative distribution function (CDF) of the resilience scheduling weights, the weight distribution is divided into several intervals, each interval corresponding to a priority level. The boundary can be set as a specific percentile of the CDF (such as 20%, 50%, 80%), and these percentiles will be dynamically adjusted according to the actual distribution of the resilience scheduling weights. Another implementation is to use machine learning models, such as clustering algorithms or classifiers, to determine the adaptive classification boundary. This model takes the resilience scheduling weights as input and combines the dynamic priority spectrum as supervision information or constraints to learn how to automatically classify logical elements into different priority levels. The model can be trained based on historical data and fine-tuned at runtime according to the new weight distribution.

[0165] Furthermore, in the process of dividing each logical element to be executed into a corresponding priority level based on the adaptive classification boundary, forming this priority partition, the priority partition refers to assigning all logical elements to be executed in the decision task to various predefined priority levels in the dynamic priority spectrum according to their resilience scheduling weights and the determined adaptive classification boundary, thereby forming a structured priority grouping. One implementation is to obtain the corresponding resilience scheduling weight for each logical element to be executed and compare it with the determined adaptive classification boundary. For example, if the resilience scheduling weight of a logical element is higher than the first boundary but lower than the second boundary, it is assigned to the intermediate priority level. Another implementation is to encode the adaptive classification boundary into a series of judgment rules by constructing a decision tree or rule engine. Each logical element is assigned to the corresponding priority level according to these rules based on its resilience scheduling weight, thereby completing the priority partition.

[0166] Through the above technical solution, this application can overcome the shortcomings of existing priority partitioning methods in terms of dynamism and adaptability. Specifically, by analyzing resource allocation focusing signals to extract a dynamic priority spectrum, the system can flexibly adjust the priority structure according to the current resource status and environmental changes, avoiding the rigidity of static priority settings and ensuring that the priority hierarchy can respond to changes in resource allocation signals in real time. On this basis, an adaptive classification boundary is determined based on the distribution of the dynamic priority spectrum and resilient scheduling weights. This ensures that priority partitioning not only considers the preset priority levels but also incorporates the actual importance of logical elements (resilient scheduling weights), thereby enabling more precise focus on key elements and preventing resource waste or failure to guarantee critical tasks due to "one-size-fits-all" boundary partitioning during resource fluctuations. By partitioning each logical element into its corresponding priority level based on the adaptive classification boundary, forming priority partitions, the resource allocation strategy can be executed efficiently according to dynamically adjusted boundaries, optimizing resource utilization efficiency and enhancing the decision-making stability and reliability of the edge AI large model in complex, dynamic, and resource-constrained environments, enabling it to better achieve system-level resilience.

[0167] In some of the embodiments described above in this application, a method is proposed to fuse resilient scheduling weights, environmental semantic features, and the dependencies of logical elements in task logical features to generate a multi-dimensional policy feature vector, which supports the matching of resilient execution policies. However, in its implementation, feature fusion may lack context-aware calibration, and the dependency context may not be effectively focused and encoded, resulting in insufficient feature representation, which in turn affects the accuracy and efficiency of policy synthesis.

[0168] To address this, this application further proposes a method for generating a multi-dimensional policy feature vector for the logical element. This method aims to optimize the feature fusion process, thereby improving the matching accuracy and efficiency of resilient execution policies. Specifically, the method includes: calibrating the resilient scheduling weights for context-aware weight execution based on the environmental semantic features to obtain a context-aware decision weight vector; using this context-aware decision weight vector, performing attention-focused encoding on the dependency context of the logical element extracted from the task's logical features to generate a dependency-aware feature vector; and inputting the context-aware decision weight vector, the dependency-aware feature vector, and the environmental semantic features into a feature generation network for cross-domain feature synthesis to obtain the multi-dimensional policy feature vector of the logical element.

[0169] Based on the environmental semantic features, context-aware weight calibration is performed on the resilience scheduling weights to obtain a contextualized decision weight vector. This aims to dynamically adjust the resilience scheduling weights of logical elements according to the current external environment and user intent (represented by environmental semantic features), making them more accurately reflect their actual importance in a specific context. For example, when the environmental risk level is high or the user intent is urgent, the weights of logical elements related to safety or critical tasks should be increased. In practice, an adaptive weight adjustment module can be constructed. This module receives environmental semantic features as input and performs multiplicative or additive corrections to the resilience scheduling weights according to a preset rule set or learned mapping relationships. For example, a function can be defined to generate a calibration factor, and then the resilience scheduling weights can be multiplied by this calibration factor. Another approach is to utilize a lightweight neural network model, such as a multilayer perceptron, which takes environmental semantic features and resilience scheduling weights as input, learns through nonlinear transformations of the network, and outputs a contextualized decision weight vector. This network can be trained offline, enabling it to finely adjust the resilience scheduling weights according to different environmental contexts, thus obtaining more context-adaptive decision weights.

[0170] Furthermore, using this contextualized decision weight vector, attention-focused encoding is applied to the dependency context of the logical element extracted from the task's logical features, generating a dependency-aware feature vector. The aim is to selectively focus on and encode the dependency context of the logical element within the task's logical features using the context-calibrated contextualized decision weight vector. The goal is to highlight dependencies that are crucial to the execution of the decision task in the current context, thereby generating a more information-dense and targeted dependency-aware feature vector. In practice, an attention-based approach can be employed. For example, the dependency context of the logical element can be represented as a graph structure or sequence, and the contextualized decision weight vector can be used as a query to calculate attention to each node or edge in the dependency context, generating attention scores. These scores are used to weight and aggregate dependency features, giving higher attention to dependencies related to high-weight logical elements, thus resulting in a stronger representation in the generated dependency-aware feature vector. Alternatively, this can be achieved by constructing a simplified graph neural network module. In this module, the contextualized decision weight vector can be used as a feature of nodes or edges to participate in the process of graph convolution or message passing. This guides the network to exert a greater influence on the key dependency paths indicated by the contextualized decision weight vector when aggregating dependency information, and outputs a dependency-aware feature vector that can effectively capture contextualized key dependency information.

[0171] The contextualized decision weight vector, the dependency-aware feature vector, and the environmental semantic features are input into a feature generation network for cross-domain feature synthesis, resulting in a multi-dimensional policy feature vector for this logical element. The aim is to effectively fuse features from different domains and modalities (contextualized decision weight vector, dependency-aware feature vector, and environmental semantic features) to generate a unified and comprehensive multi-dimensional policy feature vector. This vector serves as the input to the subsequent policy synthesis model, providing rich and accurate data for matching differentiated elastic execution policies. Specifically, a feature generation network can be used. A common approach is to use a deep neural network, such as a multilayer perceptron or a convolutional neural network. This network concatenates or inputs the contextualized decision weight vector, dependency-aware feature vector, and environmental semantic features in parallel. Through multi-layer nonlinear transformations and feature learning, it automatically extracts and fuses the latent correlations between different features, generating a high-dimensional, semantically rich multi-dimensional policy feature vector. Another approach is to use an attention-based fusion network. This network can assign different attention weights to each input feature vector to dynamically adjust their contribution to the synthesized features. For example, a self-attention module can be designed to allow the network to learn how to optimally combine contextual decision weights, dependency awareness, and environmental semantic information based on the characteristics of the input features, thereby generating a policy feature vector that can comprehensively represent the current decision context and task requirements.

[0172] Through the above technical solutions, this application can solve the problems of insufficient context-aware calibration and inadequate context-focused encoding of dependencies during feature fusion. Specifically, context-aware weight calibration of resilience scheduling weights based on environmental semantic features enables decision weights to dynamically adapt to current environmental risks, task criticality, and user intent, avoiding decision bias caused by rigid weight allocation and ensuring that the importance assessment of decisions is highly consistent with the actual context. Furthermore, attention-focused encoding of the dependency context of logical elements using contextualized decision weight vectors can identify and highlight key dependency paths in task execution, effectively improving the expression efficiency and information density of dependency features and avoiding interference from irrelevant information. Cross-domain feature synthesis of contextualized decision weight vectors, dependency-aware feature vectors, and environmental semantic features through a feature generation network achieves deep fusion and semantic unification of heterogeneous features, generating a more comprehensive, robust, and context-aware multi-dimensional policy feature vector. This optimized feature representation provides more accurate and reliable input for subsequent policy synthesis models to match differentiated elastic execution strategies, improving the accuracy, adaptability, and overall resilience of edge AI large models in making autonomous decisions in complex dynamic environments.

[0173] In some of the solutions described above in this application, differentiated flexible execution strategies are proposed to match each logical element of the decision-making task in order to optimize the decision-making process. However, in this process, the strategy synthesis may not directly integrate resource constraint information, which may lead to the feasibility risk of the generated strategy in a resource-constrained environment. It is impossible to ensure that the configuration parameters of the strategy components meet the actual resource constraints, thereby affecting the reliability and efficiency of decision execution.

[0174] To address this, this application further proposes a method for inputting the multi-dimensional policy feature vector of each logical element into a policy synthesis model, synthesizing and outputting a flexible execution policy customized for each logical element. This method includes: concatenating the multi-dimensional policy feature vector of each logical element with the resource constraint encoding corresponding to the priority partition to form a synthesizer input vector; inputting this synthesizer input vector into the feature decoupling and policy planning sub-network of the policy synthesis model, through which the policy planning sub-network outputs policy component selection instructions and component configuration parameter sets in parallel; inputting the policy component selection instructions and component configuration parameter sets into the resource feasible region verification and tuning sub-network of the policy synthesis model, through which the resource feasible region verification and tuning sub-network iteratively fine-tunes the component configuration parameters while satisfying resource constraints, and outputting the flexible execution policy.

[0175] The multi-dimensional policy feature vector for each logical element is a comprehensive representation describing the characteristics of a single logical element in the decision-making task, its association with the environment, and its dependencies within the task logic. Its purpose is to provide sufficient information for the policy synthesis model to generate customized resilient execution policies for that logical element. This vector can be a high-dimensional numerical vector containing information such as resilient scheduling weights, environmental semantic features, and dependency encodings. For example, it can be a vector of floating-point numbers, with each dimension representing a specific feature, such as importance score, risk level, or parent node dependency strength.

[0176] A policy synthesis model is a computational model capable of generating specific execution policies based on input features. Its core function is to transform abstract feature information into actionable policy instructions. This model can be implemented using various artificial intelligence technologies. For example, it can be a deep neural network model, such as the Transformer or graph neural network, which learns from a large amount of historical data to understand the mapping relationship between features and policies. Alternatively, it can be a rule-based reasoning expert system that generates policies through predefined logical rules and optimization algorithms.

[0177] A flexible execution strategy is a customized execution plan generated for specific logical elements. It aims to guide large AI models to execute decision-making tasks flexibly and adaptably under resource-constrained and dynamically changing environmental conditions. This strategy typically includes several aspects, such as specifying the computational precision of the logical element (e.g., FP32, FP16, INT8), the type of computing unit used (e.g., CPU, GPU, NPU), memory caching strategies (e.g., preloading, cache size), and execution time limits. Its goal is to maximize resource utilization efficiency and system resilience while ensuring task completion.

[0178] The resource constraint encoding corresponding to priority partitioning is data that quantifies the upper limit of resource usage allowed at a specific priority level. Its purpose is to provide explicit resource boundary information for the policy synthesis model, ensuring that the generated policy is feasible in resource allocation. This encoding can be a numerical vector, for example, containing upper limits for CPU utilization, memory usage, and power consumption. Alternatively, it can be a discrete categorical encoding representing different resource levels, such as "high priority - high resources," "medium priority - medium resources," and "low priority - low resources."

[0179] Vector concatenation is an operation that joins two or more vectors sequentially to form a longer new vector. Its purpose is to integrate interrelated feature information from different sources and of different types into a unified input format, allowing policy synthesis models to consider this information simultaneously when making decisions. For example, a vector representing the characteristics of logical elements can be directly concatenated with a vector representing resource constraints to form a comprehensive input vector containing all necessary information.

[0180] The synthesizer input vector is a comprehensive input data that, after concatenation, includes multi-dimensional policy features of logical elements and resource constraint codes corresponding to priority partitions. It forms the basis for policy generation by the policy synthesis model. The completeness of this vector directly affects the accuracy and effectiveness of policy synthesis, ensuring that the model can simultaneously consider task requirements and resource constraints when planning policies.

[0181] The feature decoupling and policy planning subnetwork is a key component of the policy synthesis model. Its function is to decompose the complex policy generation task into more manageable subtasks and plan the initial policy component selection and parameter configuration. This subnetwork can be a multi-head attention neural network, capable of simultaneously focusing on different feature dimensions in the input vector and decoupling them into independent feature streams. Alternatively, it can be a system based on decision trees or rule engines, triggering different policy planning logics based on specific combinations of input features.

[0182] The strategy component selection instruction is part of the output of the strategy planning subnetwork. It explicitly specifies which predefined resilient components should be used when executing a particular logical element. These components can be algorithm modules, computational optimization techniques, or resource management strategies. For example, it can be a binary vector where each position corresponds to a preset component, with a value of 1 indicating selection and 0 indicating no selection. Alternatively, it can be a list containing component IDs or names.

[0183] The component configuration parameter set is another part of the policy planning sub-network output, containing the specific operational settings or values ​​of the selected policy components. These parameters directly determine the component's behavior and resource consumption. For example, if the "Computation Precision Adjustment" component is selected, its parameters might include "FP16" or "INT8". If the "Memory Cache" component is selected, its parameters might include "Cache Size: 128MB" or "Cache Policy: LRU".

[0184] Parallel output refers to the policy planning subnetwork's ability to simultaneously generate policy component selection instructions and component configuration parameter sets. This parallelism improves the efficiency of policy generation and avoids the latency that may be caused by serial processing. For example, the subnetwork can be designed with two independent output layers: one for classification and component selection, and the other for regression and parameter generation, thus achieving synchronous output.

[0185] The resource feasible region verification and tuning subnetwork is another important component of the policy synthesis model. Its core function is to ensure that the generated policy is feasible under real-world resource constraints and to optimize the policy parameters accordingly. This subnetwork can be a reinforcement learning-based agent that learns how to adjust parameters to meet resource constraints and maximize performance by interacting with the environment. Alternatively, it can be a gradient descent-based optimizer that iteratively adjusts parameters by calculating the gradient of resource consumption.

[0186] "Under the premise of meeting resource constraints" means that during the adjustment of strategy parameters, preset resource limits must be strictly adhered to, such as CPU utilization, memory usage, and power consumption. This means that any modification to parameters must not cause actual or predicted resource consumption to exceed the range that the system can provide. This is a fundamental requirement to ensure the feasibility of the strategy and the stability of the system.

[0187] Iterative fine-tuning is a method that gradually optimizes policy parameters by repeatedly performing parameter adjustments and verification processes. Its purpose is to improve the performance or efficiency of the policy as much as possible while satisfying resource constraints. For example, resource consumption can be calculated based on initial parameters; if it exceeds the constraints, adjustments are made, and then the calculation and adjustment are repeated until resource consumption is within the constraints and performance reaches its optimal level or convergence conditions are met.

[0188] Through the above technical solution, this application solves the problem that when matching flexible execution strategies for each logical element of a decision-making task, the strategy synthesis process may not directly integrate resource constraint information, leading to feasibility risks in resource-constrained environments. Specifically, by concatenating the multi-dimensional strategy feature vector of each logical element with the resource constraint encoding corresponding to the priority partition to form the synthesizer input vector, this application integrates resource constraint information into the input at the initial stage of strategy synthesis, ensuring resource awareness in strategy planning. By using feature decoupling and the parallel output of strategy component selection instructions and component configuration parameter sets by the strategy planning sub-network, the collaborative and efficient generation of strategy component selection and parameter configuration is achieved. Furthermore, by using the resource feasible domain verification and tuning sub-network to iteratively fine-tune the component configuration parameters while satisfying resource constraints, this application can dynamically optimize strategy parameters, ensuring that the output flexible execution strategy not only meets task requirements but also has high feasibility and robustness in resource-constrained edge environments. This phased, resource-aware strategy synthesis mechanism improves the reliability and efficiency of autonomous decision-making by edge AI models in complex and dynamic environments, especially when resources are scarce or the environment changes suddenly, ensuring the stable execution of critical tasks.

[0189] In some of the solutions mentioned above in this application, a set of policy component selection instructions and component configuration parameters is output by a policy planning sub-network to synthesize a flexible execution policy. However, in this process, the output process may be inefficient due to sequential processing or lack of dynamic adaptation mechanism, and the component selection does not fully consider the matching degree with the feature flow, resulting in insufficient policy generation and slow response.

[0190] To address this, this application further proposes a method that outputs policy component selection instructions and component configuration parameter sets in parallel through a policy planning sub-network. Specifically, the method includes: mapping the synthesizer input vector to a structured policy feature space via the policy planning sub-network, and decoupling it into a component selection feature stream and a parameter configuration feature stream. The component selection feature stream is input into a dynamic component selection head, which generates the policy component selection instruction by calculating the adaptation attention weights between the component selection feature stream and each component in a predefined elastic component library. The parameter configuration feature stream is input into a parameter generation network, which generates initialization configuration parameters corresponding to each selected component in the policy component selection instruction, forming the component configuration parameter set.

[0191] The policy planning subnetwork is a key component of the policy synthesis model. Its main function is to process and transform the input vector to the synthesizer to generate the information needed for subsequent policy component selection and parameter configuration. This subnetwork is typically composed of multiple layers of neural networks, such as fully connected layers, convolutional layers, or recurrent neural network layers, designed to extract high-dimensional features from the input vector and perform abstract representation.

[0192] The synthesizer input vector is the raw input received by the policy synthesis model. It contains multi-dimensional policy feature vectors of the logical elements of the decision task and resource constraint encodings corresponding to priority partitions. Mapping this vector to a structured policy feature space involves transforming the raw input vector into a more organized and easily parsed internal representation through a series of nonlinear transformations. This structured space can be a high-dimensional vector space, where different dimensions or subspaces represent different aspects of the policy, such as component type, performance parameters, and resource consumption. In terms of implementation, a multilayer perceptron (MLP) can be used to progressively project the input vector onto the target feature space, or an autoencoder can be used to learn a low-dimensional or structured representation of the input vector.

[0193] After mapping to the structured policy feature space, the policy planning sub-network further decouples this structured representation into a component selection feature stream and a parameter configuration feature stream. The purpose of this decoupling is to separate the information used to select policy components from the information used to configure the parameters of these components, enabling subsequent parallel processing. The component selection feature stream focuses on describing preferences and constraints regarding component types and functional requirements under the current task and environmental conditions, while the parameter configuration feature stream focuses on describing how the selected components should be specifically configured to meet performance, resource, and other requirements. In terms of implementation, two independent output branches can be set after the output layer of the structured policy feature space, each responsible for extracting and outputting the corresponding feature stream. For example, one branch generates the component selection feature stream using a set of weight matrices, and the other branch generates the parameter configuration feature stream using a different set of weight matrices.

[0194] The dynamic component selection head is a module specifically designed to select suitable components from a predefined library of flexible components. It receives a component selection feature stream as input and interacts with the representations of each component in the predefined library to calculate fit attention weights. These fit attention weights reflect the degree of matching or relevance between the current component selection feature stream and each component in the library. Higher weights indicate a more suitable component for the current context. In implementation, the dynamic component selection head can be a neural network module based on an attention mechanism. For example, it can use the component selection feature stream as a query vector, and the embedding representation of each component in the predefined library as a key vector and a value vector. Attention weights are generated by calculating the similarity between the query and the key, and then the value vectors are weighted and summed, or the component with the highest weight is directly selected. Another implementation is to use a classifier network that takes the component selection feature stream as input, directly outputs the probability distribution of each component, and then selects the component with the highest probability.

[0195] This predefined elasticity component library is a collection of various policy components to choose from. These components can be different algorithm implementations, different model structures, different optimizer configurations, etc., each with different performance and resource consumption characteristics.

[0196] The strategy component selection instruction is the output generated by the dynamic component selection header based on the fit attention weights. It explicitly indicates which specific strategy components should be selected from the predefined resilient component library. This instruction can be a list of component IDs, a binary mask vector, or structured data containing component names and quantities.

[0197] This parameter generation network is a module specifically designed to generate initial configuration parameters for selected policy components. It receives the parameter configuration feature stream as input and, combined with the selected component information indicated in the policy component selection instruction, generates initial configuration parameters corresponding to each selected component. These parameters are specific values ​​or configuration options that the component needs to set before execution, such as learning rate, number of iterations, number of model layers, cache size, etc. In terms of implementation, the parameter generation network can be a conditional generation network, taking the parameter configuration feature stream and the identifier of the selected component as conditional inputs to generate parameters for the corresponding component. For example, a small parameter generation sub-network can be designed for each component, or a unified generation network can be used, activating different parameter generation paths based on the type of the selected component through attention or gating mechanisms.

[0198] Through the above technical solution, this application improves the efficiency and accuracy of policy generation by outputting policy component selection instructions and component configuration parameter sets in parallel through a policy planning sub-network. Mapping the synthesizer input vector to a structured policy feature space and decoupling it into a component selection feature stream and a parameter configuration feature stream effectively organizes the input data using the structured feature space and lays the foundation for subsequent parallel processing, avoiding the latency that may be caused by traditional sequential processing, thereby improving overall processing efficiency. The component selection feature stream is input into a dynamic component selection head, which dynamically evaluates the matching degree between each component in the predefined elastic component library and the feature stream by calculating the fit attention weight, thereby generating accurate and adaptable policy component selection instructions and solving the problem of insufficient matching degree between component selection and feature stream. The parameter configuration feature stream is input into a parameter generation network, which generates initial configuration parameters corresponding to each selected component in the policy component selection instruction, ensuring the relevance and consistency of the configuration. This provides a high-quality initial configuration for subsequent resource feasible domain verification and optimization, further accelerating the generation process of elastic execution policies. Overall, this solution, through structured mapping, dynamic adaptation evaluation, and parallel parameter generation, enables large AI models to generate highly adaptable and resilient execution strategies more quickly and accurately in resource-constrained and dynamically changing edge scenarios, thereby improving the resilience and reliability of decision-making.

[0199] In some of the solutions mentioned above in this application, it is proposed to iteratively fine-tune the component configuration parameters through resource feasible domain verification and tuning sub-network to output elastic execution strategies, so as to ensure that the strategies are feasible under resource constraints. However, in this process, there may be problems such as low fine-tuning efficiency, inability to converge quickly, or inaccurate control of resource over-limit, which leads to excessively long strategy optimization process and inaccurate resource allocation, thereby affecting the real-time nature of decision-making and the resilience of the system in dynamic environments.

[0200] To address this, this application further proposes a method for iteratively fine-tuning component configuration parameters under resource constraints through a resource feasible domain verification and tuning sub-network, outputting an elastic execution strategy. This method includes: combining strategy component selection instructions and a set of component configuration parameters into a current strategy configuration; calculating the multidimensional resource excess of the current strategy configuration under resource constraints using a differentiable resource consumption prediction model; calculating the gradient update direction of the component configuration parameters based on the multidimensional resource excess through backpropagation of the resource consumption prediction model; adjusting the component configuration parameters along the gradient update direction to obtain the updated strategy configuration; and iteratively executing the above calculation and adjustment process until the multidimensional resource excess meets a preset convergence condition; and outputting the strategy configuration that meets the convergence condition as an elastic execution strategy.

[0201] Specifically, combining strategy component selection instructions with a set of component configuration parameters to form the current strategy configuration aims to clearly define the strategy instance to be optimized. The strategy component selection instructions indicate the specific functional modules or algorithm components selected by the large AI model when performing decision-making tasks; for example, it may choose to use low-precision floating-point units or high-precision fixed-point units. The set of component configuration parameters contains detailed operational parameters for these selected components, such as computational precision, number of concurrent threads, memory allocation size, and model pruning rate. Combining these two elements creates a complete, executable snapshot of the strategy configuration, serving as the starting point for subsequent resource evaluation and optimization. For example, the current strategy configuration can be represented by encapsulating the component selection instructions (such as a list of component IDs) and the set of component configuration parameters (such as a key-value dictionary) into a unified data structure, such as a JSON object or a Protobuf message. Alternatively, a strategy configuration class can be defined, containing the strategy component selection instructions and the set of component configuration parameters as its member variables; the current strategy configuration object is created by instantiating this class.

[0202] A differentiable resource consumption prediction model is used to calculate the multidimensional resource overruns under resource constraints for the current policy configuration. This model is a trained machine learning model or a simulation model based on physical laws, its core characteristic being "differentiability," meaning the gradient of its output (such as CPU utilization, memory usage, power consumption, latency, etc.) relative to its input (component configuration parameters) can be calculated. This model can predict the amount of various resources consumed by the current policy configuration when running on the terminal device. Comparing the predicted resource consumption with preset resource constraints (e.g., CPU utilization limit, memory capacity limit, power budget, latency requirements) yields the multidimensional resource overruns, i.e., the portion exceeding the constraints in each resource dimension. For example, a deep neural network (such as a multilayer perceptron or graph neural network) can be trained, using policy configuration parameters as input, to predict the consumption of various resources such as CPU, memory, and power consumption, ensuring that the network structure supports automatic differentiation. Alternatively, a resource consumption simulator based on a physical model or empirical formula can be built. This simulator can simulate resource usage according to component configuration parameters, and the differentiability of the overall model can be achieved by designing the simulator's internal calculation process as a differentiable operation.

[0203] Based on multidimensional resource overruns, the gradient update direction of component configuration parameters is calculated through backpropagation of the resource consumption prediction model. Since the resource consumption prediction model is differentiable, the multidimensional resource overrun can be considered a loss function, and the goal is to minimize this loss. Using the backpropagation algorithm, the partial derivatives of the loss function with respect to each component configuration parameter can be efficiently calculated; these partial derivatives constitute the gradient. The gradient update direction indicates how to adjust the component configuration parameters to most effectively reduce resource overruns. For example, the automatic differentiation capabilities provided by mainstream deep learning frameworks such as TensorFlow and PyTorch can be used to build the resource consumption prediction model within these frameworks, defining resource overruns as the loss; the framework will automatically calculate the gradient of the loss with respect to the component configuration parameters. Alternatively, for simpler differentiable models, the gradient can be calculated through numerical differentiation (such as the finite difference method) or symbolic differentiation (derived through mathematical expressions).

[0204] The component configuration parameters are adjusted along the gradient update direction to obtain the updated policy configuration. This calculation and adjustment process is iteratively executed until the multi-dimensional resource overruns meet the preset convergence condition. After obtaining the gradient update direction, an optimization algorithm (such as gradient descent, Adam, RMSprop, etc.) is used to fine-tune the component configuration parameters along this direction with a certain learning rate. After each adjustment, a new policy configuration is formed, and the new multi-dimensional resource overruns are calculated again using the resource consumption prediction model. The gradient calculation and parameter adjustment process is repeated. This iterative process continues until the resource overruns fall within a preset tolerance range, or the preset maximum number of iterations is reached, i.e., the convergence condition is met. For example, the convergence condition can be set as the overruns of all resource dimensions being less than a certain minimum threshold (such as 1%), or the change in resource overruns in N consecutive iterations being less than a certain threshold.

[0205] The strategy configuration that meets the convergence criteria is output as a flexible execution strategy. Once the iteration process converges, it means that the current strategy configuration has been optimized and can be executed while meeting resource constraints. At this point, this optimized strategy configuration is determined as the flexible execution strategy, which can be used by the AI ​​large model for actual decision execution. For example, the converged strategy configuration (including component selection instructions and adjusted component configuration parameters) can be stored in a standard format (such as Protobuf or YAML) and passed to the execution module of the AI ​​large model. Alternatively, the optimized flexible execution strategy can be output as a return value through a predefined API interface for use by upper-layer scheduling or execution modules.

[0206] Through the above technical solution, this application introduces a differentiable resource consumption prediction model and a backpropagation mechanism, which efficiently solves the efficiency and accuracy problems in iterative fine-tuning, ensuring that the strategy quickly converges to the resource feasible region, thereby improving the real-time performance and system resilience of the decision-making process. Specifically, the strategy component selection instructions and component configuration parameter set are combined into the current strategy configuration, defining the initial starting point for fine-tuning, avoiding the uncertainty caused by random initialization, and ensuring that the optimization process starts from a reasonable configuration. The differentiable resource consumption prediction model calculates the multidimensional resource excess under resource constraints of the current strategy configuration, and directly outputs quantitative indicators using the differentiable properties of the model, solving the inefficiency problem caused by the unavailability of gradients in traditional methods. Based on this multidimensional resource excess, the gradient update direction is calculated through backpropagation. With the help of the mathematical optimization characteristics of backpropagation, the parameter adjustment direction is accurately guided, avoiding the time-consuming trial-and-error search. The component configuration parameters are adjusted along this gradient update direction to obtain the updated strategy configuration. The gradient direction ensures that resource excess is systematically reduced in each iteration. The calculation and adjustment process is iteratively executed until the convergence condition is met. This iterative mechanism, combined with the preset convergence condition, ensures that the fine-tuning process is completed within a finite number of steps, preventing infinite loops. The strategy configuration that meets the convergence condition is output as a flexible execution strategy. The optimized strategy ensures feasibility under resource constraints, thus maintaining decision reliability in dynamic environments. This method enables more precise matching of differentiated flexible execution strategies to various logical elements of a decision-making task in resource-constrained edge environments, improving the autonomous decision-making ability and system resilience of large AI models in complex and dynamic scenarios.

[0207] In response, this application further proposes steps for controlling the execution decisions of large AI models based on flexible execution strategies. This addresses the problem that existing solutions may lack a real-time dynamic adjustment mechanism for strategy execution, making it unable to effectively respond to real-time fluctuations in environmental semantic features and resource constraints. Consequently, the decision-making process suffers from insufficient stability and low resource utilization efficiency during environmental changes or resource shortages, thus affecting the overall resilience and reliability of the system.

[0208] Specifically, see Figure 7 The method for controlling the execution decisions of large AI models based on flexible execution strategies includes: 701. Analyze the elastic execution strategy and extract the execution control parameters for each logical element in the decision task.

[0209] 702. Based on the execution control parameters, dynamically configure the inference computation graph of the AI ​​large model, including assigning differentiated computation precision, computation units and memory caching strategies to different logical elements.

[0210] 703. During the execution and inference process of this large AI model, the execution control parameters are fine-tuned online based on the changes in the semantic features of the environment and the resource constraint features.

[0211] The elastic execution strategy is parsed to extract the execution control parameters for each logical element in the decision-making task. This step aims to transform the abstract elastic execution strategy into concrete instructions that the AI ​​model can directly manipulate. For example, a predefined strategy parsing module can be used, which parses the strategy according to its structured format (such as JSON, XML, or a domain-specific language DSL) to identify parameter fields corresponding to each logical element, such as computational precision level, computational unit type, and memory allocation ratio. Alternatively, a machine learning-based parser can be used, trained to identify different strategy patterns and extract key parameters, which is particularly suitable for situations where the strategy format may have some flexibility. This is the foundation for dynamic configuration, ensuring the operability of the strategy.

[0212] Based on these execution control parameters, the inference computation graph of the large AI model is dynamically configured, including assigning differentiated computational precision, computational units, and memory caching strategies to different logical elements. This step adjusts the allocation of computational resources and execution methods within the large AI model in real time according to the extracted control parameters. For example, for computational precision, the model quantization level can be adjusted (e.g., from FP32 to FP16 or INT8), or different precision pre-trained sub-models can be selected. For computational units, the allocation ratio of CPU cores, GPU cores, or dedicated AI accelerators (such as NPUs) can be dynamically scheduled. For memory caching strategies, the residence time of data for specific logical elements in the cache, the allocated cache size, or different cache eviction algorithms can be adjusted. Another implementation method is to integrate a dynamic configuration interface into the large AI model framework layer (such as TensorFlow Lite or PyTorch Mobile), allowing modification of graph node attributes, operator implementations, or memory manager parameters at runtime. For example, different operator implementations (high precision / low precision) can be selected based on parameters, or the parallelism of specific layers can be adjusted. This achieves efficient resource utilization and priority guarantees, and is a direct means of addressing resource constraints and environmental changes.

[0213] During the inference process of this large AI model, the execution control parameters are fine-tuned online based on changes in environmental semantic features and resource constraint features. This step aims to adjust the configured execution parameters in real-time and with fine granularity based on the real-time perceived environmental and resource status during model operation. For example, this can be achieved through a lightweight online learning or adaptive control module. This module continuously monitors environmental semantic features (such as increased risk level) and resource constraint features (such as decreased available memory), and makes small adjustments or switches to execution control parameters such as computational accuracy, computational unit allocation, or memory caching strategies according to preset fine-tuning rules or reinforcement learning strategies. Alternatively, a feedback control mechanism can be used, where environmental semantic features and resource constraint features are used as feedback signals input to a PID controller or fuzzy controller. The controller outputs parameter adjustments that directly affect the execution control parameters. For example, when the environmental risk level suddenly increases, the computational accuracy of critical logic elements is increased. When the equipment temperature is too high, the computational unit allocation of some non-critical logic elements is reduced. This ensures that the system can maintain stable and efficient operation in dynamic and uncertain environments, which is key to improving system resilience.

[0214] Through the above technical solution, this application can solve the problem of the lack of a real-time dynamic adjustment mechanism for strategy execution in related technologies. Specifically, by parsing the elastic execution strategy and extracting the execution control parameters for each logical element, the abstract strategy instructions are transformed into specific configurations that the AI ​​large model can directly operate on. This allows each logical element to receive customized control based on its importance, avoiding resource waste or delays in critical tasks caused by uniform processing. Based on this, the inference computation graph of the AI ​​large model is dynamically configured according to these execution control parameters, including assigning differentiated computational precision, computational units, and memory caching strategies to different logical elements, achieving efficient resource allocation and priority processing. This ensures that when resources are limited, critical logical elements can obtain more computational resources, while non-critical elements can save energy by reducing precision, thereby optimizing overall decision-making efficiency. Furthermore, during the execution and inference process of the AI ​​large model, the execution control parameters are fine-tuned online based on real-time changes in environmental semantic features and resource constraint features. This allows the system to monitor the external environment and internal resource status in real time and dynamically adjust parameters to cope with sudden changes. This online fine-tuning mechanism effectively prevents fixed parameters from failing in dynamic scenarios, maintains the continuity and reliability of the decision-making process, and greatly enhances the system's anti-interference capability and overall resilience. In summary, by introducing dynamic configuration and online fine-tuning mechanisms, this application ensures that the AI ​​large model can adapt to environmental changes and resource constraints in real time when executing decisions, improving the system's response speed, decision stability, and resource utilization efficiency under uncertain conditions.

[0215] The following example will provide a more detailed explanation of the above technical solution: Imagine a large-scale edge AI model deployed on a drone, whose core task is to perform autonomous inspections in complex urban environments. This drone faces multiple challenges, including limited battery life, scarce computing resources, volatile external environments (such as sudden weather changes and unexpected obstacles), and complex inspection task logic (such as path planning, target recognition, anomaly detection, and data transmission). Related technologies often suffer from insufficient inter-module coordination when dealing with these constraints, leading to decreased inspection efficiency and decision-making reliability during resource fluctuations or environmental changes. This solution aims to address these issues, ensuring that the drone achieves system-level resilient autonomous decision-making under conditions of multiple uncertainties.

[0216] This method acquires and processes the drone's internal operational status data, the logical structure data of the decision-making tasks to be executed by the AI ​​large model, external environment perception data, and user command data, respectively obtaining resource constraint features, task logical features, and environmental semantic features. Specifically: Internal operational data of the drone, such as real-time CPU / GPU load, memory usage, battery level, and sensor health status, undergoes predictive and performance degradation analysis. For example, time-series prediction is performed on real-time resource monitoring sequences to obtain predicted resource availability values ​​for a future period. Performance degradation trend models are built on historical operational parameters to generate equipment health assessment values. These quantitative indicators, such as predicted resource availability values ​​and equipment health assessment values, are weighted and fused to form resource constraint features, used to characterize the current and future resource scarcity of the drone.

[0217] The logical structure data of the decision-making tasks to be executed by the AI ​​large-scale model, such as the decomposition of an inspection task into multiple logical elements such as path planning, target recognition, data analysis, and communication feedback, as well as the order and dependencies between them, are analyzed. This method performs causal deconstruction analysis on this logical structure data, identifies each logical element and its causal dependencies, and constructs a directed acyclic graph (DAG) to represent the reasoning logic. Simultaneously, based on historical data or domain knowledge, initial importance weights are assigned to each logical element in the graph. This DAG and the associated initial importance weights are collectively defined as the task logical features.

[0218] External environmental perception data includes multimodal data from cameras, LiDAR, GPS, weather sensors, etc., as well as user command data (e.g., commands sent by users via ground stations such as "emergency return" or "prioritize completing area A inspection"). This method extracts and fuses multimodal features from this data to generate a comprehensive environmental feature vector. This vector is then input into a miniature risk identification model deployed on the edge, outputting an environmental risk level (e.g., "low risk," "medium risk," "high risk"). This level is matched against a pre-defined scenario safety rule base to output the task scenario criticality (e.g., "general area," "critical area," "no-fly zone"). Real-time semantic parsing of user commands determines the urgency of the user's intent. These environmental risk levels, task scenario criticality, and urgency of the user's intent collectively constitute the environmental semantic features.

[0219] Based on environmental semantic features, the importance of each logical element in the task logical features is modulated to generate environmentally adaptive importance weights. Based on resource constraint features, the resource sensitivity of the environmentally adaptive importance weights is calibrated to generate resilient scheduling weights and resource allocation focusing signals.

[0220] In the environmentally adaptive importance weight generation stage, this method performs dynamic semantic gain calculation based on the environmental risk level and task scenario criticality in the environmental semantic features, generating a semantically perceptual gain factor. For example, when a drone enters a high-risk area (high environmental risk level) or performs critical area inspection (high task scenario criticality), a larger semantically perceptual gain factor is calculated. Then, based on this gain factor and the initial importance weights of each logical element, a non-uniform weight modulation field is constructed. This modulation field defines differentiated modulation intensities applicable to different initial importance weights. For example, when the risk is high, the initial importance weights of safety-related logical elements (such as obstacle avoidance) will be significantly amplified, while weights related to non-critical data analysis may be moderately suppressed. This weight modulation field is used to perform field-driven modulation on the initial importance weights of each logical element to generate environmentally adaptive importance weights.

[0221] In the resilient scheduling weight and resource allocation focus signal generation stage, this method generates dynamic focus control parameters based on resource constraint characteristics (such as low battery power and CPU overload) and the distribution characteristics of environmentally adaptive importance weights. For example, when resources are scarce, the focus control parameters instruct the system to concentrate resources. Based on these focus control parameters, a Pareto-based nonlinear sharpening process is performed on the environmentally adaptive importance weights to generate resilient scheduling weights. This sharpening process makes the distribution of resilient scheduling weights more Pareto-like when resource constraints increase, meaning that a few key logical elements receive extremely high weights, while the weights of most non-key elements are significantly compressed. This differs from the simple linear weight adjustment in related technologies; this scheme can more effectively highlight core tasks when resources are limited. For example, when battery power is extremely low, the resilient scheduling weight of the "emergency landing" logical element in path planning is sharpened to the highest level, while the weight of "refined target recognition" is significantly reduced. Simultaneously, based on the distribution of resilience scheduling weights and focus control parameters, a resource allocation focus signal is generated. This signal is used to prioritize high-weight logical elements during resource allocation. For example, logical elements with resilience scheduling weights greater than the dynamic focus threshold are identified as a set of key logical elements and encapsulated as structured instructions.

[0222] Based on resilient scheduling weights, resource allocation focus signals, task logical characteristics, and environmental semantic characteristics, differentiated elastic execution strategies are matched for each logical element of the decision-making task.

[0223] Based on resource allocation focus signals, priority partitioning of logical elements to be executed in decision-making tasks is determined. For example, according to the focus signals, "emergency landing" is classified as the highest priority, "obstacle avoidance" as a high priority, "routine target identification" as a medium priority, and "non-critical data feedback" as a low priority. Compared with related technologies that lack dynamic priority partitioning, this approach can more flexibly respond to unexpected situations.

[0224] For any logical element among the logical elements to be executed, such as "target recognition," this method fuses its corresponding resilience scheduling weights, environmental semantic features related to that logical element (such as environmental risk level and task scenario criticality), and the logical element's dependency in the task logical features (such as dependency on the "image acquisition" logical element) to generate a multi-dimensional policy feature vector for that logical element. For example, during the fusion process, context-aware weight calibration is performed on the resilience scheduling weights based on environmental semantic features to obtain a contextualized decision weight vector. This vector is then used to perform attention-focused encoding on the dependency context extracted from the task logical features to generate a dependency-aware feature vector.

[0225] The multi-dimensional policy feature vector of each logical element is input into the policy synthesis model, which synthesizes and outputs a flexible execution policy customized for each logical element. This policy synthesis model constrains the resource consumption limit of the synthesized policy based on priority partitioning. For example, for a high-priority "obstacle avoidance" logical element, the policy synthesis model might match an execution policy of "high precision, low latency, and GPU core usage." For a low-priority "non-critical data backhaul," a policy of "low power consumption, background operation, and CPU edge core usage" might be matched. During policy synthesis, the policy synthesis model outputs policy component selection instructions (such as selecting a "lightweight object detection algorithm" or a "full-featured object detection algorithm") and component configuration parameter sets (such as detection frame rate and model quantization accuracy) in parallel through feature decoupling and policy planning sub-networks. These instructions and parameters are input into the resource feasible domain verification and tuning sub-network, which iteratively fine-tunes the component configuration parameters while satisfying resource constraints until the resource limits are exceeded and the preset convergence condition is met, at which point a flexible execution policy is output. For example, if the initial configuration leads to resource overruns, the subnetwork will backpropagate to calculate gradients and fine-tune parameters (such as reducing the detection frame rate or further quantizing the model) until the drone's battery and computing power constraints are met.

[0226] Control the execution decisions of the AI ​​large model based on a flexible execution strategy.

[0227] This method analyzes the flexible execution strategy and extracts the execution control parameters for each logical element in the decision-making task. For example, the "obstacle avoidance" logical element is allocated FP16 computational precision, uses the NPU for computation, and reserves a large memory cache. Conversely, the "background environment analysis" logical element is allocated INT8 computational precision, uses the CPU for computation, and employs a smaller memory cache strategy. This differentiated configuration maximizes the use of limited edge resources, ensuring the performance of critical tasks. During the execution and inference of the large AI model, the execution control parameters are fine-tuned online based on changes in environmental semantic features and resource constraints. For example, if the drone suddenly detects a further decrease in battery power, the system immediately fine-tunes the strategy, further reducing the computational precision of non-critical tasks or suspending their execution to ensure the continuous operation of core safety tasks. This online fine-tuning mechanism enables the drone to continuously adapt to dynamically changing environmental and resource conditions, improving its decision-making resilience and operational stability in complex scenarios, overcoming the limitations of static or simple dynamic adjustments in related technologies.

[0228] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0229] Figure 8 This is a schematic diagram of the structure of an intelligent autonomous decision-making system based on a large edge AI model provided in an embodiment of this application. See also... Figure 8 The system includes: The acquisition module 801 is used to acquire and process the internal operating status data of the terminal device, the logical structure data of the decision-making tasks to be executed by the AI ​​big model, the external environment perception data, and the user instruction data, and obtain resource constraint features, task logic features, and environmental semantic features respectively.

[0230] The modulation module 802 is used to modulate the importance of each logical element in the task's logical features based on the semantic features of the environment, generate environment-adaptive importance weights, and perform resource sensitivity calibration on the environment-adaptive importance weights based on the resource constraint features, generating resilient scheduling weights and resource allocation focusing signals.

[0231] The matching module 803 is used to match differentiated elastic execution strategies for each logical element of the decision task based on the resilience scheduling weight, the resource allocation focus signal, the task logical characteristics, and the environmental semantic characteristics.

[0232] The execution module 804 is used to control the execution decisions of the AI ​​large model according to the elastic execution strategy.

[0233] It should be noted that the intelligent autonomous decision-making system based on edge AI large model provided in the above embodiments is only illustrated by the division of the above functional modules when making autonomous decisions. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. In addition, the intelligent autonomous decision-making system based on edge AI large model provided in the above embodiments and the intelligent autonomous decision-making method embodiment based on edge AI large model belong to the same concept. The specific implementation process is detailed in the method embodiment, and will not be repeated here.

[0234] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including a computer program that can be executed by a processor to perform the intelligent autonomous decision-making method based on a large edge AI model in the above embodiments. For example, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, and optical data storage device, etc.

[0235] In an exemplary embodiment, a computer program product or computer program is also provided, which includes program code stored in a computer-readable storage medium. The processor of a computer device reads the program code from the computer-readable storage medium and executes the program code, causing the computer device to execute the above-described intelligent autonomous decision-making method based on an edge AI big model.

[0236] In some embodiments, the computer program involved in the present application embodiments may be deployed and executed on a computer device, or executed on multiple computer devices located in one location, or executed on multiple computer devices distributed in multiple locations and interconnected through a communication network. Multiple computer devices distributed in multiple locations and interconnected through a communication network may constitute a blockchain system.

[0237] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0238] The above are merely optional embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An intelligent autonomous decision-making method based on a large edge AI model, characterized in that, The method includes: The system acquires and processes the internal operating status data of the terminal device, the logical structure data of the decision-making tasks to be executed by the AI ​​big model, the external environment perception data, and the user instruction data, and obtains resource constraint features, task logic features, and environmental semantic features, respectively. Based on the environmental semantic features, the importance of each logical element in the task logical features is modulated to generate an environment-adaptive importance weight. Based on the resource constraint features, the resource sensitivity of the environment-adaptive importance weight is calibrated to generate a resilient scheduling weight and a resource allocation focus signal. Based on the resilient scheduling weights, the resource allocation focus signals, the task logic features, and the environmental semantic features, differentiated elastic execution strategies are matched for each logical element of the decision task. The AI ​​large model is controlled to make execution decisions based on the elastic execution strategy.

2. The method according to claim 1, characterized in that, The acquisition and processing of the terminal device's internal operating status data, the logical structure data of the decision-making tasks to be executed by the AI ​​large model, the external environment perception data, and the user instruction data yields resource constraint features, task logical features, and environmental semantic features, including: Predictive analysis and performance degradation analysis are performed on the internal operating status data to generate quantitative indicators, including resource availability prediction values ​​and equipment health assessment values, to form the resource constraint characteristics. The logical structure data is modeled using causal logic and analyzed for importance to generate logical relationships and weight information, including the dependencies between each step in the decision-making task and the initial importance weights, in order to form the logical features of the task. Risk semantic recognition, scenario criticality assessment, and intent parsing are performed on the external environment perception data and the user instruction data to generate semantic assessment results including environmental risk level, task scenario criticality, and user intent urgency, thus forming the environmental semantic features.

3. The method according to claim 1, characterized in that, The step of modulating the importance of each logical element in the task logical features based on the environmental semantic features to generate environmentally adaptive importance weights includes: Based on the environmental risk level and task scenario criticality in the environmental semantic features, dynamic gain calculation of semantic perception is performed to generate semantic perception gain factor. Based on the semantically aware gain factor, a non-uniform weight modulation field is constructed, which defines a differentiated modulation intensity applicable to different initial importance weights. The initial importance weights of each logical element in the task logic features are field-driven modulated using the weight modulation field to generate the environment-adaptive importance weights.

4. The method according to claim 3, characterized in that, The step of using the weight modulation field to perform field-driven modulation on the initial importance weights of each logical element in the task logical features to generate the environment-adaptive importance weights includes: The initial importance weights of each logical element are input into the weight modulation field, and the field intensity modulation values ​​acting on each logical element are parsed out. Based on the field emphasis control value, a dynamic recalibration operation is performed on the initial importance weight of each logical element to generate the environment adaptive weight corresponding to each logical element. The environment-adaptive weights of all logical elements are aggregated in an ordered manner to obtain the environment-adaptive importance weights.

5. The method according to claim 1, characterized in that, The step of calibrating the importance weights of the environment adaptation based on the resource constraint characteristics to generate resilient scheduling weights and resource allocation focusing signals includes: Based on the distribution characteristics of the resource constraint features and the importance weights of the environmental adaptation, dynamic focusing control parameters are generated. Based on the focus control parameters, a nonlinear sharpening process based on the Pareto principle is performed on the importance weights of the environment adaptation to generate the resilient scheduling weights. The nonlinear sharpening process is configured to make the distribution of the resilient scheduling weights satisfy the Pareto distribution when resource constraints are enhanced. Based on the distribution of the resilience scheduling weights and the focus control parameters, the resource allocation focus signal is generated. The resource allocation focus signal is used to prioritize high-weight logical elements during resource allocation.

6. The method according to claim 5, characterized in that, The process of performing Pareto-based nonlinear sharpening on the importance weights of the environment adaptation based on the focus control parameters to generate the resilience scheduling weights includes: Based on the focus control parameters, the sharpening intensity parameters used to modulate the nonlinear transform function are determined; Based on the sharpening intensity parameter, the transformation curve of the nonlinear transformation function is configured, wherein the transformation curve is configured such that the output gain for high importance weights is higher than the output gain for low importance weights. The importance weights of the environment adaptation are transformed and calculated using the configured nonlinear transformation function to obtain the resilience scheduling weights.

7. The method according to claim 1, characterized in that, The step of matching differentiated elastic execution strategies for each logical element of the decision-making task based on the resilient scheduling weights, the resource allocation focus signals, the task logical features, and the environmental semantic features includes: Based on the resource allocation focus signal, the priority partitioning of the logical elements to be executed in the decision task is determined; For any logical element among the logical elements to be executed, the resilient scheduling weight corresponding to the logical element, the environmental semantic features related to the logical element, and the dependency relationship of the logical element in the task logical features are fused to generate a multi-dimensional policy feature vector of the logical element. The multi-dimensional strategy feature vector of each logical element is input into the strategy synthesis model, which synthesizes and outputs a flexible execution strategy customized for each logical element. The strategy synthesis model constrains the resource consumption limit of the synthesized strategy based on the priority partition.

8. The method according to claim 7, characterized in that, The process of inputting the multi-dimensional policy feature vector of each logical element into the policy synthesis model, and synthesizing and outputting a flexible execution policy customized for each logical element, includes: The multi-dimensional strategy feature vector of each logical element is concatenated with the resource constraint code corresponding to the priority partition to form the synthesizer input vector. The synthesizer input vector is input into the feature decoupling and policy planning subnetwork of the policy synthesis model, and the policy planning subnetwork outputs policy component selection instructions and component configuration parameter sets in parallel. The strategy component selection instruction and the set of component configuration parameters are input into the resource feasible domain verification and tuning subnetwork of the strategy synthesis model. The resource feasible domain verification and tuning subnetwork iteratively fine-tunes the component configuration parameters under the premise of satisfying resource constraints, and outputs the elastic execution strategy.

9. The method according to claim 8, characterized in that, The step of iteratively fine-tuning the component configuration parameters through the resource feasible domain verification and tuning sub-network under the premise of satisfying resource constraints, and outputting the elastic execution strategy, includes: The strategy component selection instruction and the component configuration parameter set are combined to form the current strategy configuration; The multidimensional resource over-limit of the current strategy configuration under the resource constraints is calculated using a differentiable resource consumption prediction model. Based on the multidimensional resource overlimit, the gradient update direction of the component configuration parameters is obtained through backpropagation calculation of the resource consumption prediction model; The component configuration parameters are adjusted along the gradient update direction to obtain the updated strategy configuration, and the above calculation and adjustment process is iteratively executed until the multidimensional resource exceeds the limit and meets the preset convergence condition. The strategy configuration that satisfies the convergence condition is output as the elastic execution strategy.

10. The method according to claim 1, characterized in that, The process of controlling the execution decisions of the AI ​​large model based on the elastic execution strategy includes: The elastic execution strategy is analyzed to extract the execution control parameters for each logical element in the decision task; Based on the execution control parameters, the inference computation graph of the AI ​​large model is dynamically configured, including assigning differentiated computation precision, computation units and memory caching strategies to different logical elements; During the execution and inference process of the AI ​​large model, the execution control parameters are fine-tuned online based on the changes in the environmental semantic features and the resource constraint features.

Images