Adaptive lightweight task execution and optimization method, device, equipment and medium

Abstract

The application relates to the technical field of artificial intelligence, can be applied to business scenarios such as financial technology and medical health, and discloses a self-adaptive light-weight task execution and optimization method, device, equipment and medium, which comprises the following steps: collecting multi-dimensional sensing data to generate a scene, a task and resource characteristics, inputting a multi-target decision model to obtain light-weight strategy parameters, reconstructing and parameter configuring an initial model to generate a light-weight execution model, using the light-weight execution model to complete reasoning to obtain a task processing result, generating running performance data and generating a running effect analysis level according to the processing result, and adjusting a strategy or updating a decision model based on the analysis level for subsequent task processing. The application dynamically generates and continuously updates a light-weight strategy by driving multi-dimensional sensing, so that the model keeps the balance between precision and resource consumption under different environments, and the task execution capability under limited computing power is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an adaptive lightweight task execution and optimization method, apparatus, device, and medium. Background Technology

[0002] With the rapid penetration of embodied intelligence and intelligent task processing technologies into multiple industries, real-time inference capabilities, lightweight deployment requirements, and edge execution efficiency are facing increasingly stringent demands.

[0003] In the fintech business, such as smart teller assistants, mobile risk control terminals, remote loan approval and field visit terminals, a large number of tasks need to be completed in real time on mobile devices or resource-constrained terminals for intelligent identification, processing judgment, and execution of action instructions. However, most existing lightweight technologies rely on static configuration or pre-built model compression strategies, which cannot adjust the execution process according to changes in task priority, fluctuations in data complexity, or device resource status, resulting in wasted computing power, excessive energy consumption, or decreased accuracy of business judgments. In scenarios with increased task density or aging equipment, these statically configured models often experience response delays or even inference timeouts, seriously affecting the continuity of intelligent field visit decisions, risk assessment, and credit processes.

[0004] In the healthcare field, applications such as ward patrol robots, rehabilitation assistive arms, and intelligent diagnostic and treatment auxiliary equipment are crucial due to significant dynamic changes in the environment, diverse task types, and high human safety constraints. Effective management of hardware resources and real-time adaptation of inference models are paramount. However, existing lightweight solutions often lack the ability to perceive the varying real-time inference capabilities required for different task objectives (such as localization, trajectory planning, and state assessment). They also fail to comprehensively consider key factors such as the complexity of the ward environment, the difficulty of obtaining human motion information, and device power consumption. This results in model generation strategies being highly dependent on experience and uncertain, leading to insufficient model accuracy at times and redundant, idle computing power at other times. Furthermore, medical intelligent devices typically operate on battery power, and the lack of dynamic adaptation can easily trigger insufficient battery life or forced task interruptions, thus affecting the continuous execution of medical assistance tasks. Summary of the Invention

[0005] The main objective of this invention is to provide an adaptive lightweight task execution and optimization method, apparatus, device, and storage medium, aiming to solve the technical problem that the existing technology lacks the ability to dynamically adjust the lightweight strategy based on real-time scenarios, task requirements, and device resource status, thus failing to achieve an adaptive balance between accuracy, computing power consumption, and energy consumption.

[0006] To achieve the above objectives, the present invention provides an adaptive lightweight task execution and optimization method, comprising: Collect multidimensional sensing data and extract features from the multidimensional sensing data to generate scene complexity features, task attribute features and resource redundancy features. The scenario complexity features, the task attribute features, and the resource redundancy features are input into a multi-objective decision model, and a lightweight strategy parameter set is generated through multi-objective optimization processing. Based on the lightweight strategy parameter set, the initial model is reconstructed and the running parameters are configured to generate a lightweight execution model. Obtain the target task input data, and use the lightweight execution model to perform inference processing on the target task input data to output the target task processing result; The task execution accuracy value is determined based on the target task processing result, and the computing power consumption and energy consumption value of the lightweight execution model during the execution inference process are monitored. Based on the task execution accuracy value, the computing power consumption value and the energy consumption value, the running performance data is generated. The operational performance data is compared with a preset performance benchmark threshold, and an operational performance analysis level is generated based on the comparison results. In response to the operational performance analysis level, a correction operation is performed on the lightweight strategy parameter set or a weight update operation is performed on the multi-objective decision model. The corrected parameters or the updated model are then applied to the lightweight execution model generation process at the next time step to optimize the processing of subsequent target tasks.

[0007] Furthermore, to achieve the above objectives, the present invention provides an adaptive lightweight task execution and optimization device, comprising: The multidimensional perception feature extraction module is used to collect multidimensional perception data and extract features from the multidimensional perception data to generate scene complexity features, task attribute features and resource redundancy features. The multi-objective decision generation module is used to input the scene complexity features, the task attribute features and the resource redundancy features into the multi-objective decision model, and generate a lightweight strategy parameter set through multi-objective optimization processing. The lightweight model construction module is used to perform model reconstruction and runtime parameter configuration on the initial model according to the lightweight strategy parameter set, and generate a lightweight execution model. The task reasoning and execution module is used to acquire target task input data, and use the lightweight execution model to perform reasoning processing on the target task input data to output the target task processing result; The execution performance monitoring module is used to determine the task execution accuracy value based on the target task processing result, and monitor the computing power consumption and energy consumption value of the lightweight execution model during the execution inference process, and generate running performance data based on the task execution accuracy value, the computing power consumption value and the energy consumption value; The performance evaluation module is used to compare the performance data with a preset performance benchmark threshold and generate a performance analysis level based on the comparison results. The strategy adaptive optimization module is used to perform a correction operation on the lightweight strategy parameter set or a weight update operation on the multi-objective decision model in response to the performance analysis level, and apply the corrected parameters or updated model to the lightweight execution model generation process at the next time step to optimize the processing of subsequent target tasks.

[0008] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and an adaptive lightweight task execution and optimization program stored in the memory and executable on the processor, wherein when the adaptive lightweight task execution and optimization program is executed by the processor, it implements the steps of the adaptive lightweight task execution and optimization method as described above.

[0009] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing an adaptive lightweight task execution and optimization program, wherein the adaptive lightweight task execution and optimization program, when executed by a processor, implements the steps of the adaptive lightweight task execution and optimization method as described above.

[0010] Beneficial Effects: This invention relates to the field of artificial intelligence technology and can be applied to business scenarios such as fintech and healthcare. It discloses an adaptive lightweight task execution and optimization method, apparatus, device, and medium, comprising: collecting multi-dimensional perception data to generate scene complexity features, task attribute features, and resource redundancy features; inputting these three types of features into a multi-objective decision model to generate a lightweight strategy parameter set; reconstructing and configuring the initial model based on the strategy parameters to obtain a lightweight execution model; using the lightweight execution model to perform inference to obtain the target task processing result; determining the task execution accuracy based on the processing result and monitoring computing power usage and energy consumption to generate operational performance data; comparing the operational performance data with a performance benchmark to generate an operational effect analysis level; and performing strategy parameter correction or decision model weight update based on the analysis level to optimize subsequent task processing. This invention generates a lightweight strategy by integrating real-time scenarios, task requirements, and resource status, forming a dynamic update closed loop, enabling the lightweight execution model to adaptively balance accuracy, computing power, and energy consumption, ensuring task execution performance in resource-constrained environments. Attached Figure Description

[0011] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of an application environment for an adaptive lightweight task execution and optimization method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating an embodiment of the adaptive lightweight task execution and optimization method of the present invention; Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the adaptive lightweight task execution and optimization device of the present invention; Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0012] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0013] The adaptive lightweight task execution and optimization method provided in this invention can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server can collect multi-dimensional perception data from the client to generate scene complexity features, task attribute features, and resource redundancy features. These three types of features are input into a multi-objective decision model to generate a lightweight strategy parameter set. Based on the strategy parameters, the initial model is reconstructed and parameters are configured to obtain a lightweight execution model. The lightweight execution model is used to perform inference to obtain the target task processing result. Based on the processing result, the task execution accuracy is determined, and computing power consumption and energy consumption are monitored to generate runtime performance data. The runtime performance data is compared with a performance benchmark to generate a runtime effect analysis level. Based on the analysis level, strategy parameters are corrected or decision model weights are updated to optimize subsequent task processing. This invention generates a lightweight strategy by integrating real-time scene, task requirements, and resource status, forming a dynamic update closed loop. This enables the lightweight execution model to adaptively balance accuracy, computing power, and energy consumption, ensuring task execution performance in resource-constrained environments. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The following detailed description of specific embodiments further illustrates this invention.

[0014] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the adaptive lightweight task execution and optimization method provided by the present invention. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0015] like Figure 2 As shown, the adaptive lightweight task execution and optimization method proposed in this invention includes the following steps: S10, collect multi-dimensional perception data, and extract features from the multi-dimensional perception data to generate scene complexity features, task attribute features and resource redundancy features. In this embodiment, the process of collecting multidimensional sensing data integrates inputs from three sources: environment, task, and device. "Multidimensional" refers to different source types, attributes, or time spans; these can be real-time sensor signals, task scheduling systems, or resource monitoring modules. Environmental sensing data describes spatial state and dynamic changes and may include visual images, distance information, or illumination measurements. Task scheduling data reflects the priority, requirements, or limitations of the current execution objective and can be formed from scheduling instructions, task descriptions, or parameter configurations. Resource monitoring data characterizes the device's available computing power, power consumption, or storage status, and its operating status can be queried through the system interface.

[0016] Feature extraction transforms raw input into structured representations. Scene complexity features originate from collected environmental data, which can be determined by quantity measurements, speed changes, or lighting intensity to determine the magnitude of environmental variations. These variations can be weighted or combined to form a single representation. Task attribute features are derived from task scheduling information, extracting key parameters from task type, priority, or required accuracy to describe the current task's requirements and stringency on the model output. Resource redundancy features are derived from operational status information, generated from hardware occupancy, remaining power, or computing power margin, reflecting the actual capacity of the device when performing tasks.

[0017] The three features remain independent of their source but consistent in purpose during the extraction process, respectively capturing environmental pressure, task requirements, and equipment capabilities, enabling the input data to form a three-dimensional representation that can be directly used for subsequent inference and adjustment.

[0018] In deployments dominated by environmental variables, multidimensional data is primarily collected from cameras, light sensors, and distance sensors. Convolutional structures can be used to extract the number of visual targets, inter-frame differences can be used to assess speed changes, and standardization on the illumination curve can yield scene complexity. In deployments with strong task-driven characteristics, task scheduling data can be analyzed in detail, mapping task priority fields to interval values ​​and converting task accuracy requirements into threshold indicators to form task attribute features. In resource-constrained operating environments, lightweight monitoring strategies can sample CPU utilization, power consumption changes, and memory usage, generating resource redundancy representations through weighted combinations.

[0019] Different implementation methods can adjust the proportion of data sources. For example, when the task changes drastically, the frequency of scheduling and parsing can be increased, and when the resource changes are sensitive, the collection latency can be reduced, thereby maintaining the accuracy of feature representation under different conditions.

[0020] This embodiment extracts scene complexity features, task attribute features, and resource redundancy features simultaneously, forming a multi-dimensional input that reflects environmental requirements, task needs, and equipment capability status. This enables subsequent inference and execution to be based on real-time and comprehensive data, helping to avoid resource waste in simple scenarios and insufficient capability in complex scenarios.

[0021] S20, input the scene complexity features, the task attribute features and the resource redundancy features into the multi-objective decision model, and generate a lightweight strategy parameter set through multi-objective optimization processing; In this embodiment, scene complexity features, task attribute features, and resource redundancy features are input into the action of the multi-objective decision-making model to fuse the three types of features into a unified decision space. Scene complexity features express the intensity of changes in environmental factors, and may include spatial congestion, the number of dynamic targets, or the magnitude of visual interference. Task attribute features express the expected accuracy and response of the current task, defined by task level, priority indicators, or target accuracy. Resource redundancy features express the available computing power, remaining energy, or processing capacity of the device when executing the task.

[0022] After receiving three types of features, the multi-objective decision-making model constructs an evaluation basis through parameter mapping, weight selection, or rule combination, enabling the input to trigger different trade-off strategies. Multi-objective optimization incorporates accuracy requirements, computational burden, and energy pressure into the same evaluation system, forming a more suitable adjustment result through screening, comparison, or convergence in the candidate strategy space. The lightweight strategy parameter set is generated from the optimization results and describes one or more configuration dimensions, including quantization bit width, pruning size, inference speed, or module start / stop status.

[0023] In one implementation, the multi-objective decision-making model adopts a hierarchical structure. The first layer standardizes and unifies the scale of three types of features. The second layer maps task attribute features to performance requirement weights, scene complexity features to accuracy risk weights, and resource redundancy features to hardware and software tolerance scales. The third layer generates candidate strategies through parameter combinations. In another implementation, a cost-driven approach can be used to set mutual checks and balances for different optimization objectives, automatically lowering the pruning threshold to improve operational efficiency when resources are scarce. An incremental configuration approach can also be used, generating a new parameter set only when the input features cross a preset range, thereby reducing the additional overhead caused by frequent reconfiguration.

[0024] In this embodiment, the three types of features are uniformly input and participate in multi-objective optimization, enabling the model configuration selection to simultaneously consider environmental changes, task requirements, and resource capabilities. This allows lightweight strategy generation to no longer rely solely on a single dimension, thereby reducing the risk of accuracy degradation and minimizing computational and energy waste.

[0025] S30, perform model reconstruction and runtime parameter configuration on the initial model according to the lightweight strategy parameter set to generate a lightweight execution model; In this embodiment, the model reconstruction action performed on the initial model according to the lightweight strategy parameter set is used to match the inference capability with the current environment, task, and resource conditions. The lightweight strategy parameter set may include quantization bit width, pruning ratio, inference speed parameters, and module start / stop flags. Each item is used to limit the structural form of the model in execution or the internal resource allocation method. The initial model is an unadjusted inference network, containing complete weights, nodes, and inference paths. Model reconstruction of the initial model can be carried out from directions such as weight precision compression, operator pruning, disabling subnetworks, or modifying feature paths to reduce computational burden. The runtime parameter configuration action is used to implement the content in the parameter set that does not affect the structure but affects the execution environment, such as the inference interval, the number of execution threads, or the cache size setting.

[0026] Quantization reduces the storage and arithmetic costs of weights and activations by decreasing the precision of numerical representation. Pruning reduces computational paths by removing neural layers, channels, or attention branches. Module management controls the participation of unnecessary branches in the network to adapt to real-time resource availability. Runtime parameter configuration loads inference-related time, scheduling, or hardware utilization items, enabling the restructured model to enter the real-world execution state. The lightweight execution model, formed by the combined effects of model reconstruction and runtime configuration, possesses lower resource consumption while meeting task requirements for inference capabilities.

[0027] In one implementation, the quantization bit width converts floating-point weights to a fixed bit width through a linear mapping, and integer multiplication and addition replace floating-point operations in inference. In another implementation, the pruning ratio can be adjusted by filtering the weight importance distribution and removing nodes with lower weight contributions, allowing the network size to dynamically change with resource conditions. Explicit start / stop control can also be used in module management to enable critical inference paths and disable auxiliary modules, such as disabling feature enhancement paths to reduce latency. Regarding runtime parameter configuration, a fixed value can be set for the inference call interval to strictly control the continuous inference interval, or an upper limit for execution threads can be set via a hardware interface to avoid resource contention.

[0028] This embodiment drives model reconstruction and runtime configuration through a lightweight strategy parameter set, enabling the inference network to remain operational even under scenarios with limited hardware resources or changing task pressure, thereby reducing computing power overhead while maintaining output quality and response speed.

[0029] S40, Obtain target task input data, and use the lightweight execution model to perform inference processing on the target task input data to output target task processing results; In this embodiment, the action of acquiring target task input data serves to provide a real-time information source for inference execution. The data source may include sensor acquisition streams, message middleware push streams, or the latest observation records cached locally. Target task input data is generally presented as an unprocessed dataset, potentially including image frames, LiDAR point fragments, audio clips, text states, or multimodal combinations, reflecting the entity, environment, and task states in the application scenario. Data acquisition can be achieved through device drivers, communication interfaces, kernel buffer reading, or remote data transmission. Dimensional adjustments are performed according to the input dimension requirements of the lightweight execution model, such as pruning the spatial range, compressing the number of channels, or performing temporal stitching.

[0030] Lightweight execution models are used to perform inference processing, mapping input data to interpretable results. These models consist of a resource-controlled network structure, including a weight matrix, model layer topology, data flow paths, and a set of enabled modules. Inference processing propagates the input data hierarchically through the network, generating internal representations through convolutions, attention transformations, feature aggregation, or probability mapping, ultimately forming decision, prediction, or action parameters at the output. The raw inference output may include classification probabilities, behavior control vectors, risk scores, or entity detection box coordinates. The target task processing result is generated from the output through necessary filtering, concatenation, or format transformation, supporting upstream decision-making actions or serving as input for the next pipeline step.

[0031] In one approach, input data is directly read from image frames by the camera driver, and after scaling and normalization, it is fed into a lightweight execution model, outputting environment understanding labels or trajectory prediction results. In another approach, input data is received via a communication interface after sensor fusion of time series data, and then truncated or zero-padding is applied before entering the inference network to obtain the action control distribution. Alternatively, images can be combined with structured vectors for hybrid inference via a joint encoding module, outputting data results with greater decision-making significance.

[0032] This embodiment acquires target task input data in real time and completes inference within a resource-controlled model, enabling the inference output to closely follow changes in the environment and task, avoiding latency accumulation, thereby improving responsiveness and reducing the waste of ineffective computing resources.

[0033] S50, determine the task execution accuracy value based on the target task processing result, and monitor the computing power consumption and energy consumption value of the lightweight execution model during the execution inference process, and generate running performance data based on the task execution accuracy value, the computing power consumption value and the energy consumption value; In this embodiment, the task execution accuracy value determined based on the target task processing result is used to extract a quantitative indicator representing the completion effect from the inference output. The target task processing result may include classification output, numerical prediction, control deviation information, or confidence intervals. The task execution accuracy value can be formed by extracting key result dimensions and mapping them to an accuracy scale. The scale interval can be set according to task requirements and supports the differential expression of different task outputs. To adapt to different application forms, the process of forming the task execution accuracy value may include extracting the main output value, comparing external reference signals, or calculating the internal consistency between output sequences to measure execution performance.

[0034] Monitoring the computational load of the lightweight execution model during inference processing reflects the utilization of computing resources. This can be obtained by sampling processor load, memory usage, or the workload of graphics acceleration units. The computational load value formation process includes reading resource status at intervals, filtering inference-related metrics, and weighted fusion to form a single metric. The sampling time point can be synchronized with the inference process to ensure that the collected resource load information accurately corresponds to the inference execution stage.

[0035] Monitoring the energy consumption of the lightweight execution model during inference processing reflects the energy cost of operation and can be obtained from the power feedback interface, battery status logs, or power consumption sensing modules. Energy consumption values ​​can be derived from instantaneous energy monitoring or inferred from the relationship between battery level decrease and execution duration, thus representing the energy expenditure required to complete the inference process per unit time.

[0036] Performance data is generated based on task execution accuracy, computing power consumption, and energy consumption to combine information from these three dimensions into a structured representation. The generation process may include data alignment, format organization, and vectorization encapsulation, enabling the simultaneous presentation of task execution performance, resource load, and energy consumption costs for subsequent use. Performance data can retain timestamps, task identification identifiers, or execution batch indexes to support analysis needs that accumulate over time.

[0037] In one approach, the task execution accuracy value is calculated by taking the maximum confidence probability of the classification output and mapping it to a percentage. The computing power utilization value is calculated by sampling the processor utilization rate and acceleration unit utilization rate, adding them proportionally, and then averaging them over the number of samples. The energy consumption value is calculated by querying the difference between the current battery level and the battery level at the start of inference, along with the execution time, to form the energy cost per unit time. These three metrics are then organized into parallel data items corresponding to timestamps.

[0038] In another approach, the task execution accuracy value is generated by the difference between the predicted trajectory and the observed trajectory, the computing power occupancy value is formed by reading the kernel running statistics table and filtering the records during inference, and the energy consumption value is formed by continuous sampling and integration by the power feedback interface. The three indicators are then concatenated into a vector form for subsequent processing.

[0039] This embodiment generates task execution accuracy, computing power usage, and energy consumption values ​​in real time and combines them into operational performance data, so that execution effect, resource consumption, and energy consumption performance are presented simultaneously. This supports balancing performance and consumption under limited computing resources and provides a basis for the next stage of operational decision-making.

[0040] S60, compare the operating performance data with a preset performance benchmark threshold, and generate an operating effect analysis level based on the comparison result; In this embodiment, the comparison between the running performance data and the performance benchmark threshold is used to determine whether the execution performance meets the preset quality, resource and energy consumption requirements.

[0041] The performance data includes three indicators: task execution accuracy, computing power usage, and energy consumption. These three indicators reflect the quality of inference results, hardware resource usage, and energy consumption, respectively, and can measure the current execution performance from different perspectives.

[0042] The performance benchmark thresholds include three boundary standards: the lower limit of the accuracy benchmark, the upper limit of the computing power benchmark, and the upper limit of the energy consumption benchmark.

[0043] The lower limit of the accuracy benchmark is used to limit the minimum requirements for task execution quality, ensuring that performance deviations do not accumulate to an unacceptable range; the upper limit of the computing power benchmark is used to limit the proportion of hardware resources used, avoiding excessive usage that could lead to task blocking or excessive system load; the upper limit of the energy consumption benchmark is used to control the level of energy consumption to adapt to mobile computing scenarios or power-sensitive scenarios.

[0044] The comparison process compares the three performance metrics in the operational data with their corresponding values ​​in the performance benchmark thresholds to determine whether the metrics meet or fail to meet the standards. To ensure reliable judgment, a metric is recorded as meeting the standards when the accuracy value is greater than or equal to the lower limit of the benchmark, while a metric is recorded as meeting the standards when the computing power and energy consumption are less than or equal to the upper limit of the benchmark, thus generating status results for each of the three metrics.

[0045] The performance analysis level is generated by combining three target states, used to abstractly express the overall performance differentiation. Meeting all three states simultaneously corresponds to stable performance and is marked as Level 1; meeting accuracy while exceeding resource limits corresponds to slight performance deviation and is marked as Level 2; failing to meet accuracy or exceeding both computing power and energy consumption limits corresponds to unbalanced performance and is marked as Level 3. The performance analysis level serves as a reference for subsequent processing, enabling differentiated responses to be triggered for different performance levels.

[0046] In one implementation, the performance data is split into three independent values. Then, three reference values ​​from the performance benchmark threshold are read, and logical judgments are used to form three Boolean states. These states correspond to accuracy, computing power, and energy consumption, respectively. The three states are then read and combined to generate a level number. For example, all true values ​​generate level one, true accuracy with any resource-related fields present generates level two, and false accuracy generates level three. The level is represented by an integer encoding or a text label.

[0047] In another implementation, the performance data is organized into a vector form, the performance benchmark threshold is organized into a threshold vector, a state structure is formed by dimensional conditional judgment, and then the overall state structure is mapped to the level label table to obtain the level classification result.

[0048] This embodiment generates a level by judging execution performance and resource consumption through thresholds, enabling the execution quality and cost status to be quickly abstracted and identified, achieving real-time classification of the running status, and providing a clear basis for subsequent actions.

[0049] S70, in response to the operational effect analysis level, perform a correction operation on the lightweight strategy parameter set or a weight update operation on the multi-objective decision model, and apply the corrected parameters or updated model to the lightweight execution model generation process at the next moment to optimize the processing of subsequent target tasks.

[0050] In this embodiment, the performance analysis level is used to distinguish whether the current execution performance needs to be maintained, adjusted, or updated. The lightweight strategy parameter set is used to control the actual parameter configuration of the model during quantization, pruning, and inference scheduling. The multi-objective decision model is used to generate the lightweight strategy parameter set based on the feature input, and its internal weights are used to express the importance ratio between different objectives.

[0051] The correction operation is used to adjust the lightweight strategy parameter set incrementally when parameter performance deviates from expectations but does not reach structural failure. The adjustment targets include numerical parameters in the lightweight strategy parameter set that express model size, processing frequency, or submodule start-up and shutdown strategies. Small adjustments are made to achieve deviation convergence without changing the model architecture.

[0052] The weight update operation is used to update the weights within the multi-objective decision-making model when the performance analysis level indicates that the current control strategy is no longer suitable. This update changes the inherent bias trend of the lightweight strategy parameter set generated by the model, so that the lightweight strategy parameter set generated in the next round is realigned with the trade-off between the current task's accuracy, computing power, and energy consumption.

[0053] The revised lightweight strategy parameter set or the updated multi-objective decision model is used to generate a new lightweight execution model in the next time step, ensuring that the next round of inference behavior depends on the adjusted control signals, so that parameter updates and actual operation form a closed-loop evolution process.

[0054] In one implementation, when the performance analysis level reflects a slight deviation between execution accuracy and resource consumption, the original multi-objective decision-making model can be retained, and the lightweight strategy parameters used to control model compression and inference scheduling can be slightly adjusted at fixed steps, and then saved as the corrected lightweight strategy parameter set.

[0055] In another implementation, when the performance analysis level indicates that the overall execution quality is insufficient or resources are significantly exceeded, the expression structure of the lightweight strategy parameter set can be retained, the weights used to represent the priority of the objectives within the multi-objective decision-making model can be adjusted, and a new lightweight strategy parameter set can be generated using the new weights.

[0056] Furthermore, by combining the two types of adjustments, a small parameter correction can be performed after the weight update, enabling the strategy to form a gradual convergence mechanism.

[0057] The revised lightweight strategy parameter set or the multi-objective decision model with updated weights is used to drive the generation of the lightweight execution model in the next time step, so that the new parameter configuration truly reflects the dynamic control decision.

[0058] This embodiment enables the control parameters to dynamically adapt through differentiated adjustment operations triggered by different levels. When the deviation is slight, it can be quickly fine-tuned, and when the deviation is severe, the decision source can be updated, so that the execution state of the next round can continue to evolve with the running feedback.

[0059] In one embodiment, step S10 includes: S101, collect environmental sensing data through sensors, and use a feature extraction network to extract the number of scene targets, target movement speed, ambient light intensity and obstacle density from the environmental sensing data; S102, according to the preset scene complexity quantification module, the number of scene targets, the target movement speed, the ambient light intensity and the obstacle density are weighted and summed to generate scene complexity features; S103, Obtain task scheduling data through the task scheduling module, parse the task scheduling data to extract task priority features and precision threshold features, and generate task attribute features based on the task priority features and precision threshold features; S104, collect equipment resource monitoring data through the equipment resource monitoring module, and extract central processing unit occupancy rate, memory remaining amount, battery power and computing power utilization rate from the equipment resource monitoring data; S105, Based on the preset resource redundancy determination module, the central processing unit occupancy rate, the remaining memory, the battery power and the computing power utilization rate are weighted and processed to generate resource redundancy features.

[0060] In this embodiment, the multi-dimensional sensing data collection does not involve a single-source data stream, but rather a multi-dimensional dataset including at least environmental sensing data, task scheduling data, and device resource monitoring data. Environmental sensing data is collected by sensors installed on the embodied intelligent carrier, such as visual sensors, depth sensors, infrared sensors, or light sensors. These sensors output raw observations at a preset frequency, accompanied by time stamps for subsequent alignment with task scheduling data and device resource monitoring data. Task scheduling data is generated or forwarded by the task scheduling module and can originate from upper-layer business systems, task queues, or scene scheduling centers. It describes the category, urgency, and business objectives of the currently pending task. Device resource monitoring data is extracted by the device resource monitoring module from the operating system interface or hardware monitoring registers, reflecting the current operating status of the processor, memory, power supply, and computing units. By organizing these three types of data under a unified time base, the multi-dimensional sensing data forms an input set covering the three axes of environment, task, and resources, providing a foundation for subsequent extraction of numerical features.

[0061] Feature extraction processing applies environmental sensing data through a feature extraction network. This network can employ multi-layer convolutional structures, temporal convolutional structures, or other feedforward network structures, receiving images, depth maps, or other environmental observation data frame by frame or time slice. The network aggregates pixels or point clouds in the spatial dimension and observes changes between consecutive frames in the temporal dimension, transforming low-level perception results into statistical quantities such as the number of scene targets, target motion speed, ambient light intensity, and obstacle density. The number of scene targets can be obtained from the number of target instances output by the target detection branch; target motion speed can be derived from the ratio of the change in target position to the time interval in adjacent time slices; ambient light intensity can be uniformly quantified based on the average image brightness or the light sensor response value; and obstacle density can be expressed by statistically representing the proportion of obstacle cells in the spatial grid or by the proportion of point clouds within a distance threshold in the depth map. Through this type of network processing, complex environmental observations are compressed into a small number of scene indicators that can be directly used in numerical calculations.

[0062] The scene complexity quantification module receives four types of indicators: the number of scene targets, target movement speed, ambient light intensity, and obstacle density. It then generates a scene complexity feature through weighted summation. In the weighted summation process, a weight coefficient is assigned to each type of indicator. These weights can be pre-set through offline parameter tuning, historical operational statistics, or empirical rules, or they can be issued by the upper-level management system based on business strategies. During the calculation, each indicator is first mapped to a unified numerical range according to a unified or normalized dimension rule, then multiplied by its corresponding weight and summed to obtain a single scene complexity feature, used to express the overall degree of scene perception load and decision-making difficulty. The coefficient values ​​can be adjusted according to different application scenarios. For example, in scenes with many dynamic obstacles, the weights of target movement speed and obstacle density can be increased; in scenes with large lighting fluctuations, the weight of ambient light intensity can be increased.

[0063] The task scheduling module receives task scheduling data from the business orchestration system or task queue, which describes the type and priority of tasks to be executed. Task scheduling data may include fields such as task type markers, business priority tags, service level requirements, latency limits, and accuracy level configurations. The system extracts task priority features and accuracy threshold features by parsing the task scheduling data. Task priority features can be represented using integer levels, tag codes, or continuous weight values ​​to distinguish the scheduling order and resource allocation between different tasks. Accuracy threshold features represent the minimum acceptable accuracy level for a task in the identification, decision-making, or control stages. These can be preset based on historical business experience or industry standards; for example, a higher threshold can be set for life safety-related tasks, while a relatively lenient threshold can be set for low-risk inspection tasks. Task attribute features, composed of task priority features and accuracy threshold features, can be formed into a single representation through vector concatenation, weight fusion, or other combinations. This representation is used to determine the trade-off between accuracy and resource objectives during subsequent multi-objective optimization.

[0064] The device resource monitoring module collects device resource monitoring data periodically or through event-triggered methods. CPU utilization can be obtained through counters provided by the operating system, such as collecting the percentage of user-mode and kernel-mode runtime per unit time to form a normalized utilization metric. Remaining memory can be obtained by querying the remaining capacity of the memory allocator or the number of unused page frames, reflecting the currently available memory space. Battery power can be obtained through the power management chip or the operating system's power management interface, expressed as a percentage, voltage, or remaining capacity. Computing power utilization measures the usage of dedicated acceleration units such as graphics processing units, tensor processing units, or neural network accelerators, and can be obtained through the underlying driver interface, such as the current execution queue fill rate, execution unit activity ratio, or power consumption mapping coefficient. By simultaneously collecting CPU utilization, remaining memory, battery power, and computing power utilization, the device resource monitoring data covers the operating status of four dimensions: processor, memory, power supply, and acceleration units, providing real-time basis for resource constraint judgment.

[0065] After obtaining central processing unit (CPU) utilization, remaining memory, battery power, and computing power utilization, the resource redundancy determination module generates resource redundancy features through weighted processing. Weighted processing can be achieved by first unifying the direction of each resource indicator, and then combining them into a single continuous value using weights and offset parameters. For example, CPU utilization and computing power utilization can be mapped inversely, with lower utilization corresponding to higher redundancy scores; while remaining memory and battery power can be mapped forward, with higher remaining space or battery power corresponding to higher redundancy scores. Subsequently, each score is multiplied by a configured weight coefficient and summed to obtain the resource redundancy feature. The weight coefficients can be set according to the device form factor; for example, increasing the weight of battery power on mobile devices and increasing the weight of computing power utilization on edge nodes. After the resource redundancy feature is generated, downstream processes can use a single value to determine whether the current device has the space to support high-precision model operation or high-frequency inference.

[0066] Through the above-mentioned sequential processing, the multidimensional perception data is decomposed and recombined into three types of numerical representations: scene complexity features, task attribute features, and resource redundancy features. These represent the current operating conditions from the perspectives of environmental structure and dynamics, task business requirements, and equipment resource status, respectively, providing directly usable input for subsequent multi-objective optimization and model reconstruction.

[0067] This embodiment transforms environmental observation, task scheduling information, and equipment resource status into three complementary representation chains—scene complexity features, task attribute features, and resource redundancy features—through joint acquisition and hierarchical extraction of multi-dimensional sensing data. This enables lightweight decision-making to be based on quantifiable scene complexity, clear task accuracy requirements, and real-time resource redundancy levels. Consequently, it provides fine and adjustable input conditions for subsequent multi-objective optimization and lightweight execution model construction, thereby improving scene adaptability and resource matching rationality.

[0068] In one embodiment, step S20 above includes: S201, Define a multi-objective optimization objective, which includes an accuracy objective, a computing power consumption objective, and an energy consumption objective; S202, Define a lightweight strategy parameter set, which includes quantization bit width parameters, pruning ratio parameters, inference frame rate parameters, and module activation parameters; S203, parse the task attribute features to extract the core precision threshold and task priority, set precision constraints based on the core precision threshold, and generate weight coefficients based on the task priority; S204, Analyze the resource redundancy characteristics to determine the computing power constraint boundary and energy consumption constraint boundary; S205, based on the multi-objective optimization objective, the accuracy constraint, the computing power constraint boundary, and the energy consumption constraint boundary, a multi-objective decision model is constructed, and the scenario complexity feature, the core accuracy threshold, and the resource redundancy feature are used as inputs to the multi-objective decision model; S206, Based on the multi-objective decision model and the weight coefficients, a weighted objective function is formed, and an optimization search is performed on the parameter space of the lightweight strategy parameter set to obtain a solution vector through iterative optimization; S207, the solution vector is decoded into the quantization bit width parameter, the pruning ratio parameter, the inference frame rate parameter, and the module activation parameter to obtain the final lightweight strategy parameter set.

[0069] In this embodiment, scene complexity features, task attribute features, and resource redundancy features have been obtained during the multi-dimensional perception stage. These three types of features are uniformly fed into a multi-objective decision model to derive a lightweight strategy parameter set that balances accuracy, computing power, and energy consumption. The multi-objective optimization objective consists of three parts: accuracy objective, computing power consumption objective, and energy consumption objective. The accuracy objective describes the desired task completion quality in numerical form, such as recognition accuracy and trajectory tracking error upper limit. It constrains the lightweight nature of the model by maintaining a relationship of not being lower than a previously given accuracy threshold. The computing power consumption objective reflects the degree of computing power resource consumption of a single inference process by aggregating indicators such as the processor occupancy ratio and acceleration unit utilization level during inference. The energy consumption objective quantifies the energy consumption of a single inference, and the energy cost of each task execution can be derived based on power sampling results and inference duration. By defining these three objectives simultaneously, subsequent optimization no longer pursues only a single indicator but seeks a compromise solution in a multi-dimensional space.

[0070] In this stage, the lightweight strategy parameter set is explicitly defined as a set including quantization bit width parameters, pruning ratio parameters, inference frame rate parameters, and module activation parameters. The quantization bit width parameter specifies the numerical precision bit width used for weights, feature maps, or intermediate tensors in each layer; for example, choosing different bit ranges affects storage usage and numerical error. The pruning ratio parameter describes the proportion of channels, nodes, or branches retained in each network layer or module during structural pruning; its value range can be configured according to the network structure. The inference frame rate parameter controls the frequency of inference triggered per unit time, which can be achieved by setting time intervals or frame intervals. In dynamic scenarios, reducing the frame rate alleviates real-time computing pressure. The module activation parameter marks the start / stop status of optional functional modules; for example, it controls whether high-precision branches, redundant detection branches, and redundant perception branches participate in this round of inference, expressing different activation levels through binary flags or multi-level strength weights. These four types of parameters together constitute the direct input for subsequent model reconstruction and runtime configuration stages, therefore, explicit values ​​need to be provided in this stage.

[0071] The core precision threshold and task priority contained in the task attribute features are further analyzed. The core precision threshold provides the minimum allowable execution precision value for the task. It can be pre-configured based on business strategies, with higher thresholds for security-sensitive tasks and slightly lower thresholds for tasks with higher fault tolerance. Task priority describes the importance of the task in the entire task queue and can be represented using a level label or continuous weights. When analyzing task attribute features, the original encoding is first mapped to numerical precision thresholds and priority weights. Then, precision constraints are constructed based on the precision thresholds, such as stipulating that the task execution precision value must not be lower than the threshold, or stipulating that the range of values ​​for quantization bit width parameters and pruning ratio parameters must not cause the precision estimate to be lower than the threshold. Simultaneously, task priority is converted into weight coefficients to influence the trade-off between precision and resource objectives in the multi-objective decision model. When the priority is high, the proportion of the precision objective in the weighted objective function is increased; when the priority is low, the weights of computing power consumption and energy consumption objectives can be appropriately increased.

[0072] Resource redundancy characteristics are used to derive computing power and energy consumption constraints. These characteristics are derived from a weighted combination of CPU utilization, remaining memory, battery power, and computing power utilization, and can be considered a comprehensive indicator of the device's currently available resource space. When deriving the constraint boundaries, the system can determine the maximum allowable computing power consumption and maximum energy consumption levels based on the relationship between the resource redundancy characteristics and a preset resource carrying capacity curve. For example, when the resource redundancy characteristics are high, the computing power and energy consumption constraints can be relaxed, allowing for higher-precision but more resource-intensive parameter configurations during optimization; conversely, when the resource redundancy characteristics are low, the computing power and energy consumption constraints are tightened, causing the optimization process to favor parameter combinations with lower resource consumption.

[0073] The multi-objective decision-making model is constructed as a mapping structure that maps scene complexity features, core accuracy thresholds, and resource redundancy features to a lightweight policy parameter set. The model input receives scene complexity features, core accuracy thresholds, and resource redundancy features. Through a set of adjustable parameters or functional relationships, each point in the input space is mapped to a set of candidate configurations in the parameter space. Scene complexity features reflect the difficulty of perception and decision-making in the current environment. When the complexity is high, the model tends to output combinations with higher quantization bit widths, smaller pruning ratios, or higher inference frame rates to maintain task execution accuracy. The core accuracy threshold is applied as a hard constraint in the mapping relationship, limiting the maximum extent of lightweighting. Resource redundancy features limit the resource overhead of the parameter combinations, ensuring they meet computational and energy consumption constraints. Through this input combination, the multi-objective decision-making model can output differentiated parameter configurations under different scene and resource conditions.

[0074] The weighted objective function is determined by a multi-objective decision model and weighting coefficients. The accuracy objective, computing power consumption objective, and energy consumption objective each correspond to a sub-objective expression, and the weighting coefficients determine the proportion of each sub-objective in the overall objective based on task priority and other strategies. When forming the weighted objective function, each sub-objective is multiplied by its corresponding weight and aggregated to obtain a single scalar optimization index. The weighting coefficients can be dynamically adjusted; for example, the weight of the accuracy objective can be increased when power is sufficient and the task is extremely important, while the weight of the energy consumption objective can be increased when power is scarce and the task is relatively less important.

[0075] When iteratively optimizing in the parameter space, the lightweight strategy parameter set is considered as a set of variables to be explored. Each combination of quantization bit width, pruning ratio, inference frame rate, and module activation parameters corresponds to a point in the parameter space. The system can employ gradient-based, contrastive search-based, or evolutionary search-based update mechanisms to explore the parameter space in multiple rounds. In each iteration, an update direction and update step size are generated based on the performance of the current parameter combination on the weighted objective function, and then applied to the parameter combination to gradually approach the minimum or near-minimum value of the weighted objective function. The results obtained through multiple iterations are represented in the form of a solution vector, where each component of the solution vector corresponds to actual parameters such as quantization bit width, pruning ratio, inference frame rate, and module activation state.

[0076] After the solution vector is generated, it needs to be decoded to be converted into practically applicable quantization bit width parameters, pruning ratio parameters, inference frame rate parameters, and module activation parameters. During decoding, each component in the solution vector is converted into a specific value according to a preset mapping relationship. For example, continuous values ​​are mapped to discrete bit width sets, normalization ratios are mapped to legal pruning ratio ranges, dimensionless time factors are mapped to specific frame intervals or trigger intervals, and module-related components are mapped to enabled or disabled states. Through the decoding process, the resulting lightweight strategy parameter set satisfies accuracy constraints, computational power constraints, and energy consumption constraints, while also considering the trade-offs brought about by task priorities, providing directly usable control parameters for subsequent model reconstruction and runtime configuration.

[0077] This embodiment introduces a multi-objective optimization framework that includes accuracy targets, computing power consumption targets, and energy consumption targets. It also maps scene complexity features, task attribute features, and resource redundancy features to quantization bit width parameters, pruning ratio parameters, inference frame rate parameters, and module activation parameters. This forms a lightweight strategy parameter set that is jointly constrained by accuracy constraints, computing power constraints, and energy consumption constraints. This enables adaptive parameter optimization under different task priorities and resource states, allowing the embodied intelligence system to dynamically balance computing power usage and energy consumption while ensuring task execution accuracy.

[0078] In one embodiment, step S30 above includes: S301, obtain the quantization bit width parameter, pruning ratio parameter, inference frame rate parameter and module activation parameter from the lightweight strategy parameter set; S302, Perform quantization processing on the initial model based on the quantization bit width parameter to generate a quantized model; S303, Perform a pruning process on the quantized model based on the pruning ratio parameter and the scene complexity feature to generate a pruned model; S304, Configure inference scheduling for the pruned model based on the inference frame rate parameter, and generate the scheduled model; S305, Based on the module activation parameters, the activation status of the model configuration module after the scheduling configuration is adjusted to generate a lightweight execution model.

[0079] In this embodiment, during the process of reconstructing the initial model and configuring runtime parameters based on the lightweight strategy parameter set, the lightweight strategy parameter set is first parsed to read four types of control variables: quantization bit width parameter, pruning ratio parameter, inference frame rate parameter, and module activation parameter. During parsing, the parameter set is mapped to an internal configuration table that corresponds one-to-one with the initial model structure. For example, parameter entries are created for each layer, each functional branch, and each inference channel, enabling precise location of the necessary adjustments layer-by-layer and module-by-module on the graph structure or operator sequence during subsequent processing. The quantization bit width parameter indicates the data precision of different layer weights and intermediate features; the pruning ratio parameter indicates the retention ratio of each layer channel or node; the inference frame rate parameter indicates the number of times inference can be triggered per unit time; and the module activation parameter indicates the start / stop status of auxiliary branches, redundant branches, and optional functional modules. These four types of parameters are organized into a configuration set that can directly drive the reconstruction process during the parsing phase.

[0080] The quantization process uses the quantization bit width parameter as a control input to convert high-precision numerical representations into a lower bit-width data format while maintaining the initial model topology. Specifically, the distribution range of weights and intermediate features at each layer can be statistically analyzed first. Scaling and offset factors are then generated based on this distribution range, and the original floating-point values ​​are mapped to the target integer range through scaling and truncation. Different layers can use different quantization bit width parameters during this process. For example, a higher bit width can be maintained for precision-sensitive perception layers, while a lower bit width can be used for precision-insensitive post-processing layers, reducing storage space and memory access bandwidth usage. The quantized parameters, along with the operator definitions, are written into a new model description, forming the quantized model and providing a foundation for subsequent pruning and scheduling configurations.

[0081] The pruning process utilizes both pruning ratio parameters and scene complexity features, introducing scene-aware information to guide structural pruning. In implementation, a pruning intensity coefficient is first generated based on scene complexity features to adjust the original pruning ratio parameter; for example, increasing the retention ratio when complexity is high and increasing the pruning ratio when complexity is low. Then, the model is traversed layer by layer after quantization. For each layer, channels or nodes to be retained are selected based on the final pruning ratio. These can be sorted according to signal response strength, channel weight norm, or pre-calculated contribution, generating retention and removal lists. The corresponding channels are then deleted from the network topology, input and output dimensions are updated, and the connection relationships of subsequent layers are updated synchronously. Through this pruning process, redundant structures are reduced without changing the overall functional path, resulting in a pruned model that reduces computational and storage consumption.

[0082] The inference scheduling configuration stage uses the inference frame rate parameter as the control basis, binding the pruned model to the runtime triggering mechanism. Specifically, the inference frame rate parameter can be converted into a time interval or data frame interval, and a timed trigger or event trigger can be set in the execution environment. Inference requests are only sent to the pruned model when the triggering conditions are met. For continuous data stream input scenarios, frame dropping control can be implemented based on timestamps or frame sequences when receiving data, generating inference tasks only for samples that meet the frame rate configuration requirements, thus avoiding frequent inference process initiation when resources are scarce. After the inference scheduling configuration is completed, the pruned model and the scheduling strategy form a unified whole, marked as the configured model, providing a unified entry point for subsequent module activation configurations.

[0083] The module activation configuration process, based on module activation parameters, divides the optional branches and functional modules in the configured model into enabled and disabled sets. In implementation, a switch flag and priority information can be set for each optional branch, mapping the module activation parameters to the specific values ​​of these flags. When constructing the execution graph, execution nodes and resources are allocated only to modules in the enabled set, while modules in the disabled set are removed from the execution graph or kept silent. For modules that need to be dynamically enabled based on scenarios, such as high-precision re-detection branches or redundant security detection branches, conditional activation rules can be set based on scenario complexity characteristics. When specific conditions are met, runtime control switches the module flag from disabled to enabled. After module activation configuration, the configured model is transformed into a lightweight execution model that can truly run under the current scenario and resource conditions. The model structure, data precision, running frequency, and module composition are consistent with the lightweight strategy parameter set.

[0084] This embodiment reads quantization bit width parameters, pruning ratio parameters, inference frame rate parameters, and module activation parameters from the lightweight strategy parameter set, and sequentially performs quantization processing, structure pruning, inference scheduling configuration, and module activation configuration. This transforms the initial model into a lightweight execution model tailored to the current scenario and resource status, achieving coordinated adjustment of data accuracy, structure size, execution frequency, and functional module range. While ensuring task processing capabilities, it significantly reduces computing power and energy consumption, providing a more suitable model form for subsequent task execution.

[0085] In one embodiment, step S40 above includes: S401, determine the data sampling interval based on the inference frame rate parameter in the lightweight strategy parameter set, collect real-time environmental data through the sensor interface according to the data sampling interval, and generate target task input data; S402, Load the lightweight execution model and activate the corresponding functional modules based on the module activation parameters in the lightweight strategy parameter set; S403, input the target task input data into the lightweight execution model; S404, the lightweight execution model is used to perform inference processing on the target task input data, and the original output data is obtained from the output layer of the lightweight execution model; S405, Post-process the original output data based on the task attribute features to generate the target task processing result.

[0086] In this embodiment, the inference frame rate parameter is derived from the aforementioned lightweight strategy parameter set and is used to limit the frequency of inference triggers allowed per unit time. In implementation, the system can convert the inference frame rate parameter into a time-domain sampling interval, for example, by determining the sampling interval length based on a preset time window length and the desired number of inferences; alternatively, it can be based on a data frame sequence, triggering acquisition every certain number of frames, establishing a one-to-one correspondence between the frame interval and the inference frame rate parameter. After the data sampling interval is determined, the sensor interface, during runtime, acts as an external access point, connecting to cameras, depth sensors, LiDAR, torque sensors, or fusion sensing units, and initiating acquisition commands at each sampling moment through a unified calling protocol. The acquired real-time environmental data is organized using timestamps, ordered buffer queues, or batch identifiers, and can be multi-channel image data, point cloud data, time-series signal data, or multimodal combined data. After necessary formatting and normalization processing, it is ready in the buffer as input data for the target task.

[0087] The lightweight execution model requires a loading process before receiving input data for the target task. This loading can be completed during device startup or lazily loaded before receiving the first input data for the target task. In implementation, the system reads the model structure description and weight files matching the current lightweight strategy parameter set from local storage or a remote model repository, maps these contents into memory, and constructs the execution graph and operator scheduling order in the inference engine. Module activation parameters, also derived from the lightweight strategy parameter set, determine the start / stop status of each functional module within the lightweight execution model. Functional modules can include high-precision re-detection branches, redundant safety monitoring branches, optional attention enhancement branches, multi-scale feature fusion branches, or task-specific output headers. During the loading phase, the system parses the module activation parameters, binding each functional module to a switch flag. When constructing the execution graph, execution nodes and runtime resources are allocated only to active functional modules; inactive modules do not participate in inference computation, or only retain their structural definitions without participating in the current inference process.

[0088] Once the target task input data is ready and the lightweight execution model is fully loaded and in a stable active state, the system sends the target task input data into the lightweight execution model via a data channel. The data channel can be a shared memory queue, a message queue, or a direct function parameter. During the data transmission process, the target task input data needs to be dimensionally aligned and its data type converted according to the model's input specifications. For example, image data is organized into a specified batch size, height, width, and number of channels; time-series data is aligned to a uniform time step length; and the numerical type is ensured to match the quantization bit width. After receiving the target task input data, the lightweight execution model executes each network layer and functional module according to the pre-arranged execution graph. The data is passed layer by layer from the low-level feature extraction units to the high-level semantic units, and then a set of intermediate results corresponding to the current task is generated through the task-related output header.

[0089] During inference, the output layer of the lightweight execution model transforms internal features into interpretable results. The output layer can take the form of a fully connected structure, a convolutional output head, a sequence decoder, or a graph-structured readout unit, outputting raw output data such as category probability distributions, regression value sequences, segmentation masks, pose parameters, or trajectory point sets, depending on the task type. After the lightweight execution model completes one forward propagation, the system reads this set of raw output data from the output layer and appends timestamps, task identifiers, and scene identifiers corresponding to the target task input data, ensuring that subsequent processing can correctly correlate these data in both temporal and task dimensions.

[0090] Task attribute features are derived from the parsing of task scheduling data in the preceding stages, including information such as task priority, precision threshold, latency tolerance, fault tolerance range, and result expression format. The post-processing stage takes the raw output data and task attribute features as input, performing result organization and format conversion. In implementation, it can filter the scores or confidence values ​​in the raw output data based on the precision threshold, eliminating low-confidence candidate results; it can adjust the display order or retention range of multi-target outputs based on task priority, placing high-priority results first; it can choose whether to perform more time-consuming fine-grained resampling or interpolation processing based on latency tolerance; and it can convert numerical results into structured records, tag sets, control commands, or text displayable on a human-machine interface based on the result expression format. After this series of processing steps, the raw output data is organized into target task processing results that meet the constraints of the current task, stored in structured or sequential form, and output to subsequent control units or business units.

[0091] This embodiment directly determines the data sampling interval using the inference frame rate parameter and collects real-time environmental data at this interval on the sensor interface, ensuring that the target task input data and the running rhythm of the lightweight execution model are consistent, thus preventing invalid data from entering the inference channel. By combining the module activation parameter with the model loading stage to complete the start and stop configuration of functional modules, only branches and output heads related to the current task are retained during the inference process, reducing redundant calculations. After the inference is completed, the original output data is extracted from the output layer and targeted post-processing is performed in combination with task attribute characteristics, mapping the same model output to appropriate result forms under different task requirements. This achieves coordinated control of data acquisition frequency, model structure usage, and result processing methods on the same link, thereby still outputting target task processing results that meet business constraints under limited computing power and energy consumption conditions.

[0092] In one embodiment, step S50 above includes: S501, parse the target task processing result to obtain an analysis index containing confidence score or action feedback error, map the analysis index to a preset precision quantization range, and obtain a task execution precision value that characterizes the current task completion quality. S502, within the time window during which the lightweight execution model performs the inference process, the CPU load rate, GPU computing power utilization rate, and memory bandwidth utilization rate are sampled through the system monitoring interface, and a weighted average is performed on the CPU load rate, GPU computing power utilization rate, and memory bandwidth utilization rate to obtain the computing power utilization value. S503, monitor the instantaneous power or battery power decay rate recorded by the power management module of the monitoring device, determine the power consumption level per unit time based on the duration of the inference processing and the instantaneous power or battery power decay rate, and obtain the energy consumption value. S504, the task execution accuracy value, computing power occupancy value and energy consumption value are timestamped and vectorized to generate runtime performance data.

[0093] In this embodiment, during the process of processing the target task results and generating runtime performance data, it is first necessary to extract analytical indicators that reflect the quality of task completion from the result data. The target task processing results generally come from the output layer of the lightweight execution model. For classification tasks, this may include category confidence scores; for detection, segmentation, or trajectory prediction tasks, it may include values ​​such as bounding box positions and keypoint coordinates; and for control tasks, it may include deviation information between control commands and execution feedback. During the parsing phase, confidence scores or action feedback errors are extracted from the result data structure through predefined mapping relationships. These values ​​are then normalized into a unified set of analytical indicators according to task type. For example, the highest category confidence score is selected for classification results, and the end-effector pose error or trajectory offset is selected for control results, forming the basic quantities used for accuracy measurement.

[0094] To ensure the comparability of analytical metrics across different tasks and scenarios, a precision quantization interval is introduced to map the analytical metrics to intervals. The precision quantization interval can be pre-defined as a closed interval or a segmented interval based on business needs. For example, the confidence score interval can be mapped to a precision scale from zero to one, and the action feedback error interval can be mapped to a reverse scale from zero to one. In implementation, the system loads the corresponding upper and lower limits of the quantization interval and mapping rules for each task type during configuration loading. During runtime, the system selects the appropriate rule based on the current task type and converts the original analytical metrics into values ​​on a uniform scale through linear interpolation, piecewise linear, or nonlinear mapping methods. This yields the task execution precision value, using a single numerical value to characterize the quality of the current task completion.

[0095] During task execution, it is necessary to synchronously monitor the computational resource usage of the lightweight execution model. To this end, a monitoring time window is set, aligned with the inference process. For example, the start time of an inference operation can be used as the starting point, and the end time as the ending point, or a sliding time window can be used to cover multiple consecutive inferences. The system monitoring interface interfaces with the operating system or runtime environment, obtaining CPU load rate from the CPU performance counter, GPU utilization rate from the GPU driver layer or inference engine, and memory bandwidth usage from the memory controller or performance monitoring unit. The monitoring interface reads these indicator values ​​multiple times within the time window according to a preset sampling period, forming a resource usage sequence that changes over time. To integrate the three types of indicators into a single computational resource usage value, the system configures weighting coefficients for CPU load rate, GPU utilization rate, and memory bandwidth usage rate. These weights can be adjusted according to different hardware platforms or application scenarios. At the end of the window, the collected sequence is weighted according to the weighting coefficient combination to balance short-term spikes with long-term averages, thus obtaining the comprehensive computational resource usage value within the current time window.

[0096] Energy consumption monitoring is accomplished using the device's power management module. This module typically provides instantaneous power data or battery percentage changes over time. For devices supporting instantaneous power output, voltage and current are periodically read within the inference time window. An instantaneous power curve is generated via a power sensing path, and then time-weighted by the inference duration to derive the average power consumption per unit time. For devices where battery level is the primary monitoring indicator, the battery percentage is recorded at the beginning and end of the inference time window. The battery decay rate is determined by combining the duration and battery capacity parameters, and then mapped to energy consumption intensity. Both methods allow for the output of energy consumption values ​​in a uniform numerical format, reflecting the energy consumption during the inference process, regardless of hardware capabilities.

[0097] Since the accuracy, computational power, and energy consumption values ​​of task execution are not perfectly synchronized in terms of generation time, timestamp alignment is required. In implementation, a unique timestamp is assigned to each inference operation, and the accuracy result, computational power window, and energy consumption window corresponding to that inference are bound to the same timestamp. When the computational power monitoring window and energy consumption monitoring window cover multiple inferences, a local index can be assigned to each inference within the window. The window-level values ​​are then distributed to the corresponding inference operations through interpolation or in-window averaging, ensuring that the result of each target task processing is associated with a unique triplet of accuracy, computational power, and energy consumption. After timestamp alignment, the system organizes the task execution accuracy, computational power, and energy consumption values ​​into a fixed-length numerical vector according to a predetermined order. Timestamps or task identifiers can be appended to the vector for subsequent querying and archiving. Through this vectorized encapsulation, the three types of indicators from multiple task samples can form a matrix or time series structure, ultimately generating runtime performance data, providing a unified input for any subsequent data-driven judgments, recording, or optimizations.

[0098] This embodiment extracts confidence scores or action feedback errors from the target task processing results and maps them to a precision quantification range to obtain a task execution precision value, thus providing a unified measurement scale for task completion quality across different task types. Within the inference time window, it collects and weights CPU load rate, GPU computing power utilization, and memory bandwidth usage through the system monitoring interface, compressing scattered resource usage information into a single computing power usage value. It then determines the power consumption level per unit time by combining instantaneous power data or battery power decay rate with inference duration, obtaining an energy consumption value. Finally, through timestamp alignment and vectorized encapsulation, the task execution precision value, computing power usage value, and energy consumption value are uniformly recorded as runtime performance data, achieving synchronous quantification of task quality and resource consumption within the same data structure. This ensures that subsequent analysis and adjustment of the runtime status are based on a complete, comparable, and structured foundation.

[0099] In one embodiment, step S60 above includes: S601, parse the running performance data to extract the task execution accuracy value, computing power occupancy value and energy consumption value respectively, and obtain the preset performance benchmark threshold, the performance benchmark threshold including the accuracy benchmark lower limit, computing power benchmark upper limit and energy consumption benchmark upper limit; S602, determine whether the task execution accuracy value is greater than or equal to the lower limit of the accuracy benchmark, and determine whether the computing power occupancy value and the energy consumption value are less than or equal to the upper limit of the computing power benchmark and the upper limit of the energy consumption benchmark, respectively, so as to determine the compliance status of the task execution accuracy value, the computing power occupancy value and the energy consumption value respectively; S603, if the target status of the task execution accuracy value, the computing power occupancy value and the energy consumption value are all yes, then generate the first level of the indication strategy maintenance as the operation effect analysis level; S604, if the target status of the task execution accuracy value is yes, but the target status of either the computing power occupancy value or the energy consumption value is no, then generate a second level of indicator parameter fine-tuning as the running effect analysis level. S605, if the target status of the task execution accuracy value is not met, or the target status of the computing power occupancy value and the energy consumption value are both not met, then the third level of the indicator model reconstruction is generated as the operation effect analysis level.

[0100] In this embodiment, the process of comparing runtime performance data with performance benchmark thresholds and generating runtime performance analysis levels first requires parsing the structure of the runtime performance data. Runtime performance data can be stored according to time slices or task rounds. Each record contains at least three indicators: task execution accuracy, computing power usage, and energy consumption. Additional fields such as task identifier, timestamp, and device identifier may also be included. The parsing process accurately extracts these three types of values ​​from the records using fixed field positions or key-value mapping, ensuring that subsequent processing always revolves around these three key indicators and avoids introducing fields unrelated to the runtime status.

[0101] Performance baseline thresholds are pre-set during system configuration or model deployment to characterize acceptable performance boundaries. The accuracy baseline lower limit constrains the minimum required accuracy value for task execution and can be determined based on business rules; for example, critical tasks can be set to a precision baseline close to the upper limit, while general tasks can be set to a relatively lower precision baseline. The computing power baseline upper limit restricts the maximum range of computing power usage and can be selected based on the device's rated load, heat dissipation conditions, and the number of concurrent tasks. The energy consumption baseline upper limit restricts energy consumption and can be configured based on battery capacity, recharging cycle, or power supply capability. These three baseline thresholds can be configured separately for different device types and task types in tabular form. At runtime, the corresponding configuration set is selected based on the current environment, ensuring both uniformity and adaptability to different scenarios.

[0102] After obtaining the task execution accuracy, computing power utilization, energy consumption, and performance benchmark threshold, each item needs to be compared to form discrete achievement status markers. The task execution accuracy is compared to the lower limit of the accuracy benchmark using a greater than or equal to relationship. If the task execution accuracy is not lower than the lower limit, the accuracy achievement status is marked as positive; otherwise, it is marked as negative. The computing power utilization is compared to the upper limit of the computing power benchmark using a less than or equal to relationship. If the computing power utilization does not exceed the upper limit, the computing power achievement status is positive; if it exceeds the upper limit, it is marked as negative. Similarly, the energy consumption is compared to the upper limit of the energy consumption benchmark using a less than or equal to relationship. If the energy consumption is not higher than the upper limit, the energy consumption achievement status is positive; if it exceeds, it is marked as negative. Through this rule, the continuous numerical space can be compressed into three binary states, corresponding to whether the task accuracy is achieved, whether computing power resources are controlled within the target range, and whether energy consumption is at an acceptable level.

[0103] After obtaining the three target states, different state combinations need to be mapped to a finite number of operational performance analysis levels to provide clear adjustment signals for subsequent control processes. When all three target states are positive, it indicates that the current task execution accuracy meets the requirements, and the computing power and energy consumption values ​​have not exceeded the expected upper limits, indicating that the system is operating stably and efficiently. At this point, the first level of the indicator strategy is generated, indicating that the current lightweight strategy does not need to be modified in the existing scenario and can continue to be used.

[0104] When the task execution accuracy is positively assessed, but either the computing power consumption or energy consumption is negatively assessed, it indicates that the task result quality still meets the requirements, but resource usage has deviated from the target range. Excessive computing power consumption may indicate insufficient lightweighting, while excessive energy consumption may indicate increased power pressure or reduced battery life. In this case, a second level of indicator parameter fine-tuning is generated, prompting subsequent stages to adjust parameters such as quantization bit width, pruning ratio, or inference frame rate to alleviate resource pressure while maintaining task execution accuracy within acceptable limits. This level prioritizes accuracy constraints over resource constraints without ignoring resource-side anomalies.

[0105] When the target accuracy of a task is negative, or when both the target computing power and energy consumption are negative, it indicates that the system is in an unbalanced state. Insufficient accuracy indicates a decline in the reliability of task results; even if resource usage is within a reasonable range, business needs cannot be guaranteed. Exceeding both computing power and energy consumption indicates severe overload or high energy consumption on the resource side; even if accuracy is temporarily met, continued operation is difficult. Therefore, a third level indicating model reconstruction is generated when any condition is triggered, serving as a trigger signal for subsequent, more significant adjustments. This guides the decision-making side to reselect lightweight strategy parameters or even adjust the model structure. This rule divides all possible numerical combinations into three mutually exclusive levels, giving different severity levels of operational states a clear location in the discrete level space.

[0106] In batch tasks or continuous time windows, the above comparison and ranking generation processes can be repeated for each time slice or each task round. The system can parse operational performance data in real time as the data stream enters, load the corresponding performance benchmark thresholds, generate three compliance states and operational effect analysis levels, and write the ranking results to the status record. Alternatively, historical operational performance data can be centrally processed in an offline analysis environment to obtain long-term ranking distributions, providing a basis for subsequent parameter configuration. Through unified comparison rules, discrete state marking, and ranking mapping logic, operational performance data is no longer just raw numerical values, but is transformed into ranking information that can directly drive decision-making and adjustment.

[0107] This embodiment analyzes the task execution accuracy, computing power consumption, and energy consumption values ​​in the operational performance data, and compares them item by item with the lower limit of accuracy benchmark, the upper limit of computing power benchmark, and the upper limit of energy consumption benchmark. It transforms continuous performance indicators into accuracy compliance status, computing power compliance status, and energy consumption compliance status. Then, based on the combination of the three compliance statuses, it generates an operational effect analysis level indicating strategy maintenance, parameter fine-tuning, or model reconstruction. This allows quality constraints and resource constraints to be considered simultaneously under unified rules, realizing the hierarchical judgment of operational status. This provides clear and easy-to-execute control signals for subsequent lightweight strategy adjustments and model updates.

[0108] In one embodiment, an adaptive lightweight task execution and optimization apparatus is provided, which corresponds one-to-one with the adaptive lightweight task execution and optimization method described in the above embodiments. (Refer to...) Figure 3 , Figure 3This is a schematic diagram of the functional modules of a preferred embodiment of the adaptive lightweight task execution and optimization device of the present invention. The modules include a multi-dimensional perception feature extraction module 10, a multi-objective decision generation module 20, a lightweight model construction module 30, a task reasoning and execution module 40, an execution performance monitoring module 50, a running effect evaluation module 60, and a strategy adaptive optimization module 70. Detailed descriptions of each functional module are as follows: The multidimensional perception feature extraction module 10 is used to collect multidimensional perception data and extract features from the multidimensional perception data to generate scene complexity features, task attribute features and resource redundancy features. The multi-objective decision generation module 20 is used to input the scene complexity features, the task attribute features and the resource redundancy features into the multi-objective decision model, and generate a lightweight strategy parameter set through multi-objective optimization processing. The lightweight model construction module 30 is used to perform model reconstruction and runtime parameter configuration on the initial model according to the lightweight strategy parameter set, and generate a lightweight execution model. The task reasoning and execution module 40 is used to acquire target task input data and use the lightweight execution model to perform reasoning processing on the target task input data to output the target task processing result. The execution performance monitoring module 50 is used to determine the task execution accuracy value based on the target task processing result, and monitor the computing power consumption and energy consumption value of the lightweight execution model during the execution inference process, and generate running performance data based on the task execution accuracy value, the computing power consumption value and the energy consumption value. The operation performance evaluation module 60 is used to compare the operation performance data with a preset performance benchmark threshold and generate an operation performance analysis level based on the comparison result. The strategy adaptive optimization module 70 is used to perform a correction operation on the lightweight strategy parameter set or a weight update operation on the multi-objective decision model in response to the running effect analysis level, and apply the corrected parameters or updated model to the lightweight execution model generation process at the next time step to optimize the processing of subsequent target tasks.

[0109] Specific limitations regarding the adaptive lightweight task execution and optimization device can be found in the aforementioned limitations of the adaptive lightweight task execution and optimization method, and will not be repeated here. Each module in the aforementioned adaptive lightweight task execution and optimization device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0110] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides deterministic and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of an adaptive lightweight task execution and optimization method on the server side.

[0111] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements client-side functions or steps of an adaptive lightweight task execution and optimization method.

[0112] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Collect multidimensional sensing data and extract features from the multidimensional sensing data to generate scene complexity features, task attribute features and resource redundancy features. The scenario complexity features, the task attribute features, and the resource redundancy features are input into a multi-objective decision model, and a lightweight strategy parameter set is generated through multi-objective optimization processing. Based on the lightweight strategy parameter set, the initial model is reconstructed and the running parameters are configured to generate a lightweight execution model. Obtain the target task input data, and use the lightweight execution model to perform inference processing on the target task input data to output the target task processing result; The task execution accuracy value is determined based on the target task processing result, and the computing power consumption and energy consumption value of the lightweight execution model during the execution inference process are monitored. Based on the task execution accuracy value, the computing power consumption value and the energy consumption value, the running performance data is generated. The operational performance data is compared with a preset performance benchmark threshold, and an operational performance analysis level is generated based on the comparison results. In response to the operational performance analysis level, a correction operation is performed on the lightweight strategy parameter set or a weight update operation is performed on the multi-objective decision model. The corrected parameters or the updated model are then applied to the lightweight execution model generation process at the next time step to optimize the processing of subsequent target tasks.

[0113] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, and a computer program is stored thereon, which, when executed by a processor, performs the following steps: Collect multidimensional sensing data and extract features from the multidimensional sensing data to generate scene complexity features, task attribute features and resource redundancy features. The scenario complexity features, the task attribute features, and the resource redundancy features are input into a multi-objective decision model, and a lightweight strategy parameter set is generated through multi-objective optimization processing. Based on the lightweight strategy parameter set, the initial model is reconstructed and the running parameters are configured to generate a lightweight execution model. Obtain the target task input data, and use the lightweight execution model to perform inference processing on the target task input data to output the target task processing result; The task execution accuracy value is determined based on the target task processing result, and the computing power consumption and energy consumption value of the lightweight execution model during the execution inference process are monitored. Based on the task execution accuracy value, the computing power consumption value and the energy consumption value, the running performance data is generated. The operational performance data is compared with a preset performance benchmark threshold, and an operational performance analysis level is generated based on the comparison results. In response to the operational performance analysis level, a correction operation is performed on the lightweight strategy parameter set or a weight update operation is performed on the multi-objective decision model. The corrected parameters or the updated model are then applied to the lightweight execution model generation process at the next time step to optimize the processing of subsequent target tasks.

[0114] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0115] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0116] It should be noted that if any AI models, software tools, or components not belonging to this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use. The above-described embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

[0117] The user personal information involved in this application embodiment is all authorized (knowing and consenting) by the relevant parties or fully authorized by all parties, and the executing entity can obtain it through various open, legal and compliant means. The collection, storage, use, processing, transmission, provision and disclosure of the information, data and signals involved all comply with the relevant laws and regulations of the relevant countries and regions, and do not violate public order and good morals.

Claims

1. An adaptive lightweight task execution and optimization method, characterized in that, Includes the following steps: Collect multidimensional sensing data and extract features from the multidimensional sensing data to generate scene complexity features, task attribute features and resource redundancy features; The scenario complexity features, the task attribute features, and the resource redundancy features are input into a multi-objective decision model, and a lightweight strategy parameter set is generated through multi-objective optimization processing. Based on the lightweight strategy parameter set, the initial model is reconstructed and the running parameters are configured to generate a lightweight execution model. Obtain the target task input data, and use the lightweight execution model to perform inference processing on the target task input data to output the target task processing result; The task execution accuracy value is determined based on the target task processing result, and the computing power consumption and energy consumption value of the lightweight execution model during the execution inference process are monitored. Based on the task execution accuracy value, the computing power consumption value and the energy consumption value, the running performance data is generated. The operational performance data is compared with a preset performance benchmark threshold, and an operational performance analysis level is generated based on the comparison results. In response to the operational performance analysis level, a correction operation is performed on the lightweight strategy parameter set or a weight update operation is performed on the multi-objective decision model. The corrected parameters or the updated model are then applied to the lightweight execution model generation process at the next time step to optimize the processing of subsequent target tasks.

2. The adaptive lightweight task execution and optimization method as described in claim 1, characterized in that, Collect multidimensional sensing data, and extract features from the multidimensional sensing data to generate scene complexity features, task attribute features, and resource redundancy features, including: The system collects environmental sensing data through sensors and uses a feature extraction network to extract the number of scene targets, target movement speed, ambient light intensity, and obstacle density from the environmental sensing data. Based on the preset scene complexity quantification module, the number of scene targets, the target movement speed, the ambient light intensity, and the obstacle density are weighted and summed to generate scene complexity features; The task scheduling module obtains task scheduling data, parses the task scheduling data to extract task priority features and precision threshold features, and generates task attribute features based on the task priority features and precision threshold features. The device resource monitoring module collects device resource monitoring data and extracts central processing unit utilization, memory remaining, battery power and computing power utilization from the device resource monitoring data. Based on the preset resource redundancy determination module, the CPU utilization rate, the remaining memory, the battery power, and the computing power utilization rate are weighted and processed to generate resource redundancy characteristics.

3. The adaptive lightweight task execution and optimization method as described in claim 1, characterized in that, The scenario complexity features, task attribute features, and resource redundancy features are input into a multi-objective decision model, and a lightweight strategy parameter set is generated through multi-objective optimization processing, including: Define a multi-objective optimization objective, which includes an accuracy objective, a computing power consumption objective, and an energy consumption objective. Define a lightweight strategy parameter set, which includes quantization bit width parameters, pruning ratio parameters, inference frame rate parameters, and module activation parameters; The task attribute features are analyzed to extract the core precision threshold and task priority. Precision constraints are set based on the core precision threshold, and weight coefficients are generated based on the task priority. The resource redundancy characteristics are analyzed to determine the computing power constraint boundary and energy consumption constraint boundary; Based on the multi-objective optimization objective, the accuracy constraint, the computing power constraint boundary, and the energy consumption constraint boundary, a multi-objective decision model is constructed, and the scenario complexity feature, the core accuracy threshold, and the resource redundancy feature are used as inputs to the multi-objective decision model. Based on the multi-objective decision model and the weight coefficients, a weighted objective function is formed, and an optimization search is performed on the parameter space of the lightweight strategy parameter set to obtain a solution vector through iterative optimization. The solution vector is decoded into the quantization bit width parameter, the pruning ratio parameter, the inference frame rate parameter, and the module activation parameter to obtain the final lightweight strategy parameter set.

4. The adaptive lightweight task execution and optimization method as described in claim 1, characterized in that, Based on the lightweight strategy parameter set, the initial model is reconstructed and runtime parameters are configured to generate a lightweight execution model, including: The quantization bit width parameter, pruning ratio parameter, inference frame rate parameter, and module activation parameter are obtained from the lightweight strategy parameter set. Based on the quantization bit width parameter, the initial model is quantized to generate the quantized model. Based on the clipping ratio parameter and the scene complexity feature, the quantized model is clipped to generate a clipped model. Based on the inference frame rate parameter, configure inference scheduling for the pruned model to generate a scheduled and configured model. Based on the module activation parameters, the activation status of the model configuration module after scheduling configuration is determined, and a lightweight execution model is generated.

5. The adaptive lightweight task execution and optimization method as described in claim 1, characterized in that, Acquire target task input data, and use the lightweight execution model to perform inference processing on the target task input data to output the target task processing result, including: The data sampling interval is determined based on the inference frame rate parameter in the lightweight strategy parameter set. Real-time environmental data is collected through the sensor interface according to the data sampling interval to generate target task input data. Load the lightweight execution model and activate the corresponding functional modules based on the module activation parameters in the lightweight strategy parameter set; Input the target task input data into the lightweight execution model; The lightweight execution model is used to perform inference processing on the target task input data, and the raw output data is obtained from the output layer of the lightweight execution model. The original output data is post-processed based on the task attribute features to generate the target task processing result.

6. The adaptive lightweight task execution and optimization method as described in claim 1, characterized in that, Based on the target task processing result, the task execution accuracy value is determined, and the computing power consumption and energy consumption value of the lightweight execution model during the execution inference process are monitored. Based on the task execution accuracy value, the computing power consumption value, and the energy consumption value, runtime performance data is generated, including: The target task processing results are analyzed to obtain analytical indicators including confidence scores or action feedback errors. The analytical indicators are then mapped to a preset precision quantization range to obtain a task execution precision value that characterizes the quality of the current task completion. During the time window in which the lightweight execution model performs the inference process, the CPU load rate, GPU computing power utilization rate, and memory bandwidth utilization rate are sampled through the system monitoring interface, and a weighted average is performed on the CPU load rate, GPU computing power utilization rate, and memory bandwidth utilization rate to obtain the computing power utilization value. The instantaneous power or battery charge decay rate recorded by the power management module of the monitoring device is used to determine the power consumption level per unit time based on the duration of the inference process and the instantaneous power or battery charge decay rate, thereby obtaining the energy consumption value. The task execution accuracy value, computing power occupancy value, and energy consumption value are timestamped and vectorized to generate runtime performance data.

7. The adaptive lightweight task execution and optimization method as described in claim 1, characterized in that, The operational performance data is compared with a preset performance benchmark threshold, and an operational performance analysis level is generated based on the comparison results, including: The running performance data is analyzed to extract the task execution accuracy value, computing power consumption value and energy consumption value respectively, and a preset performance benchmark threshold is obtained. The performance benchmark threshold includes the accuracy benchmark lower limit, computing power benchmark upper limit and energy consumption benchmark upper limit. Determine whether the task execution accuracy value is greater than or equal to the lower limit of the accuracy benchmark, and determine whether the computing power occupancy value and the energy consumption value are less than or equal to the upper limit of the computing power benchmark and the upper limit of the energy consumption benchmark, respectively, so as to determine the compliance status of the task execution accuracy value, the computing power occupancy value and the energy consumption value. If the target status of the task execution accuracy value, the computing power occupancy value, and the energy consumption value are all met, then the first level of the indication strategy is generated as the operation effect analysis level. If the target status of the task execution accuracy value is yes, but the target status of either the computing power occupancy value or the energy consumption value is no, then a second level of indicator parameter fine-tuning is generated as the operation effect analysis level. If the target status of the task execution accuracy value is not met, or the target status of both the computing power occupancy value and the energy consumption value is not met, then the third level of the indicator model reconstruction is generated as the operation effect analysis level.

8. An adaptive lightweight task execution and optimization device, characterized in that, The adaptive lightweight task execution and optimization device includes: The multidimensional perception feature extraction module is used to collect multidimensional perception data and extract features from the multidimensional perception data to generate scene complexity features, task attribute features and resource redundancy features. The multi-objective decision generation module is used to input the scene complexity features, the task attribute features and the resource redundancy features into the multi-objective decision model, and generate a lightweight strategy parameter set through multi-objective optimization processing. The lightweight model construction module is used to perform model reconstruction and runtime parameter configuration on the initial model according to the lightweight strategy parameter set, and generate a lightweight execution model. The task reasoning and execution module is used to acquire target task input data, and use the lightweight execution model to perform reasoning processing on the target task input data to output the target task processing result; The execution performance monitoring module is used to determine the task execution accuracy value based on the target task processing result, and monitor the computing power consumption and energy consumption value of the lightweight execution model during the execution inference process, and generate running performance data based on the task execution accuracy value, the computing power consumption value and the energy consumption value; The performance evaluation module is used to compare the performance data with a preset performance benchmark threshold and generate a performance analysis level based on the comparison results. The strategy adaptive optimization module is used to perform a correction operation on the lightweight strategy parameter set or a weight update operation on the multi-objective decision model in response to the performance analysis level, and apply the corrected parameters or updated model to the lightweight execution model generation process at the next time step to optimize the processing of subsequent target tasks.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and an adaptive lightweight task execution and optimization program stored in the memory and executable on the processor. When executed by the processor, the adaptive lightweight task execution and optimization program implements the steps of the adaptive lightweight task execution and optimization method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores an adaptive lightweight task execution and optimization program, which, when executed by a processor, implements the steps of the adaptive lightweight task execution and optimization method as described in any one of claims 1-7.

Images