Air unmanned equipment-oriented large model structure pruning and control integrated method

Abstract

This invention provides an integrated method for pruning and controlling large-scale models of unmanned aerial vehicles (UAVs), belonging to the field of model lightweighting technology. The proposed method significantly reduces model size and computational load, greatly improves inference speed and control response performance, and enables the deployment of complex neural network controllers on resource-constrained airborne platforms. Simultaneously, the lightweight model maintains or approaches the control accuracy and stability of the original uncompressed model. This method has wide applicability and strong scalability, covering various scenarios from basic flight control to complex mission control, providing a feasible, efficient, and cost-effective solution for applying large models to airborne controllers of UAVs, and possesses high engineering practical value.

Description

Technical Field

[0001] This invention relates to the field of lightweight model technology, specifically to an integrated method for pruning and controlling the structure of large models for unmanned aerial vehicles. Background Technology

[0002] With the rapid development of artificial intelligence technology, deep neural networks (DNNs) have demonstrated outstanding performance in various tasks such as classification, detection, recognition, and perception. Their application scenarios have gradually expanded to embedded systems, edge devices, mobile devices, and unmanned equipment. Onboard systems of unmanned equipment, aerial platforms, and unmanned aerial vehicles (UAVs) urgently need to rely on deep neural networks to achieve perception, decision-making, and control functions in complex environments to meet the demands of intelligent operation. However, performance improvements in deep neural networks often depend on increased model complexity, number of parameters, and computational load. These large models place extremely high demands on computing resources, storage space, power consumption, and inference latency, creating a significant contradiction with the hardware constraints of resource-constrained platforms such as onboard systems. Model lightweighting has become a core prerequisite for the practical application of deep neural networks in these scenarios, making model compression technology a research hotspot in this field. Among these, structured pruning and related compression / encoding methods, with their excellent lightweighting effects, have become important technical paths for achieving efficient model deployment.

[0003] To address the compatibility issues of large models with resource-constrained platforms, various model compression schemes have been developed in the existing technology. Among them, the "deep compression" method is widely used. This method achieves DNN model compression through a three-stage process of pruning, training quantization, and Huffman coding: First, redundant connections in the network are pruned and removed. Then, the remaining weights are quantized so that multiple connections share the same weights. Finally, Huffman coding is used to further reduce storage overhead. Using this method, classic networks such as AlexNet and VGG-16 can achieve compression of 35 to 49 times with almost no loss of model accuracy, effectively reducing model size and lowering inference power consumption and latency. Based on the structured pruning approach, filter-level pruning methods, such as the ThiNet algorithm, have been developed. ThiNet evaluates and prunes the entire filter in a convolutional neural network (CNN), significantly reducing computational cost (FLOPs) and parameter size without altering the overall network structure. Experiments on VGG-16 have demonstrated that this method can drastically reduce model size and computational cost while incurring only minor performance loss, balancing compression efficiency with applicability to standard deep learning tasks. This provides a feasible path for deploying lightweight models on resource-constrained devices.

[0004] However, while existing structured pruning and model compression techniques have achieved the transformation of large models into lightweight models and can adapt to offline inference or soft real-time tasks in embedded and edge devices, there are significant technological gaps when it comes to strictly real-time control or feedback control. Most published research focuses on tasks such as perception, inference, classification, and recognition, with a core focus on model compression ratio, storage efficiency, and inference speed, neglecting the compatibility of pruned and compressed models with complex requirements in control tasks, such as stability, response time, system dynamics, sensor input adaptation, and multi-input / output (MIMO) collaboration. Especially for onboard controllers such as flight controllers for unmanned equipment, aerial platforms, and UAVs, which have extremely high requirements for real-time performance, stability, and safety, existing technologies cannot provide a systematic solution of "large model + pruning and compression + onboard deployment + real-time control." This makes it difficult to reliably apply high-performance deep neural networks to such complex control scenarios and fails to meet the actual needs of intelligent real-time control of unmanned equipment. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an integrated method for large-scale model pruning and control of unmanned aerial vehicles, solving the technical problem that existing structured pruning and model compression technologies cannot meet the actual needs of intelligent real-time control of unmanned equipment.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides an integrated method for pruning and controlling the structure of a large model of unmanned aerial vehicles, comprising: Initial structured pruning is performed on the pre-trained large neural network to obtain a domain-adaptive neural network; The domain-adaptive neural network is used as a controller for unmanned aerial vehicles, and a set of control sensitivity indicators for unmanned aerial vehicles is obtained. The set of control sensitivity indicators is used as hard or soft constraints for the retraining and pruning of the domain-adaptive neural network. The domain-adaptive neural network, which serves as the controller, is retrained and pruned during the training process to obtain the optimal lightweight model for unmanned aerial vehicles. Intermediate models from the retraining process are retained as candidate models. During pruning, the control sensitivity index of the structural units of the domain-adaptive neural network is analyzed. The control loss and model complexity are modeled simultaneously during pruning, with the goal of minimizing the control task loss. The optimal lightweight model is subjected to safety constraint acceptance. If the acceptance is successful, the optimal lightweight model is selected as the model to be deployed. Otherwise, a candidate model with lower pruning intensity is automatically selected for safety constraint acceptance until the model passes the acceptance test and is obtained as the model to be deployed. Deploy the model to be deployed to the controller.

[0007] Preferably, the analysis of the control sensitivity index of the structural units of the domain-adaptive neural network includes: Define the first in the control sensitivity index set Control sensitivity index of each structural unit , Used for direct quantization of the The contribution of each structural unit to the control output and controller response is expressed as follows: in, For the first Intermediate feature output of each structural unit; To control the task loss function; It is the expectation operation, representing the distribution of task data. Above all samples The average value reflects the average sensitivity of the domain-adaptive neural network in typical task scenarios.

[0008] Preferably, the step of simultaneously modeling control loss and model complexity during pruning includes: The objective function of the model is: in, In the representation of the domain-adaptive neural network, the first The retention flag for each structural unit: 0 indicates not to be retained, and 1 indicates to be retained. For the first The total cost of each structural unit As a weighting factor, To control the task loss function.

[0009] Preferably, the real-time constraints include: Single inference delay of the pruned model satisfy: ,in, This is the maximum control cycle allowed by the controller; Meanwhile, the closed-loop end-to-end time is less than the maximum control cycle allowed by the controller.

[0010] Preferably, the security constraints include: The stability and robustness of the model under extreme scenarios were verified using a boundary condition test set.

[0011] Preferably, after deploying the model to be deployed to the controller, the integrated method of large model structure pruning and control further includes: The controller is monitored in real time, and the pruning mode is switched online based on the controller's control error and health indicators. Specifically: when or At that time, we switch from the current lightweight model to a candidate model that retains more parameters, is more conservative in computation, and has more stable performance; among them, To track errors in real time, To track the error threshold, To control the output variance, To control the output threshold.

[0012] Preferably, after deploying the model to be deployed to the controller, the integrated method of large model structure pruning and control further includes: During controller operation, closed-loop metrics are continuously collected, and online calibration is performed periodically. Online calibration includes: If a long-term deviation is detected, online fine-tuning is triggered or the pruning threshold is adjusted directly by reconstructing the model structure or updating the parameters. If a short-term deviation is detected, the pruning decision parameters and control sensitivity estimates are updated in real time using sliding window statistics or recursive least squares estimation methods.

[0013] Secondly, the present invention provides an integrated system for pruning and controlling large-scale models of unmanned aerial vehicles (UAVs), which is used to execute the integrated method for pruning and controlling large-scale models of UAVs as described above.

[0014] Thirdly, the present invention provides a computer-readable storage medium storing a computer program for the integrated pruning and control of large model structures for unmanned aerial vehicles, wherein the computer program causes a computer to execute the integrated pruning and control method for large model structures for unmanned aerial vehicles as described above.

[0015] Fourthly, the present invention provides an electronic device, comprising: One or more processors; Memory; and One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing integrated pruning and control of large model structures for unmanned aerial vehicles as described above.

[0016] (III) Beneficial Effects This invention provides an integrated method for pruning and controlling the structure of large-scale unmanned aerial vehicles. Compared with existing technologies, it has the following advantages: The proposed integrated method for large-scale model pruning and control of unmanned aerial vehicles (UAVs) significantly reduces model size and computational load, while greatly improving inference speed and control response performance. It enables the deployment of complex neural network controllers on resource-constrained airborne platforms. Simultaneously, the lightweight model maintains or nearly matches the control accuracy and stability of the original uncompressed model. This method has broad applicability and strong scalability, covering various scenarios from basic flight control to complex mission control. It provides a feasible, efficient, and cost-effective solution for applying large models to airborne controllers of UAVs, demonstrating significant engineering practical value. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a block diagram of an integrated method for pruning and controlling large-scale models of unmanned aerial vehicles, according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This application provides an integrated method for large-scale model pruning and control of unmanned aerial vehicles, which solves the technical problem that existing structured pruning and model compression technologies cannot meet the actual needs of intelligent real-time control of unmanned vehicles, and enables the deployment of complex neural network controllers on resource-constrained onboard platforms.

[0021] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows: While existing structured pruning and model compression techniques can significantly reduce model size, computational load, and storage requirements, excessive pruning often leads to decreased model expressiveness and performance degradation, failing to meet the demands of highly complex tasks (such as perception and control tasks). Especially in aircraft or unmanned equipment controllers requiring closed-loop control, real-time response, and system stability and robustness, relying solely on existing compressed models may not meet safety and performance requirements in terms of control accuracy, response speed, and reliability.

[0022] To address the challenge of balancing high model capacity and real-time control performance when deploying on resource-constrained platforms (such as embedded systems, edge devices, or onboard systems for unmanned aerial vehicles), this invention proposes a structured pruning + real-time control method. This method combines the pruning and compression of large models with control algorithms optimized for real-time control tasks. The compressed neural network possesses sufficient expressive and decision-making capabilities while meeting the limitations of resource-constrained platforms (onboard computing / storage / power consumption), ensuring the controller maintains real-time performance, stability, and robustness. This approach enables the deployment of highly complex neural network controllers on resource-constrained, real-time-critical unmanned aerial vehicles.

[0023] Furthermore, the embodiments of the present invention also strive to solve the following specific problems: to maximize the preservation of the network's expressive ability for key features of the control task (such as adaptability to environmental input, sensor data, multi-channel fusion, and complex dynamic models) during the pruning / compression process; to ensure that the inference latency and response time meet the flight control system's requirements for real-time feedback and closed-loop control after deployment to onboard hardware; and to provide an overall solution that enables the controller to have sufficient stability and robustness in the actual operating environment of flight / unmanned equipment, and to be tolerant of environmental changes, sensor noise, hardware limitations, and dynamic disturbances.

[0024] In summary, the purpose of this invention is to fill the gap in the prior art—to provide an aerial unmanned equipment with a neural network control solution that is lightweight (resource-friendly), highly capable (with complex control logic), and meets the requirements of real-time control and stability, so that large models can be truly implemented in the onboard controller of unmanned equipment.

[0025] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0026] This invention provides an integrated method for pruning and controlling the structure of a large model of unmanned aerial vehicles, such as... Figure 1 As shown, it includes: S1. Perform initial structured pruning on the pre-trained large neural network to obtain a domain-adaptive neural network; S2. Use the domain-adaptive neural network as the controller for unmanned aerial vehicles and obtain the control sensitivity index set for unmanned aerial vehicles. Use the control sensitivity index set as the hard or soft constraint for retraining and pruning the domain-adaptive neural network. S3. Retrain the domain-adaptive neural network used as the controller and prune it during the training process to obtain the optimal lightweight model for unmanned aerial equipment, and retain the intermediate model in the retraining process as a candidate model. During pruning, analyze the control sensitivity index of the structural unit of the domain-adaptive neural network, and model the control loss and model complexity at the same time during pruning, with the goal of minimizing the control task loss. S4. Perform security constraint acceptance on the optimal lightweight model. If the acceptance is successful, the optimal lightweight model will be used as the model to be deployed. Otherwise, automatically select a candidate model with lower pruning intensity for security constraint acceptance until the model passes the acceptance and is obtained as the model to be deployed. S5. Deploy the model to be deployed to the controller.

[0027] The integrated method for large-scale model pruning and control proposed in this invention for unmanned aerial vehicles can significantly reduce model size and computational load, greatly improve inference speed and control response performance, and enable the deployment of complex neural network controllers on resource-constrained airborne platforms. Simultaneously, the lightweight model maintains or approaches the control accuracy and stability of the original uncompressed model. This method has a wide range of applications and strong scalability, covering various scenarios from basic flight control to complex mission control. It provides a feasible, efficient, and cost-effective solution for applying large models to airborne controllers of unmanned equipment, and has high engineering practical value.

[0028] In step S1, the pre-trained large neural network undergoes initial structured pruning to obtain a domain-adaptive neural network. The specific implementation process is as follows: By performing structured pruning on a large, pre-trained neural network, redundant channels, filters, or other structural units are removed (the pruning is tailored to the specific task requirements), significantly reducing the number of network parameters and computational cost. This pruning and compression allows large models, originally designed for high-performance hardware (such as GPUs / servers), to be adapted to resource-constrained, computationally limited, and power-constrained onboard controllers. Through such compression and optimization, the model becomes lightweight and can potentially be deployed on embedded platforms in aircraft / unmanned equipment.

[0029] In step S2, the domain-adaptive neural network is used as the controller for the unmanned aerial vehicle (UAV), and a set of control sensitivity indicators for the UAV is obtained. This set of control sensitivity indicators serves as the hard or soft constraint for the retraining and pruning of the domain-adaptive neural network. The specific implementation process is as follows: Neural networks that adapt to the domain Mapping relationships used directly as controllers: in, To preserve the parameter set after pruning, yes The state at any given moment, yes The output of the time-domain adaptive neural network is determined. A set of sensitivity metrics (e.g., upper limit of tracking error) is incorporated during the modeling phase. Transition time (such as the upper limit of oscillation amplitude), and use these indicators as hard or soft constraints for subsequent pruning and retraining.

[0030] Existing technologies generally employ standard system dynamics or data-driven approaches to establish state-action mappings for unmanned equipment; commonly used tools and frameworks include MATLAB / Simulink, ROS, PyTorch, and TensorFlow for simulation and data acquisition. Closed-loop controllers typically use... The form of representation, in which It's a state. It controls the input.

[0031] In this embodiment of the invention, the structured pruned neural network is directly used as the mapping relationship for the controller. This effectively solves the problem that the pruned model is used as a simple inference network and its control effect is ignored. It ensures that the pruning strategy is coupled with the control performance index, thereby ensuring that the safety and stability requirements of closed-loop control can still be met after pruning.

[0032] In step S3, the domain-adaptive neural network, which serves as the controller, is retrained, and pruning is performed during the training process to obtain the optimal lightweight model for unmanned aerial vehicles. Intermediate models from the retraining process are retained as candidate models. During pruning, the control sensitivity index of the structural units of the domain-adaptive neural network is analyzed. Control loss and model complexity are modeled simultaneously during pruning, with the goal of minimizing the control task loss. The specific implementation process is as follows: Define control sensitivity indicators in the control sensitivity indicator set. (Regarding the first) Each structural unit (channel / filter / submodule) is used to directly quantify the contribution of that structural unit to the control output and controller response. in, For the first The intermediate feature output of each structural unit, The loss function is used to measure the control performance, such as the cumulative sum of trajectory tracking errors and the penalty term for safety constraint violations. The smaller the loss, the better the control effect. It is the expectation operation, representing the distribution of task data. Above all samples The average value reflects the model's average sensitivity in typical task scenarios.

[0033] During pruning, the control loss and model complexity are modeled simultaneously, with the objective function as follows: in, Indicates the first Retention flags for each structural unit, The total cost of this structural unit includes computing cost, latency cost, and energy cost. This is a weighting factor.

[0034] During training, the sensitivity index will be controlled. By incorporating constraints or regularization terms, low-sensitivity structural units are preferentially pruned; efficient solutions can be obtained using continuous relaxation or heuristic greedy strategies.

[0035] The loss function during training is as follows: in, Used to limit and control the mutation of output. Used to constrain state deviations. , For the weighting factor; The expectation symbol represents the mathematical expectation operation. for The state reference value at any given time; for The state at any given moment; yes The output of the domain-adaptive neural network at any given time; It is the previous sampling time. The output of the domain-adaptive neural network at any given time; The sampling interval is denoted as .

[0036] During retraining, a constrained optimizer (such as projected gradient, Adam / RMSprop with penalty, or L-BFGS constrained variant) is used for retraining, and key indicators (tracking error peak, convergence time, oscillation amplitude) are verified in batches using closed-loop simulation during training, serving as the basis for early stopping or model selection.

[0037] During the training process, intermediate models from each training round are retained as candidate models.

[0038] In existing technologies, the conventional approach after pruning is to retrain the network to restore accuracy; the commonly used loss is the task loss plus conventional regularization terms such as weight decay. The retraining method of this invention not only restores inference performance but also prioritizes the restoration and guarantee of control stability; by embedding dynamic constraints, it significantly reduces the risk of control oscillations or response degradation caused by pruning. Simultaneously, the control sensitivity index designed in this invention addresses the problem that inference accuracy sensitivity (such as classification error) alone cannot reflect the impact on control performance, allowing pruning decisions to directly use the control task loss as the measurement standard, thereby more reliably guaranteeing stability and security.

[0039] Furthermore, the control-aware pruning optimization objective of this embodiment of the invention directly takes control performance as the first optimization criterion, avoiding the destruction of closed-loop stability due to simply pursuing computational compression; by introducing a cost term, the goal of "maximizing computational / energy savings while satisfying control constraints" is achieved.

[0040] In step S4, the optimal lightweight model undergoes joint acceptance testing for real-time and security constraints. If the acceptance test passes, the optimal lightweight model is selected as the model to be deployed. Otherwise, a candidate model with lower pruning intensity is automatically selected for joint acceptance testing for real-time and security constraints until the model passes the acceptance test, thus obtaining the model to be deployed. The specific implementation process is as follows: First, verify the control cycle constraint: ensure the single inference delay of the model after pruning. satisfy ,in, The maximum control cycle allowed by the controller is defined; simultaneously, it is verified whether the closed-loop end-to-end time (sensor sampling → inference → execution) is less than the maximum control cycle allowed by the controller. In this embodiment of the invention, delay constraints are used as strong constraints to exclude models that violate real-time control requirements due to computational savings.

[0041] For models that meet the control cycle constraints, further safety constraint acceptance is performed: verifying their steady-state performance and robustness under boundary conditions. This involves verifying the model's stability and robustness under extreme scenarios such as wind disturbance, load variations, and sensor noise using a boundary condition test set, ensuring control safety.

[0042] In step S5, the model to be deployed is deployed to the controller. The specific implementation process is as follows: The model to be deployed obtained in step S4 is deployed to the controller to complete the pruning model deployment.

[0043] In specific implementation, the embodiments of the present invention also include: Step S6: Monitor the controller in real time and switch the pruning mode online based on the controller's control error and health indicators. The specific implementation process is as follows: like or When switching to the low pruning / revert controller in, To track errors in real time, To track the error threshold, To control the output variance, To control the output threshold, the threshold is set by safety requirements. The above formula indicates that when the controller performance deteriorates or the controller health status is abnormal, it switches from the current lightweight (highly pruned) model to a "candidate model" that retains more parameters, performs more conservative calculations, and has more stable performance, in order to ensure the control safety of unmanned equipment.

[0044] In practical implementation, two rollback paths are supported: cold backup and hot switchover (low pruning intensity model), and the switchover latency is guaranteed to be within an acceptable range (switchover time). By using online switching based on control performance as a criterion, the safety and robustness of the controller are enhanced under sudden changes in operating conditions or disturbances, avoiding loss of control due to misjudgment.

[0045] Step S7: During controller operation, continuously collect closed-loop metrics (tracking error, control output statistics, delay, power consumption, etc.) and periodically perform online calibration: If a long-term deviation is detected, trigger online fine-tuning (short-term retraining) or adjust the pruning threshold; if a short-term deviation is detected, use sliding window statistics or recursive least squares estimation to update the pruning decision parameters and sensitivity estimates. The specific implementation process is as follows: During controller operation, closed-loop metrics (tracking error, control output statistics, delay, power consumption, etc.) are continuously collected, and online calibration is performed periodically. If a long-term deviation is detected, online fine-tuning (short-term local retraining) or direct adjustment of the pruning threshold is triggered to repair the performance deviation from the root by reconstructing the model structure or updating the parameters.

[0046] If a short-term deviation is detected (a short-term performance fluctuation caused by transient disturbances, noise, or other temporary factors), the pruning decision parameters and control sensitivity estimates are updated in real time using sliding window statistics or recursive least squares estimation methods. This allows for rapid adaptation to the current operating conditions without changing the model structure, thus smoothly suppressing transient performance fluctuations.

[0047] Short-term deviation refers to the instantaneous performance fluctuation of the controller within a short time window (such as within a few control cycles). It is usually caused by temporary factors such as environmental disturbances, sensor noise, and sudden changes in operating conditions. The deviation has a short duration and small amplitude and will not continue to deteriorate.

[0048] Long-term deviation refers to the performance shift of the controller that persists over a long time scale (such as more than dozens of control cycles). It reflects the fundamental degradation of the model itself (such as the decline in the model's generalization ability due to environmental changes, or irreversible performance degradation due to excessive pruning). The deviation will continue to accumulate, and if no intervention is made, it will lead to the controller going out of control.

[0049] This invention establishes a "dual-loop optimization" mechanism. The fast loop ensures the real-time safety and instantaneous robustness of the controller under extreme conditions through online switching triggered by thresholds and short-term deviation correction. The slow loop achieves continuous iterative optimization of the model structure and control parameters through periodic offline training and online fine-tuning of long-term deviations, avoiding irreversible performance degradation caused by one-time pruning and ensuring the performance stability and gradual improvement of the controller under long-term operation.

[0050] This invention also provides an integrated system for pruning and controlling large-scale models of unmanned aerial vehicles, which is used to execute the integrated pruning and control method for large-scale models of unmanned aerial vehicles as described above.

[0051] This invention also provides a computer-readable storage medium storing a computer program for the integrated pruning and control of large model structures for unmanned aerial vehicles, wherein the computer program causes a computer to execute the aforementioned integrated pruning and control method for large model structures for unmanned aerial vehicles.

[0052] This invention also provides an electronic device, including: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing integrated large model structure pruning and control as described above for aerial unmanned equipment.

[0053] In summary, compared with existing technologies, it has the following beneficial effects: The embodiments of this invention significantly reduce model size and computational load; significantly improve inference and control response speed; allow deployment of complex neural network controllers on resource-constrained onboard platforms; maintain or approach the control performance and stability of uncompressed models; and have a wide range of applications and strong scalability, covering everything from basic flight control to complex mission control. Thus, the embodiments of this invention provide a feasible, effective, and cost-effective solution for introducing large models into onboard controllers for unmanned equipment, and have high practical value.

[0054] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0055] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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. Such 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.

Claims

1. A method for integrating structural pruning and control of large-scale unmanned aerial vehicles, characterized in that, include: Initial structured pruning is performed on the pre-trained large neural network to obtain a domain-adaptive neural network; The domain-adaptive neural network is used as a controller for unmanned aerial vehicles, and a set of control sensitivity indicators for unmanned aerial vehicles is obtained. The set of control sensitivity indicators is used as hard or soft constraints for the retraining and pruning of the domain-adaptive neural network. The domain-adaptive neural network, which serves as the controller, is retrained and pruned during the training process to obtain the optimal lightweight model for unmanned aerial vehicles. Intermediate models from the retraining process are retained as candidate models. During pruning, the control sensitivity index of the structural units of the domain-adaptive neural network is analyzed. The control loss and model complexity are modeled simultaneously during pruning, with the goal of minimizing the control task loss. The optimal lightweight model is subjected to safety constraint acceptance. If the acceptance is successful, the optimal lightweight model is selected as the model to be deployed. Otherwise, a candidate model with lower pruning intensity is automatically selected for safety constraint acceptance until the model passes the acceptance test and is obtained as the model to be deployed. Deploy the model to be deployed to the controller.

2. The integrated method for pruning and controlling large-scale models of unmanned aerial vehicles as described in claim 1, characterized in that, The analysis of the control sensitivity index of the structural units of the domain-adaptive neural network includes: Define the first in the control sensitivity index set Control sensitivity index of each structural unit , Used for direct quantization of the The contribution of each structural unit to the control output and controller response is expressed as follows: in, For the first Intermediate feature output of each structural unit; To control the task loss function; It is the expectation operation, representing the distribution of task data. Above all samples The average value reflects the average sensitivity of the domain-adaptive neural network in typical task scenarios.

3. The integrated method for pruning and controlling large-scale models of unmanned aerial vehicles as described in claim 1, characterized in that, The method of simultaneously modeling control loss and model complexity during pruning includes: The objective function of the model is: in, In the representation of the domain-adaptive neural network, the first The retention flag for each structural unit: 0 indicates not to be retained, and 1 indicates to be retained. For the first The total cost of each structural unit As a weighting factor, To control the task loss function.

4. The integrated method for pruning and controlling large-scale models of unmanned aerial vehicles as described in claim 1, characterized in that, The real-time constraints include: Single inference delay of the pruned model satisfy: ,in, This is the maximum control cycle allowed by the controller; Meanwhile, the closed-loop end-to-end time is less than the maximum control cycle allowed by the controller.

5. The integrated method for pruning and controlling large-scale models of unmanned aerial vehicles as described in claim 1, characterized in that, The security constraints include: The stability and robustness of the model under extreme scenarios were verified using a boundary condition test set.

6. The integrated method for pruning and controlling large-scale models of unmanned aerial vehicles as described in any one of claims 1 to 5, characterized in that, After deploying the model to be deployed to the controller, the integrated large model structure pruning and control method further includes: The controller is monitored in real time, and the pruning mode is switched online based on the controller's control error and health indicators. Specifically: when or At that time, we switch from the current lightweight model to a candidate model that retains more parameters, is more conservative in computation, and has more stable performance; among them, To track errors in real time, To track the error threshold, To control the output variance, To control the output threshold.

7. The integrated method for pruning and controlling large-scale models of unmanned aerial vehicles as described in any one of claims 1 to 5, characterized in that, After deploying the model to be deployed to the controller, the integrated large model structure pruning and control method further includes: During controller operation, closed-loop metrics are continuously collected, and online calibration is performed periodically. Online calibration includes: If a long-term deviation is detected, online fine-tuning is triggered or the pruning threshold is adjusted directly by reconstructing the model structure or updating the parameters. If a short-term deviation is detected, the pruning decision parameters and control sensitivity estimates are updated in real time using sliding window statistics or recursive least squares estimation methods.

8. A large-scale model structure pruning and control integrated system for unmanned aerial vehicles, characterized in that, The integrated system for pruning and controlling large model structures for unmanned aerial vehicles is used to execute the integrated method for pruning and controlling large model structures for unmanned aerial vehicles as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, It stores a computer program for the integrated pruning and control of large model structures for unmanned aerial vehicles, wherein the computer program causes the computer to execute the integrated pruning and control method for large model structures for unmanned aerial vehicles as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: One or more processors; Memory; as well as One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing integrated pruning and control of large model structures for aerial unmanned equipment as described in any one of claims 1 to 7.

Images