Intelligent controller complexity and scheduling collaborative design method and system

Abstract

The invention discloses an intelligent controller complexity and scheduling collaborative design method and system, and relates to the field of intelligent control. The method comprises the following steps of: establishing a control performance prediction function of the intelligent controller, describing a mutual relationship among the control performance, the complexity of the intelligent controller, the end-to-end response time and the physical characteristics of a controlled object, and fitting the mutual relationship through a machine learning regression model; establishing an execution model of the intelligent controller on a heterogeneous computing device, including establishing a mapping relationship between an intelligent control algorithm and a computing task execution model on a heterogeneous platform, and seamlessly combining controller design with heterogeneous resource allocation; an end-to-end optimization method of controller complexity and scheduling is established to determine the optimal controller complexity, heterogeneous resource allocation scheme and real-time scheduling strategy, and an end-to-end solution is provided for intelligent controller design and deployment on a heterogeneous computing platform.

Description

Technical Field

[0001] The present invention relates to the field of intelligent control, and in particular to a complexity and scheduling collaborative design method and system for an intelligent controller. Background Art

[0002] Modern control systems, such as robots, drones, and autonomous vehicle systems, are gradually adopting intelligent control based on artificial intelligence and machine learning. Intelligent control is often based on data-driven or machine learning methods to meet the performance requirements of complex and diverse control tasks. For example, deep neural networks have made significant progress in feedback control in recent years by leveraging the combined advantages of machine learning's scalability and analytical traceability.

[0003] At the same time, in order to adapt to the needs of machine learning-based algorithms, the structure of computing platforms is developing towards heterogeneous computing architectures, which integrate multiple computing resources such as CPUs, GPUs, and FPGAs. By combining accelerators (such as GPUs) with traditional CPUs, the calculation of intelligent control strategies can be accelerated, which is particularly important in intelligent feedback control scenarios with strict timing constraints. However, the mechanism of intelligent control has become increasingly complex with the introduction of heterogeneous computing devices, resulting in major challenges in design work in terms of runtime latency assurance and reliability assurance.

[0004] Mainstream machine learning tasks are usually evaluated based on accuracy and are closely related to the complexity and size of the algorithm. However, for intelligent controllers, the main design goal is to improve control performance, such as minimizing the error between the physical state and the reference value or reducing the physical state convergence time. There is a modeling gap between the control performance in real systems and the accuracy of machine learning algorithms, because the control performance is affected not only by the accuracy of the intelligent controller in deriving the control strategy, but also by the characteristics of the physical objects involved. This gap is exacerbated by the difficult relationship between computational latency and the final intelligent control performance due to the real-time nature of intelligent controllers.

[0005] In addition, when multiple intelligent control tasks are executed in parallel on the same computing platform and need to meet real-time requirements, the complexity of computationally intensive algorithms must be limited and properly balanced due to limited computing resources. The scheduling methods used for these intelligent control tasks also significantly affect real-time performance. Unlike traditional homogeneous processors, the interleaved execution mode across heterogeneous processors strengthens the dependencies and competition between parallel tasks, making control performance difficult to predict.

[0006] Therefore, technicians in this field are committed to developing a method and system for co-designing the complexity and scheduling of intelligent controllers. This provides an end-to-end solution for the design of intelligent controllers and their deployment on heterogeneous computing platforms, improves control performance, and is applicable to any multi-task intelligent control scenario that shares a heterogeneous computing platform. Summary of the invention

[0007] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is to reduce the execution time delay and reliability of the intelligent controller in heterogeneous computing devices and improve the control performance.

[0008] To achieve the above objectives, the present invention provides a method for collaboratively designing the complexity and scheduling of an intelligent controller, including complexity selection and real-time scheduling of an intelligent controller algorithm based on a deep neural network.

[0009] Furthermore, a fine-grained execution model of an intelligent controller on a heterogeneous computing device based on a self-suspended segmentation model is included.

[0010] Furthermore, a mapping relationship between each layer of the intelligent control algorithm and the sub-thread execution model of the computing task on the heterogeneous platform is constructed.

[0011] Furthermore, a data-driven control performance prediction function is included to characterize the relationship between the closed-loop control performance and the complexity and accuracy of the intelligent controller algorithm and the end-to-end response time.

[0012] Further, the following steps are included:

[0013] Step 1: Establish a control performance prediction function for the intelligent controller, describe the relationship between the control performance and the complexity of the intelligent controller algorithm, the end-to-end response time, and the physical characteristics of the controlled object, and fit the relationship through a machine learning regression model;

[0014] Step 2: Establish an execution model of the intelligent controller on the heterogeneous computing device and a resource allocation model mapping, including building a mapping relationship between the intelligent control algorithm and the computing task execution model on the heterogeneous platform;

[0015] Step 3: Establish an end-to-end optimization method for controller complexity and scheduling to determine the optimal controller complexity, heterogeneous resource allocation scheme, and real-time scheduling strategy;

[0016] Step 4: Deploy the optimization solution on the heterogeneous computing platform.

[0017] Furthermore, the step 1 comprises the following steps:

[0018] Step 1.1, construct the control system model of the physical controlled object;

[0019] Step 1.2: Control performance of intelligent controller The relevant influencing variables include the complexity c of the intelligent controller i , end-to-end response time R i 、Physical characteristic parameters of the controlled objectθ i , i∈{1,2,…,N} is the control loop number, and N is the number of control system loops. Specifically, the complexity of the intelligent controller algorithm c i is the complexity of the intelligent controller neural network, including depth and various structures, and the end-to-end response time R i is the time taken by the intelligent controller task from the start to the completion of the calculation, and the physical characteristic parameter θ i is the physical dynamic model parameter of the controlled object in a specific control instance;

[0020] Step 1.3: Construct control performance prediction function

[0021] Step 1.4: Fix the physical characteristic parameter θ of the controlled object i After that, traverse the complexity c of the intelligent controller i , end-to-end response time R i All possible values ​​within the value range constitute the machine learning input set, and the control performance is tested through simulation experiments Construct the output set;

[0022] Step 1.5: After the input set and output set data are collected, they are divided into training set and test set;

[0023] Step 1.6: Train a regression model of the relationship between the output variable and the input variable;

[0024] Step 1.7: For the test set, compare the accuracy of various models and parameters, weigh and analyze the specific input-output variable combination, and select the optimal regression model;

[0025] Step 1.8, the output of the optimal regression model is expressed as an output variable form of the control performance index to form a performance analysis statistical model;

[0026] Step 1.9: Build and implement the optimal machine learning regression model based on the trained performance analysis statistical model.

[0027] Furthermore, the step 2 comprises the following steps:

[0028] Step 2.1: Intelligent controller task τ based on deep neural network iSelf-suspend segmentation model. Specifically, self-suspend means that when a process in a computer operating system is waiting for a time-consuming operation such as an I / O device response or a task being handed over to an accelerator (such as a GPU), its execution is suspended and the state is saved to the external memory to free up memory resources and improve the overall efficiency of the system. When the conditions are met, the process is resumed and continued to execute. The self-suspend segmentation model is to treat the part of the intelligent controller task that is handed over to the accelerator as the accelerator segment that contains the self-suspend operation, and the part of the task on the CPU as the CPU segment;

[0029] Step 2.2: Intelligent controller task τ based on deep neural network i A hierarchical resource scheduling model for deep neural networks. Specifically, the model uses each layer of the deep neural network as a basic unit for resource allocation to form a hierarchical scheduling model;

[0030] Step 2.3, perform an offline running time test on each network layer of the intelligent controller to obtain the empirical worst case execution time, that is, the longest execution time recorded during the test;

[0031] Step 2.4: Map the hierarchical resource scheduling model to the self-suspending segmentation model.

[0032] Furthermore, the step 3 comprises the following steps:

[0033] Step 3.1, construct the worst case response time analysis method based on the self-suspending segmentation model;

[0034] Step 3.2: Construct the final optimization problem based on resource allocation model mapping and response time analysis;

[0035] Step 3.3: Solve the optimization problem based on reinforcement learning.

[0036] Furthermore, the step 4 comprises the following steps:

[0037] Step 4.1: According to the optimization result of step 3, select the complexity c of the deep neural network intelligent controller. i ;

[0038] Step 4.2: Based on the optimization result of step 3, a heterogeneous computing resource allocation scheme is implemented for each controller. i,j};

[0039] Step 4.3: According to the optimization result of step 3, the real-time scheduler assigns a priority ρ to the task set.

[0040] The present invention also provides an intelligent controller complexity and scheduling collaborative design system, comprising the following modules:

[0041] Module M1 is configured to establish a control performance prediction function of an intelligent controller, describe the relationship between the control performance and the complexity of the intelligent controller, the end-to-end response time, and the physical characteristics of the controlled object, and fit the relationship through a machine learning regression model;

[0042] Module M2 is configured to establish an execution model of the intelligent controller on a heterogeneous computing device, including building a mapping relationship between an intelligent control algorithm and a computing task execution model on a heterogeneous platform;

[0043] Module M3 is configured to establish an end-to-end optimization method for controller complexity and scheduling, and determine the optimal controller complexity, heterogeneous resource allocation scheme and real-time scheduling strategy;

[0044] Module M4 is configured to be able to deploy the optimization solution on heterogeneous computing platforms.

[0045] The mainstream intelligent controller design does not consider the impact of real-time scheduling, and cannot be applied to heterogeneous computing devices. The execution time delay and reliability of intelligent control algorithm tasks on heterogeneous computing devices cannot be guaranteed. The present invention introduces a complexity and scheduling collaborative design method for intelligent controllers based on deep neural networks to ensure the execution time delay and reliability of intelligent control algorithm tasks on heterogeneous computing devices. The complexity and scheduling collaborative design method for intelligent controllers based on deep neural networks selects and schedules the complexity of intelligent controller algorithms based on deep neural networks in real time, ensuring the execution time delay and reliability of intelligent control algorithm tasks on heterogeneous computing devices.

[0046] The interleaved execution mode across heterogeneous processors strengthens the dependency and competition between parallel intelligent control tasks, making the control performance difficult to predict. The present invention establishes an execution model of an intelligent controller on a heterogeneous computing device, including constructing a mapping relationship between an intelligent control algorithm and a computing task execution model on a heterogeneous platform. A fine-grained execution model of an intelligent controller on a heterogeneous computing device based on a self-suspended segmentation model includes constructing a mapping relationship between each layer of the intelligent control algorithm and a computing task sub-thread execution model on a heterogeneous platform.

[0047] Mainstream machine learning task performance evaluation methods are not applicable to real-time intelligent control systems. This paper proposes a control performance prediction function, a data-driven control performance prediction function, which characterizes the relationship between closed-loop control performance and the complexity and accuracy of the intelligent controller algorithm and the end-to-end response time.

[0048] In view of the defects in the prior art, the purpose of the present invention is to provide a method and system for collaborative design of complexity and scheduling of intelligent controllers based on deep neural networks on heterogeneous computing platforms.

[0049] According to the present invention, a method for co-designing complexity and scheduling of intelligent controllers based on deep neural networks on a heterogeneous computing platform is provided, and the method comprises the following steps:

[0050] Step S1: Establish a control performance prediction function of the intelligent controller to describe the relationship between the control performance and the complexity of the intelligent controller, the end-to-end response time and the physical characteristics of the controlled object, and fit the relationship through a machine learning regression model;

[0051] Step S2: Establishing an execution model of the intelligent controller on the heterogeneous computing device, including constructing a mapping relationship between the intelligent control algorithm and the computing task execution model on the heterogeneous platform, seamlessly combining the controller design with the heterogeneous resource allocation;

[0052] Step S3: Establish an end-to-end optimization method for controller complexity and scheduling to determine the optimal controller complexity, heterogeneous resource allocation scheme and real-time scheduling strategy, providing an end-to-end solution for intelligent controller design and deployment on heterogeneous computing platforms;

[0053] Step S4: Deploy the optimization solution on the heterogeneous computing platform.

[0054] Preferably, step S1 comprises the following steps:

[0055] Step S1.1: construct a control system model of the physical controlled object;

[0056] Step S1.2: Study the control performance of the intelligent controller and the influencing variables on the control performance The influence of the intelligent controller complexity c i , end-to-end response time R i 、Physical characteristic parameters of the controlled objectθ i , i∈{1,2,…,N} is the control loop number, and N is the number of control system loops. Specifically, the complexity of the intelligent controller algorithm c i is the complexity of the intelligent controller neural network, including depth and various structures, and the end-to-end response time R i is the time taken by the intelligent controller task from the start to the completion of the calculation, and the physical characteristic parameter θ i is the physical dynamic model parameter of the controlled object in a specific control instance;

[0057] Step S1.3: Construct a control performance prediction function based on the parameters in step S1.2

[0058] Step S1.4: Fix the physical characteristic parameter θ of the controlled object i After that, traverse the complexity c of the intelligent controller i , end-to-end response time Ri All possible values ​​within the value range constitute the machine learning input set, and its control performance is tested through simulation experiments Construct the output set;

[0059] Step S1.5: After the input and output data are collected, they are divided into training sets and test sets;

[0060] Step S1.6: training a regression model of the relationship between the output variable and the input variable;

[0061] Step S1.7: For the test set, compare the accuracy of various models and parameters, weigh and analyze the specific input-output variable combination, and select the optimal regression model;

[0062] Step S1.8: Express the output of the optimal regression model as an output variable of the control performance index to form a performance analysis statistical model;

[0063] Step S1.9: Based on the performance analysis statistical model trained in step S1.7, construct and implement the optimal machine learning regression model.

[0064] Preferably, step S2 comprises the following steps:

[0065] Step S2.1: Intelligent controller task τ based on deep neural network i A segmentation model based on self-suspension is proposed. Specifically, self-suspension refers to the process in a computer operating system that pauses its execution and saves its state to external memory when the process is waiting for time-consuming operations such as I / O device response or task transfer to accelerator (such as GPU) to release memory resources and improve the overall efficiency of the system. When the conditions are met, the process is resumed and continued to execute. The self-suspension segmentation model is to treat the part of the intelligent controller task that is transferred to the accelerator as the accelerator segment containing the self-suspension operation, and the part of the task on the CPU as the CPU segment;

[0066] Step S2.2: Intelligent controller task τ based on deep neural network i A hierarchical resource scheduling model for neural networks is proposed. Specifically, the model uses each layer of the deep neural network as the basic unit of resource allocation to form a hierarchical scheduling model;

[0067] Step S2.3: Perform an offline running time test on each network layer of the intelligent controller, and obtain the empirical worst-case execution time through repeated tests, that is, the longest execution time recorded during the test;

[0068] Step S2.4: Map the hierarchical resource scheduling model to the self-suspending segmentation model.

[0069] Preferably, step S3 comprises the following steps:

[0070] Step S3.1: construct a worst-case response time analysis method based on a self-suspended segmentation model;

[0071] Step S3.2: construct the final optimization problem based on the resource allocation model mapping of step S2 and the response time analysis of step S3.1;

[0072] Step S3.3: Establish a solution method based on reinforcement learning to solve the optimization problem in step S3.2.

[0073] Preferably, step S4 comprises the following steps:

[0074] Step S4.1: Select the complexity c of the deep neural network intelligent controller based on the optimization result of step S3 i ;

[0075] Step S4.2: Implement a heterogeneous computing resource allocation scheme for each controller according to the optimization result of step S3 {l i,j};

[0076] Step S4.3: The real-time scheduler assigns a priority ρ to the task set according to the optimization result of step S3.

[0077] The present invention also provides an intelligent controller complexity and scheduling collaborative design system based on deep neural network on a heterogeneous computing platform, the system includes the following modules:

[0078] Module M1: Establish the control performance prediction function of the intelligent controller, describe the relationship between the control performance and the complexity of the intelligent controller algorithm, the end-to-end response time and the physical characteristics of the controlled object, and fit the relationship through the machine learning regression model;

[0079] Module M2: Establishing the execution model of the intelligent controller on heterogeneous computing devices, including building the mapping relationship between the intelligent control algorithm and the computing task execution model on the heterogeneous platform, seamlessly combining the controller design with heterogeneous resource allocation;

[0080] Module M3: Establish an end-to-end optimization method for controller complexity and scheduling to determine the optimal controller complexity, heterogeneous resource allocation scheme and real-time scheduling strategy, providing an end-to-end solution for intelligent controller design and deployment on heterogeneous computing platforms;

[0081] Module M4: Deployment of optimization solutions on heterogeneous computing platforms.

[0082] Preferably, in module M1:

[0083] Module M1.1 builds the control system model of the physical controlled object;

[0084] Module M1.2: Study the control performance of intelligent controllers and the influencing variables on the control performance The influence of the intelligent controller complexity c i , end-to-end response time R i 、Physical characteristic parameters of the controlled objectθ i , i∈{1,2,…,N} is the control loop number, and N is the number of control system loops. Specifically, the complexity of the intelligent controller algorithm c i is the complexity of the intelligent controller neural network, including depth and various structures, and the end-to-end response time R i is the time taken by the intelligent controller task from the start to the completion of the calculation, and the physical characteristic parameter θ i is the physical dynamic model parameter of the controlled object in a specific control instance;

[0085] Module M1.3: Construct control performance prediction function based on parameters in module M1.2

[0086] Module M1.4: Fixed physical characteristic parameters θ of the controlled object i After that, traverse the complexity c of the intelligent controller i , end-to-end response time R i All possible values ​​within the value range constitute the machine learning input set, and its control performance is tested through simulation experiments Construct the output set;

[0087] Module M1.5: After the input and output data are collected, they are divided into training sets and test sets;

[0088] Module M1.6: Train regression models of the relationship between output variables and input variables;

[0089] Module M1.7: Compare the accuracy of various models and parameters for the test set, weigh and analyze the specific input-output variable combinations, and select the optimal regression model;

[0090] Module M1.8: Express the output of the optimal regression model as an output variable of the control performance index to form a performance analysis statistical model;

[0091] Module M1.9: Build and implement the optimal machine learning regression model based on the performance analysis statistical model trained in Module M1.7.

[0092] Preferably, in module M2:

[0093] Module M2.1: Intelligent controller task τ based on deep neural network iA segmentation model based on self-suspension is proposed. Specifically, self-suspension refers to the process in a computer operating system that pauses its execution and saves its state to external memory when the process is waiting for time-consuming operations such as I / O device response or task transfer to accelerator (such as GPU) to release memory resources and improve the overall efficiency of the system. When the conditions are met, the process is resumed and continued to execute. The self-suspension segmentation model is to treat the part of the intelligent controller task that is transferred to the accelerator as the accelerator segment containing the self-suspension operation, and the part of the task on the CPU as the CPU segment;

[0094] Module M2.2: Intelligent controller task τ based on deep neural network i A hierarchical resource scheduling model for neural networks is proposed. Specifically, the model uses each layer of the deep neural network as the basic unit of resource allocation to form a hierarchical scheduling model;

[0095] Module M2.3: Perform offline running time test on each network layer of the intelligent controller, and obtain the empirical worst-case execution time through repeated tests, that is, the longest execution time recorded during the test;

[0096] Module M2.4: Mapping the hierarchical resource scheduling model to the self-suspending segmentation model.

[0097] Preferably, in module M3:

[0098] Module M3.1: Construct the worst-case response time analysis method based on the self-suspending segmentation model;

[0099] Module M3.2: Construct the final optimization problem based on the resource allocation model mapping of module M2 and the response time analysis of module M3.1;

[0100] Module M3.3: Establish a reinforcement learning-based solution method to solve the optimization problem in module M3.2.

[0101] Preferably, in module M4:

[0102] Module M4.1: Select the complexity c of the deep neural network intelligent controller based on the optimization results of module M3 i ;

[0103] Module M4.2: Implement heterogeneous computing resource allocation scheme for each controller according to the optimization results of module M3 {l i,j};

[0104] Module M4.3: The real-time scheduler assigns priority ρ to the task set based on the optimization results of module M3.

[0105] The present invention relates to a method for collaboratively designing the complexity and scheduling of an intelligent controller based on a deep neural network on a heterogeneous computing platform, comprising: establishing a control performance prediction function of an intelligent controller, describing the relationship between the control performance and the complexity of the intelligent controller algorithm, the end-to-end response time and the physical characteristics of the controlled object, and fitting the relationship through a machine learning regression model; establishing an execution model of the intelligent controller on a heterogeneous computing device, including constructing a mapping relationship between the intelligent control algorithm and the computing task execution model on the heterogeneous platform, and seamlessly combining the controller design with the heterogeneous resource allocation; establishing an end-to-end optimization method for controller complexity and scheduling to determine the optimal controller complexity, heterogeneous resource allocation scheme and real-time scheduling strategy, and providing an end-to-end solution for the design of the intelligent controller and its deployment on a heterogeneous computing platform.

[0106] Compared with the prior art, the present invention has the following obvious substantial features and significant advantages:

[0107] 1. The present invention establishes an end-to-end optimization method for controller complexity and scheduling to determine the optimal controller complexity, heterogeneous resource allocation scheme and real-time scheduling strategy, providing an end-to-end solution for intelligent controller design and deployment on heterogeneous computing platforms.

[0108] 2. The present invention constructs a mapping relationship between each layer of the intelligent control algorithm and the sub-thread execution model of the computing task on the heterogeneous platform, seamlessly combining the controller design with the heterogeneous resource allocation.

[0109] 3. The present invention characterizes the relationship between closed-loop control performance and the complexity and accuracy of the intelligent controller and the end-to-end response time, and fits the relationship through a machine learning regression model.

[0110] 4. The present invention considers establishing the relationship between the final control performance of the intelligent controller and various influencing factors through the control performance prediction function. Compared with the prior art which only considers the accuracy of the neural network of the intelligent controller, the control performance can be improved;

[0111] 5. The present invention considers the end-to-end joint design method of the complexity of the intelligent controller based on the deep neural network and the real-time scheduling scheme, which improves the control performance compared with the method that only considers one part of it;

[0112] 6. The present invention has high universality and is applicable to any multi-task intelligent control scenario of a shared heterogeneous computing platform.

[0113] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0114] Figure 1 It is a general schematic diagram of the collaborative design of complexity and scheduling of an intelligent controller based on a deep neural network according to a preferred embodiment of the present invention;

[0115] Figure 2 It is a schematic diagram of a machine learning model of an intelligent controller performance prediction function of a preferred embodiment of the present invention;

[0116] Figure 3 It is a schematic diagram of mapping a hierarchical resource scheduling model to a self-suspending segment model according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0117] The following describes several preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0118] In the drawings, components with the same structure are indicated by the same numerical reference numerals, and components with similar structures or functions are indicated by similar numerical reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present invention does not limit the size and thickness of each component. In order to make the illustration clearer, the thickness of the components is appropriately exaggerated in some places in the drawings.

[0119] The present invention proposes a method for co-designing the complexity and scheduling of intelligent controllers based on deep neural networks on heterogeneous computing platforms, which designs the complexity and real-time scheduling strategy of neural network intelligent controllers end-to-end, thereby optimizing the overall control performance. Compared with the existing neural network or intelligent controller design methods, the algorithm introduces a control performance prediction function to comprehensively link the control performance with the relationship between various parameters. At the same time, real-time scheduling and the allocation of heterogeneous resources are considered, and the overall real-time control performance of the heterogeneous computing platform is optimized end-to-end.

[0120] The technical solution adopted by the present invention to solve this technical problem is: a method for co-designing the complexity and scheduling of an intelligent controller based on a deep neural network on a heterogeneous computing platform, comprising the following steps:

[0121] Step 1: Establish a control performance prediction function for the intelligent controller to describe the relationship between the control performance and the complexity of the intelligent controller, the end-to-end response time, and the physical characteristics of the controlled object, and fit the relationship through a machine learning regression model;

[0122] 1.1) Construct a control system model of the physical controlled object;

[0123] 1.2) Study the control performance of intelligent controllers and the influencing variables on the control performance The influence of the intelligent controller complexity c i , end-to-end response time R i 、Physical characteristic parameters of the controlled objectθ i , i∈{1,2,…,N} is the control loop number, and N is the number of control system loops. Specifically, the complexity of the intelligent controller algorithm c i is the complexity of the intelligent controller neural network, including depth and various structures, and the end-to-end response time R i is the time taken by the intelligent controller task from the start to the completion of the calculation, and the physical characteristic parameter θ i is the physical dynamic model parameter of the controlled object in a specific control instance;

[0124] 1.3) Construct a control performance prediction function based on the parameters in step 1.2)

[0125] 1.4) Fixed physical characteristic parameters θ of the controlled object i After that, traverse the complexity c of the intelligent controller i , end-to-end response time R i All possible values ​​within the value range constitute the machine learning input set, and its control performance is tested through simulation experiments Construct the output set;

[0126] 1.5) After the input and output data are collected, they are divided into training sets and test sets;

[0127] 1.6) Train the regression model of the relationship between output variables and input variables;

[0128] 1.7) For the test set, compare the accuracy of various models and parameters, weigh and analyze the specific input-output variable combination, and select the optimal regression model;

[0129] 1.8) The output of the optimal regression model is expressed as an output variable form of the control performance index to form a performance analysis statistical model;

[0130] 1.9) Based on the performance analysis statistical model trained in step 1.7), build and implement the optimal machine learning regression model.

[0131] Step 2: Establish an execution model of the intelligent controller on heterogeneous computing devices, including building a mapping relationship between the intelligent control algorithm and the computing task execution model on the heterogeneous platform, seamlessly combining controller design with heterogeneous resource allocation;

[0132] 2.1) Intelligent controller task τ based on deep neural network iA segmentation model based on self-suspension is proposed. Specifically, self-suspension refers to the process in a computer operating system that pauses its execution and saves its state to external memory when the process is waiting for time-consuming operations such as I / O device response or task transfer to accelerator (such as GPU) to release memory resources and improve the overall efficiency of the system. When the conditions are met, the process is resumed and continued to execute. The self-suspension segmentation model is to treat the part of the intelligent controller task that is transferred to the accelerator as the accelerator segment containing the self-suspension operation, and the part of the task on the CPU as the CPU segment;

[0133] 2.2) Intelligent controller task τ based on deep neural network i A hierarchical resource scheduling model for neural networks is proposed. Specifically, the model uses each layer of the deep neural network as the basic unit of resource allocation to form a hierarchical scheduling model;

[0134] 2.3) Perform offline running time tests on each network layer of the intelligent controller, and obtain the empirical worst-case execution time through repeated tests, that is, the longest execution time recorded during the test;

[0135] 2.4) Map the hierarchical resource scheduling model to the self-suspending segmentation model.

[0136] Step 3: Establish an end-to-end optimization method for controller complexity and scheduling to determine the optimal controller complexity, heterogeneous resource allocation scheme, and real-time scheduling strategy, providing an end-to-end solution for intelligent controller design and deployment on heterogeneous computing platforms;

[0137] 3.1) Construct the worst case response time analysis method based on the self-suspend segmentation model;

[0138] 3.2) construct the final optimization problem based on the resource allocation model mapping in step 2 and the response time analysis in step 3.1);

[0139] 3.3) Establish a solution method based on reinforcement learning to solve the optimization problem in step 3.2).

[0140] Step 4: Deploy the optimization solution on the heterogeneous computing platform.

[0141] 4.1) According to the optimization results of step 3, select the complexity c of the deep neural network intelligent controller i ;

[0142] 4.2) Implement heterogeneous computing resource allocation scheme for each controller according to the optimization result of step 3 {l i,j};

[0143] 4.3) The real-time scheduler assigns a priority ρ to the task set according to the optimization result in step 3;

[0144] More specifically, the present invention provides a complexity and scheduling collaborative design method for an intelligent controller based on a deep neural network on a heterogeneous computing platform. Compared with existing technical solutions, the algorithm designs the complexity and real-time scheduling strategy of the neural network intelligent controller end-to-end, thereby optimizing the overall control performance.

[0145] like Figure 1 As shown, the present invention performs end-to-end joint design of controller model complexity, heterogeneous resource allocation, and real-time scheduling strategy through control performance prediction functions, thereby improving the real-time control performance of a multi-loop control task set sharing a heterogeneous computing platform.

[0146] Step 1: Establish a control performance prediction function for the intelligent controller to describe the relationship between the control performance and the complexity of the intelligent controller, the end-to-end response time, and the physical characteristics of the controlled object, and fit the relationship through a machine learning regression model;

[0147] 1.1 Construct the control system model of the physical controlled object:

[0148]

[0149] x i and u i They represent the state vector of the controlled object and the control variable vector, respectively, i (·)and They represent the physical model of the controlled object and the intelligent controller model respectively.

[0150] 1.2) Study the control performance of intelligent controllers and the influencing variables on the control performance The influence of the intelligent controller complexity c i , end-to-end response time R i 、Physical characteristic parameters of the controlled objectθ i , i∈{1,2,…,N} is the control loop number, and N is the number of control system loops. Specifically, the complexity of the intelligent controller algorithm c i is the complexity of the intelligent controller neural network, including depth and various structures, and the end-to-end response time R i is the time taken by the intelligent controller task from the start to the completion of the calculation, and the physical characteristic parameter θ i is the controlled object model f in a specific control instance i (·) physical dynamic parameters.

[0151] 1.3) Construct a control performance prediction function based on the parameters in step 1.2)

[0152] Figure 2 A visual diagram of the model is shown.

[0153] 1.4) Fixed physical characteristic parameters θ of the controlled object i After that, traverse the complexity c of the intelligent controller i , end-to-end response time R i All possible values ​​within the value range constitute the machine learning input set, and its control performance is tested through simulation experiments Construct the output set;

[0154] 1.5) After the input and output data are collected, they are divided into training sets and test sets;

[0155] 1.6) Train the regression model of the relationship between output variables and input variables;

[0156] 1.7) For the test set, compare the accuracy of various models and parameters, weigh and analyze the specific input-output variable combination, and select the optimal regression model;

[0157] 1.8) The output of the optimal regression model is expressed as an output variable form of the control performance index to form a performance analysis statistical model;

[0158] 1.9) Based on the performance analysis statistical model trained in step 1.9, build and implement the optimal machine learning regression model.

[0159] Step 2: Establish an execution model of the intelligent controller on heterogeneous computing devices, including building a mapping relationship between the intelligent control algorithm and the computing task execution model on the heterogeneous platform, seamlessly combining controller design with heterogeneous resource allocation;

[0160] 2.1) Intelligent controller task τ based on deep neural network i Propose a segmentation model based on self-suspension M i To intelligently control the number of segments of a neural network, represents the estimated worst-case execution time of the mth CPU / accelerator fragment, i.e., the expected worst-case maximum execution time. m∈{1,2,…,M i}, D i and T i Represent the deadline and period of the task respectively.

[0161] 2.2) Intelligent controller task τ based on deep neural network i Propose a neural network hierarchical resource scheduling model l i.j Indicates the type of computing resources allocated to the j-layer neural network, i.e., l i,j =1 means allocating CPU cores for calculation, l i,j= 0 indicates the allocation of accelerator cores, j∈{1,2,…,L i}.

[0162] 2.3) Perform offline running time tests on each network layer of the intelligent controller, and obtain the empirical worst-case execution time through repeated tests, including computing load Neural network model replication cost η i,j , the cost of neuron data replication Respectively represent the computing load of the network layer on the CPU and accelerator, They represent the neuron data copying cost from the accelerator to the CPU and from the CPU to the accelerator, respectively.

[0163] 2.4) Map the hierarchical resource scheduling model to the self-suspending segmentation model. Let represents the network layer group consisting of the mth consecutive layers assigned to the CPU core, represents the network layer group consisting of the mth consecutive layers assigned to the accelerator core, then

[0164] Figure 3 A visual diagram of the mapping process is shown.

[0165] Step 3: Establish an end-to-end optimization method for controller complexity and scheduling to determine the optimal controller complexity, heterogeneous resource allocation scheme, and real-time scheduling strategy, providing an end-to-end solution for intelligent controller design and deployment on heterogeneous computing platforms;

[0166] 3.1) Construct the worst case response time analysis method based on the self-suspending segmentation model. Task τ i In the task set priority assignment ρ={ρ1,ρ2,...,ρ N}Response time under By accumulating the response time of each segment Calculate, specifically is the resource blocking time of the mth accelerator fragment.

[0167] 3.2) Based on the resource allocation model mapping in step 2 and the response time analysis in step 3.1, the final optimization problem is constructed as follows:

[0168]

[0169] l i,j ∈{0,1}#(2e)

[0170] ρ i,j ∈{1,2,…,N}#(2f)

[0171]

[0172] The decision variable is the resource allocation type l i,j , task set task priority assignment ρ i (larger values ​​indicate higher priority) and the complexity of the intelligent controller neural network Candidates for different intelligent controller neural network complexity Generate in advance, where Ξ is the number of candidate complexities. The optimization objective (2a) is the overall control performance of N control tasks, w i is the specified importance of control loop i. Inequality (2b) is the constraint on the worst-case response time of each control task under the interference of other tasks, with the deadline D i (2c) and (2d) represent the CPU and accelerator resource utilization constraints, respectively. represents the total estimated worst-case execution time of the neural network on the CPU / accelerator, λ C ,λ A is a safety margin for estimating the worst-case execution time. Because the empirical worst-case execution time test may not necessarily cover the worst-case execution time in the actual experiment, λ can be adjusted according to the fluctuation of execution time or the strictness of required performance. C ,λ A The value of .

[0173] 3.3) Establish a solution method based on reinforcement learning to solve the optimization problem in step 3.2). Specifically, in the classical deep reinforcement learning framework, the optimization state variable is defined as The corresponding control performance is Optimize the action a at a certain step t t Defined as changing the optimization variable at step t The value of a certain dimension in the next step is obtained Then action a t The reward is In this way, the above settings are brought into the classic deep reinforcement learning framework to solve the problem.

[0174] Step 4: Deploy the optimization solution on the heterogeneous computing platform.

[0175] 4.1) According to the optimization results of step 3, select the complexity c of the deep neural network intelligent controller i ;

[0176] 4.2) Implement heterogeneous computing resource allocation scheme for each controller according to the optimization result of step 3 {l i,j};

[0177] 4.3) Based on the optimization results of step 3, the real-time scheduler assigns priority ρ to the task set.

[0178] In another embodiment, the present invention further provides a system for co-designing complexity and scheduling of intelligent controllers based on deep neural networks on a heterogeneous computing platform, the system comprising the following modules:

[0179] Module M1: Establish the control performance prediction function of the intelligent controller, describe the relationship between the control performance and the complexity of the intelligent controller, the end-to-end response time and the physical characteristics of the controlled object, and fit the relationship through the machine learning regression model;

[0180] Module M1.1: Construct the control system model of the physical controlled object;

[0181] Module M1.2: Study the control performance of intelligent controllers and the influencing variables on the control performance The influence of the intelligent controller complexity c i , end-to-end response time R i 、Physical characteristic parameters of the controlled objectθ i , i∈{1,2,…,N} is the control loop number, and N is the number of control system loops. Specifically, the complexity of the intelligent controller algorithm c i is the complexity of the intelligent controller neural network, including depth and various structures, and the end-to-end response time R i is the time taken by the intelligent controller task from the start to the completion of the calculation, and the parameter θ i is the physical dynamic model parameter of the controlled object in a specific control instance;

[0182] Module M1.3: Construct control performance prediction function based on parameters in module M1.2

[0183] Module M1.4: Fixed physical characteristic parameters θ of the controlled object i After that, traverse the complexity c of the intelligent controller i , end-to-end response time R i All possible values ​​within the value range constitute the machine learning input set, and its control performance is tested through simulation experiments Construct the output set;

[0184] Module M1.5: After the input and output data are collected, they are divided into training sets and test sets;

[0185] Module M1.6: Train regression models of the relationship between output variables and input variables;

[0186] Module M1.7: Compare the accuracy of various models and parameters for the test set, weigh and analyze the specific input-output variable combinations, and select the optimal regression model;

[0187] Module M1.8: Express the output of the optimal regression model as an output variable of the control performance index to form a performance analysis statistical model;

[0188] Module M1.9: Build and implement the optimal machine learning regression model based on the performance analysis statistical model trained in Module M1.7.

[0189] Module M2: Establishing the execution model of the intelligent controller on heterogeneous computing devices, including building the mapping relationship between the intelligent control algorithm and the computing task execution model on the heterogeneous platform, seamlessly combining the controller design with heterogeneous resource allocation;

[0190] Module M2.1: Intelligent controller task τ based on deep neural network i A segmentation model based on self-suspension is proposed. Specifically, self-suspension refers to the process in a computer operating system that pauses its execution and saves its state to external memory when the process is waiting for time-consuming operations such as I / O device response or task transfer to accelerator (such as GPU) to release memory resources and improve the overall efficiency of the system. When the conditions are met, the process is resumed and continued to execute. The self-suspension segmentation model is to treat the part of the intelligent controller task that is transferred to the accelerator as the accelerator segment containing the self-suspension operation, and the part of the task on the CPU as the CPU segment;

[0191] Module M2.2: Intelligent controller task τ based on deep neural network i A hierarchical resource scheduling model for neural networks is proposed. Specifically, the model uses each layer of the deep neural network as the basic unit of resource allocation to form a hierarchical scheduling model;

[0192] Module M2.3: Perform offline running time test on each network layer of the intelligent controller, and obtain the empirical worst-case execution time through repeated tests, that is, the longest execution time recorded during the test;

[0193] Module M2.4: Mapping the hierarchical resource scheduling model to the self-suspending segmentation model.

[0194] Module M3: Establish an end-to-end optimization method for controller complexity and scheduling to determine the optimal controller complexity, heterogeneous resource allocation scheme and real-time scheduling strategy, providing an end-to-end solution for intelligent controller design and deployment on heterogeneous computing platforms;

[0195] Module M3.1: Construct the worst-case response time analysis method based on the self-suspending segmentation model;

[0196] Module M3.2: Construct the final optimization problem based on the resource allocation model mapping of module M2 and the response time analysis of module M3.1;

[0197] Module M3.3: Establish a reinforcement learning-based solution method to solve the optimization problem in module M3.2.

[0198] Module M4: Deployment of optimization solutions on heterogeneous computing platforms.

[0199] Module M4.1: Select the complexity c of the deep neural network intelligent controller based on the optimization results of module M3 i ;

[0200] Module M4.2: Implement heterogeneous computing resource allocation scheme for each controller according to the optimization results of module M3 {l i,j};

[0201] Module M4.3: The real-time scheduler assigns priority ρ to the task set based on the optimization results of module M3.

[0202] In addition to implementing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, the system and its various devices, modules, and units provided by the present invention can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; the devices, modules, and units for implementing various functions can also be considered as both software modules for implementing the method and structures within the hardware component.

[0203] The present invention is not limited to the above specific implementations, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments may be combined with each other at will.

[0204] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for co-designing complexity and scheduling of intelligent controllers, characterized in that: This includes the selection of complexity and real-time scheduling of intelligent controller algorithms based on deep neural networks.

2. The intelligent controller complexity and scheduling collaborative design method according to claim 1, characterized in that: Includes a fine-grained execution model of an intelligent controller on heterogeneous computing devices based on a self-suspending segmentation model.

3. The intelligent controller complexity and scheduling collaborative design method according to claim 1, characterized in that: Construct a mapping relationship between each layer of the intelligent control algorithm and the sub-thread execution model of the computing task on the heterogeneous platform.

4. The intelligent controller complexity and scheduling collaborative design method according to claim 1, characterized in that: Includes data-driven control performance prediction functions that characterize the relationship between closed-loop control performance and the complexity and accuracy of the intelligent controller and the end-to-end response time.

5. The intelligent controller complexity and scheduling collaborative design method according to claim 1, characterized in that: The following steps are involved: Step 1: Establish a control performance prediction function for the intelligent controller, describe the relationship between the control performance and the complexity of the intelligent controller algorithm, the end-to-end response time, and the physical characteristics of the controlled object, and fit the relationship through a machine learning regression model; Step 2: Establish an execution model of the intelligent controller on the heterogeneous computing device and a resource allocation model mapping, including building a mapping relationship between the intelligent control algorithm and the computing task execution model on the heterogeneous platform; Step 3: Establish an end-to-end optimization method for controller complexity and scheduling to determine the optimal controller complexity, heterogeneous resource allocation scheme, and real-time scheduling strategy; Step 4: Deploy the optimization solution on the heterogeneous computing platform.

6. The intelligent controller complexity and scheduling collaborative design method according to claim 5, characterized in that: The step 1 comprises the following steps: Step 1.1, construct the control system model of the physical controlled object; Step 1.2: Control performance of intelligent controller The relevant influencing variables include the complexity of the intelligent controller algorithm c i , end-to-end response time R i 、Physical characteristic parameters of the controlled objectθ i , i∈{1, 2, ..., N} is the control loop number, and N is the number of control system loops; Step 1.3: Construct control performance prediction function Step 1.4: Fix the physical characteristic parameter θ of the controlled object i After that, the complexity of the intelligent controller algorithm c is traversed i , end-to-end response time R i All possible values ​​within the value range constitute the machine learning input set, and the control performance is tested through simulation experiments Construct the output set; Step 1.5: After the input set and output set data are collected, they are divided into training set and test set; Step 1.6: Train a regression model of the relationship between the output variable and the input variable; Step 1.7: For the test set, compare the accuracy of various models and parameters, weigh and analyze the specific input-output variable combination, and select the optimal regression model; Step 1.8, the output of the optimal regression model is expressed as an output variable form of the control performance index to form a performance analysis statistical model; Step 1.9: Build and implement the optimal machine learning regression model based on the trained performance analysis statistical model.

7. The intelligent controller complexity and scheduling collaborative design method according to claim 5, characterized in that: The step 2 comprises the following steps: Step 2.1: Intelligent controller task τ based on deep neural network i The self-suspended segmentation model; Step 2.2: Intelligent controller task τ based on deep neural network i Hierarchical resource scheduling model for deep neural networks; Step 2.3, perform an offline running time test on each network layer of the intelligent controller to obtain the empirical worst case execution time; Step 2.4: Map the hierarchical resource scheduling model to the self-suspending segmentation model.

8. The intelligent controller complexity and scheduling collaborative design method according to claim 5, characterized in that: The step 3 comprises the following steps: Step 3.1, construct the worst case response time analysis method based on the self-suspending segmentation model; Step 3.2: Construct the final optimization problem based on resource allocation model mapping and response time analysis; Step 3.3: Solve the optimization problem based on reinforcement learning.

9. The intelligent controller complexity and scheduling collaborative design method according to claim 5, characterized in that: The step 4 comprises the following steps: Step 4.1: According to the optimization result of step 3, select the complexity c of the deep neural network intelligent controller. i ; Step 4.2: Based on the optimization result of step 3, a heterogeneous computing resource allocation scheme is implemented for each controller. i，j }; Step 4.3: According to the optimization result of step 3, the real-time scheduler assigns a priority ρ to the task set.

10. An intelligent controller complexity and scheduling collaborative design system, characterized in that: Includes the following modules: Module M1 is configured to establish a control performance prediction function of the intelligent controller, describe the relationship between the control performance and the complexity of the intelligent controller, the end-to-end response time and the physical characteristics of the controlled object, and fit the relationship through a machine learning regression model; Module M2 is configured to establish an execution model of the intelligent controller on a heterogeneous computing device, including building a mapping relationship between an intelligent control algorithm and a computing task execution model on a heterogeneous platform; Module M3 is configured to establish an end-to-end optimization method for controller complexity and scheduling, and determine the optimal controller complexity, heterogeneous resource allocation scheme and real-time scheduling strategy; Module M4 is configured to be able to deploy the optimization solution on heterogeneous computing platforms.

Images