Linear motion unit feedforward control method and system based on physical fusion neural network, and medium

Abstract

The invention relates to a linear motion unit feedforward control method and system based on a physical fusion neural network and a medium, and the method comprises the steps: data collection and preprocessing: collecting input and output time sequence data of a linear motion unit in a closed-loop motion control system, and carrying out the normalization processing of the collected data, so as to eliminate the dimension influence; designing a neural network fusing physical information, and designing a loss function, wherein the loss function comprises data driving loss, a neural network coupling physical layer penalty term and neural network regularization loss; and a feedforward controller is designed, and a feedforward control quantity uff is generated through matrix operation. According to the invention, by introducing physical parameter constraints, the generalization ability and dynamic response performance of the controller are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of precision motion control, and particularly to a feedforward control method, system and medium for a linear motion unit based on a physics-integrated neural network. Background Art

[0002] With the continuous development of industrial automation and intelligent control, high-precision motion control technology plays a crucial role in intelligent manufacturing. As an important part of a precision motion control system, the dynamic performance and trajectory tracking accuracy of a linear motion unit directly affect the production efficiency of equipment and the quality of products. Currently, motion control systems mainly rely on the combination of a feedback controller (such as PID control) and a feedforward controller. However, these two control methods have the following problems in practical applications:

[0003] On the one hand, the feedback controller compensates by detecting the error signal, and the generation of the error inevitably introduces a response delay. In high-speed and high-precision trajectory tracking tasks, simply relying on the feedback loop is difficult to meet the dynamic performance requirements.

[0004] On the other hand, traditional feedforward controllers rely on the accurate physical modeling of dynamic systems, such as the speed and acceleration feedforward control of a ball screw drive system. However, in complex dynamic systems, the existence of friction, transmission errors, and unmodeled dynamic effects makes it difficult to avoid physical modeling errors.

[0005] Although data-driven methods can fit complex systems, they lack physical constraints, resulting in insufficient generalization ability and being difficult to apply in actual industrial environments. With the development of intelligent control, mechanism-data-driven hybrid modeling has always been a popular research direction in the control field.

[0006] In the prior art, the design of feedforward controllers usually adopts the following several methods:

[0007] Feedforward control based on a physical model: By establishing the dynamic model of the system, the required control input is calculated. This method relies on an accurate physical model, and it is difficult to obtain a perfect model in an actual system.

[0008] Feedforward control based on learning: By iterative learning control (ILC) or repetitive control (RC) methods, the historical data of repetitive tasks are used to optimize the control input. Such methods are only applicable to repetitive tasks and have limited generalization ability for non-repetitive trajectories.

[0009] Pure data-driven feedforward control: Completely rely on machine learning methods such as neural networks to establish the inverse model of the system. Such methods require a large amount of training data and have poor generalization ability outside the training data range.

[0010] Therefore, how to design a feedforward controller that can not only effectively describe the nonlinear dynamic characteristics of the system but also achieve real-time compensation has become a difficult problem to overcome in the current motion control field. Summary of the Invention

[0011] The purpose of the present invention is to provide a feedforward control method, system, and medium for a linear motion unit based on a physically integrated neural network, which can improve the generalization ability and dynamic response performance of the controller by introducing physical parameter constraints.

[0012] To achieve the above object, in the first aspect, an embodiment of the present application provides a feedforward control method for a linear motion unit based on a physically integrated neural network, including the following specific steps:

[0013] Data acquisition and preprocessing: Collect the input-output time series data of the linear motion unit in the closed-loop motion control system, including: target trajectory r(t), feedback position y(t), controller output u(t), and normalize the collected data to eliminate the influence of dimensions.

[0014] Design of a neural network integrating physical information: Construct a network structure jointly composed of a physical layer θ phy and a neural network layer C φ , where:

[0015] - Input physical conversion layer: Extract the differential features and sign function of the target trajectory;

[0016] - Hidden layer: Adopt a multi-layer fully connected neural network with the activation function tanh;

[0017] - Physical layer parameters θ phy : Introduce known physical parameters as additional constraint layer parameter features;

[0018] - Output layer: Combine the predicted output of the NNARX inverse model and the predicted output under the action of the physical layer parameters;

[0019] Design of the loss function: Include data-driven loss, neural network coupling physical layer penalty term, and neural network regularization loss;

[0020] Design of the feedforward controller: Store the trained network layer weights and offset parameters as the feedforward control matrix W, and the trained physical layer parameter vector P, receive the target trajectory r(t) in real time, generate the regression vector X(t), and generate the feedforward control quantity u ff .

[0021] In the data acquisition and preprocessing step, the data sampling frequency is the position loop servo frequency, covering the variable-speed reciprocating motion within the typical working range of the system.

[0022] The winning bidder in the design of the neural network integrating physical information. The physical parameter layer extracts physical layer parameters based on the dynamic differential equation of the linear motion unit:

[0023]

[0024] θ phy =[J eq ,B eq ,T c ,k eq T

[0025] The network output form is:

[0026]

[0027] In the design of the loss function, the loss function includes:

[0028] Data-driven loss:

[0029] Neural network coupled physical layer penalty term:

[0030] Neural network regularization loss: L reg =γ∑θ NN 2

[0031] Total loss function: L total =L data +L coupling +L reg ;

[0032] Π1(k) in the neural network coupled physical layer penalty term is the projection operator on the subspace spanned by the column vectors of U1 after the singular value decomposition of the physical layer parameter θ phy after singular value decomposition by

[0033] the projection operator on the subspace spanned by the column vectors of U1

[0034] The levenberg-marquardt least squares optimization algorithm is used for network training, and the regularization parameters λ and γ are set to 0.01 respectively.

[0035] In the design of the feedforward controller, the calculation formula of the feedforward control quantity is: u ff =X(t)W + X(t)P, where W is the network layer parameter and P is the physical layer parameter,

[0036] In the design of the feedforward controller, the regression vector is obtained through the physical transformation layer: T(r(k)) = [δ 2 ​r(k), δr(k), sign(δr(k)), r(k).

[0037] In a second aspect, a feedforward control system for a linear motion unit based on a physically integrated neural network includes:

[0038] Data acquisition and preprocessing module: Collect the input-output time series data of the linear motion unit in the closed-loop motion control system, including: the target trajectory r(t), the feedback position y(t), and the controller output u(t), and perform normalization processing on the collected data to eliminate the influence of dimensions;

[0039] Neural network design module integrating physical information: Construct a network structure jointly composed of a physical layer θ phy and a neural network layer C φ where:

[0040] - Input physical conversion layer: Extract the differential features and sign function of the target trajectory;

[0041] - Hidden layer: Adopt a multi-layer fully connected neural network with the activation function tanh;

[0042] - Physical layer parameter θ phy : Introduce known physical parameters as additional constraint layer parameter features;

[0043] - Output layer: Combine the predicted output of the NNARX inverse model and the predicted output under the action of the physical layer parameters;

[0044] Loss function design module: Include data-driven loss, neural network coupled physical layer penalty term, and neural network regularization loss;

[0045] Feedforward controller design module: Store the trained network layer weights and offset parameters as the feedforward control matrix W, and the trained physical layer parameter vector P, receive the target trajectory r(t) in real time, generate the regression vector X(t), and generate the feedforward control quantity u through matrix operations ff .

[0046] In a third aspect, an embodiment of the present application provides a computer-readable storage medium. When the program code stored in the computer-readable storage medium is executed by a processor, the steps of the above-mentioned feedforward control method for a linear motion unit based on a physically integrated neural network are implemented.

[0047] The beneficial effects of adopting the above embodiments are as follows: By integrating NNARX (Neural Network with Autoregressive Exogenous Input Model) and physical layer parameters for simultaneous prediction and control, the present invention can effectively reduce the number of hidden layers and neurons required in the neural network. Compared with the existing research where the design of the feedforward controller depends on the accurate modeling of the physical system or the lack of generalization ability in pure data-driven methods, this model has better prediction accuracy and robust performance, and is an effective modeling method for describing the nonlinear dynamic characteristics of the system. The trained inverse model can be directly used as the feedforward module of the motion controller, further improving the dynamic response ability of the control system and reducing the following error. Brief Description of the Drawings

[0048] To more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for the embodiments of the present application. It should be understood that the following drawings only show some embodiments of the present application, and thus should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0049] Figure 1 It is a schematic diagram of the mechanical structure of the linear motion unit, showing the basic composition of the linear motion unit including a stepper motor, a ball screw, and a translation stage.

[0050] Figure 2 It is a schematic diagram of the physical fusion neural network structure, showing the network architecture including a physical parameter layer and a neural network layer.

[0051] Figure 3 It is a schematic diagram of the control algorithm framework, showing the combined control structure of the feedforward controller and the feedback controller based on the physical fusion neural network.

[0052] Figure 4 It is a reference input trajectory, showing the variable-speed reciprocating motion trajectory for training and verification. Detailed Embodiments

[0053] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the following further details the present invention in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0054] It should be noted that in the present invention, the descriptions involving "first", "second", etc. are only for descriptive purposes and should not be construed as indicating or implying their relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one such feature. Additionally, the technical solutions between various embodiments can be combined with each other, but it must be based on the ability of those of ordinary skill in the art to implement. When the combination of technical solutions results in contradictions or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0055] Embodiment 1: A feedforward control method for a linear motion unit based on a physical fusion neural network according to the present invention,

[0056] As Figure 1 shown, the linear motion unit of this embodiment includes a servo motor, a ball screw drive mechanism, and a translation stage. The physical fusion neural network structure is as Figure 2 shown, and the control algorithm framework is as Figure 3 shown.

[0057] The specific implementation steps are as follows:

[0058] 1. Data acquisition and preprocessing

[0059] Collect the input-output time series data of the linear motion unit in the closed-loop motion control system, including: the target trajectory r(t), the feedback position y(t), and the controller output u(t).

[0060] The data sampling period T S is consistent with the position loop servo control period and covers the variable-speed reciprocating motion within the typical working range of the system. Normalize the collected data to eliminate the influence of dimensions.

[0061] 2. Design of the neural network integrating physical information.

[0062] Network structure: Consisted of the physical layer θ param and the neural network layer C φ jointly

[0063] Input physical conversion layer: The input features include the differential features (velocity, acceleration) of the target trajectory and the sign function, which are used to capture the change in the friction direction.

[0064] Hidden layer: Adopt a multi-layer fully connected neural network, and the activation function is tanh, which is used to extract the nonlinear dynamic characteristics of the system.

[0065] Physical parameter layer θ param: Introduce known physical parameters (such as mass, friction coefficient, etc.) as additional constraint layer parameter features. Extract physical layer parameters based on the dynamic differential equation of the linear motion unit.

[0066]

[0067] θ phy =[J eq ,B eq ,T c ,k eq T

[0068] Output layer: Combine the predicted output of the NNARX inverse model and the predicted output under the action of the physical layer parameters to generate the inverse model result of the dynamic system.

[0069] Network output: The output of the neural network is the predicted feedforward control quantity:

[0070]

[0071] (3) Loss function design

[0072] 1. Data-driven loss: The mean square error between the predicted input and the actual input:

[0073] 2. Neural network coupled physical layer penalty term: Where C φ (k) represents the output of the neural network layer, and Π1(k) represents the physical layer parameter θ param [[ID= forty-three]]After singular value decomposition:

[0074]

[0075] The projection operator on the subspace spanned by the column vectors of U1,

[0076]

[0077] 3. Neural network regularization loss: Limit the magnitude of the neural network parameters to prevent overfitting: L reg =γ∑ j θ j 2

[0078] Total loss function: L total =L data +L coupling +L reg

[0079] ​Training method: Use the levenberg-marquardt least squares optimization algorithm for network training, and set the regularization parameters λ and γ to 0.1 and 0.01 respectively.

[0080] The dataset is divided into a training set (70%) and a validation set (30%), and the network hyperparameters are adjusted through cross-validation.

[0081] (4) Design of the feedforward controller

[0082] After training is completed, store the network layer weights and bias parameters as the feedforward control matrix W, which serves as the feedforward term of the network layer. And use the trained physical layer parameter vector P as the feedforward term of the physical layer

[0083] Real-time controller implementation: For the trajectory tracking task, the controller generates a regression vector by receiving the target trajectory r(t)

[0084]

[0085] Subsequently, generate the feedforward control quantity through matrix operations

[0086] u ff = X(t)W + X(t)P

[0087] where W is the network layer parameter and P is the physical layer parameter.

[0088] A feedforward control system for a linear motion unit based on a physically fused neural network, including

[0089] Data acquisition and preprocessing module: Collect the input-output time series data of the linear motion unit in the closed-loop motion control system, including: target trajectory r(t), feedback position y(t), controller output u(t), and normalize the collected data to eliminate the influence of dimensions;

[0090] Neural network design module for fusing physical information: Construct a network structure composed of a physical layer θ phy and a neural network layer C φ together, where:

[0091] - Input physical conversion layer: Extract the differential features and sign function of the target trajectory;

[0092] - Hidden layer: Use a multi-layer fully connected neural network with the activation function tanh;

[0093] - Physical layer parameter θ phy : Introduce known physical parameters as additional constraint layer parameter features;

[0094] - Output layer: Combine the predicted output of the NNARX inverse model and the predicted output under the action of the physical layer parameters;

[0095] Loss function design module: including data-driven loss, neural network coupled physical layer penalty term, and neural network regularization loss;

[0096] Feedforward controller design module: stores the trained network layer weights and offset parameters as the feedforward control matrix W, and the trained physical layer parameter vector P, receives the target trajectory r(t) in real time, generates the regression vector X(t), and generates the feedforward control quantity u through matrix operations ff 。

[0097] The embodiment of the present application provides a computer-readable storage medium, and the computer-readable storage medium stores program codes. When the program codes are executed by a processor, the steps of the above-mentioned feedforward control method for a linear motion unit based on a physics-integrated neural network are implemented.

[0098] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0099] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0100] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured product including an instruction device, and the instruction device implements the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0101] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are executed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions for implementing the functions specified in one process or a plurality of processes and / or boxes Figure 1 one process or a plurality of processes and / or boxes Figure 1 steps for implementing the functions specified in one box or a plurality of boxes.

[0102] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0103] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM) and / or non-volatile memory such as read-only memory (ROM) or flash memory (flash RAM). The memory is an example of computer-readable media.

[0104] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can store information by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transitory media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media such as modulated data signals and carrier waves.

[0105] Those skilled in the art can implement the present invention in various variant embodiments without departing from the scope and essence of the present invention. For example, the features of one embodiment can be used in another embodiment to obtain another embodiment. Any modifications, equivalent replacements, and improvements made within the technical concept of the present invention shall fall within the scope of the rights of the present invention.

Claims

1. A feedforward control method for a linear motion unit based on a physical fusion neural network, characterized in that, It includes the following specific steps: Data acquisition and preprocessing: Acquire the input-output time series data of the linear motion unit in the closed-loop motion control system, including: target trajectory r(t), feedback position y(t), controller output u(t), and perform normalization processing on the acquired data to eliminate the influence of dimensions; Neural network design integrating physical information: building a physical layer θ phy and neural network layer C φ The network structure consists of: - Input physical conversion layer: Extract the differential features and sign function of the target trajectory; - Hidden layer: Adopt a multi-layer fully connected neural network with the activation function tanh; - Physical layer parameter θ phy : Introduce known physical parameters as additional constraint layer parameter features; - Output layer: Combine the predicted output of the NNARX inverse model and the predicted output under the action of physical layer parameters; Loss function design: Include data-driven loss, neural network coupled physical layer penalty term, and neural network regularization loss; Feedforward controller design: Store the weights and bias parameters of the trained network layer as the feedforward control matrix W, and the trained physical layer parameter vector P. Receive the target trajectory r(t) in real time, generate the regression vector X(t), and generate the feedforward control quantity u through matrix operations ff .

2. The feedforward control method for a linear motion unit based on a physical fusion neural network according to claim 1, wherein In the data acquisition and preprocessing step, the data sampling frequency is the position loop servo frequency, covering the variable-speed reciprocating motion within the typical working range of the system.

3. A feedforward control method for a linear motion unit based on a physical fusion neural network according to claim 1, characterized in that, In the design of the neural network integrating physical information, the physical parameter layer extracts physical layer parameters based on the dynamic differential equation of the linear motion unit: θ phy = [J eq , B eq , T c , k eq T ​ The network output form is:

4. A feedforward control method for a linear motion unit based on a physical fusion neural network according to claim 1, characterized in that, In the loss function design, the loss function includes: Data-driven loss Neural network coupled physical layer penalty term: Neural network regularization loss: L reg = γ∑θ NN 2 Total loss function: L total = L data + L coupling + L reg ; In the neural network coupled physical layer penalty term, Π1(k) is the physical layer parameter θ phy The projection operator on the subspace spanned by the column vectors of U1 after singular value decomposition Use the levenberg-marquardt least squares optimization algorithm for network training, and set the regularization parameters λ and γ to 0.01 respectively.

5. A feedforward control method for a linear motion unit based on a physical fusion neural network according to claim 1, characterized in that, In the design of the feedforward controller, the calculation formula of the feedforward control quantity is: u ff = X(t)W + X(t)P, where W is the parameter of the network layer and P is the parameter of the physical layer, 6. The feedforward control method for a linear motion unit based on a physically fused neural network according to claim 1, wherein In the design of the feedforward controller, the regression vector is transformed through the physical transformation layer to obtain: T(r(k)) = [δ 2 r(k), δr(k), sign(δr(k)), r(k)].

7. A feedforward control system for a linear motion unit based on a physically fused neural network, characterized in that, Include, Data acquisition and preprocessing module: Acquire the input-output time series data of the linear motion unit in the closed-loop motion control system, including: target trajectory r(t), feedback position y(t), controller output u(t), and perform normalization processing on the acquired data to eliminate the influence of dimensions; Neural Network Design Module Integrating Physical Information: Construct a network structure jointly composed of a physical layer θ phy and a neural network layer C φ wherein: - Input physical conversion layer: Extract the differential features and sign function of the target trajectory; - Hidden layer: Adopt a multi-layer fully connected neural network with the activation function tanh; - Physical layer parameter θ phy : Introduce known physical parameters as additional constraint layer parameter features; - Output layer: Combine the predicted output of the NNARX inverse model and the predicted output under the action of physical layer parameters; Loss function design module: Include data-driven loss, neural network coupled physical layer penalty term, and neural network regularization loss; Feedforward controller design module: Store the weights and bias parameters of the trained network layer as the feedforward control matrix W, and the trained physical layer parameter vector P. Receive the target trajectory r(t) in real time, generate the regression vector X(t), and generate the feedforward control quantity u through matrix operations ff .

8. A computer-readable storage medium, characterized in that, [[ID=

Images