Grinding and polishing quality digital twinning prediction method and system based on machine learning

Abstract

The invention discloses a grinding and polishing quality digital twinning prediction method and system based on machine learning, and relates to industrial big data, and the method comprises the steps: obtaining surface parameters of a current workpiece before grinding and polishing, and part material characteristics; utilizing a transfer learning model to execute training learning according to a prior task associated with the obtained surface parameters and material characteristics; inputting the process parameters of the current workpiece in the grinding and polishing process into the trained transfer learning model to predict the grinding and polishing quality of the current workpiece under the grinding and polishing process parameters, and presenting the grinding and polishing quality on the basis of a pre-constructed digital twinning model; and the grinding and polishing quality under the grinding and polishing process parameters is measured, and the transfer learning model is updated and trained. Training is carried out according to part characteristics and grinding and polishing process parameters, the trained algorithm model can output the overall dimension and the surface roughness of the ground and polished part according to the actual material characteristics of the workpiece, and the overall dimension and the surface roughness of the ground and polished workpiece can be predicted online in real time and digitally displayed.

Description

Technical Field

[0001] The present application relates to the field of industrial big data, and in particular to a method and system for digital twin prediction of grinding and polishing quality based on machine learning. Background Art

[0002] As one of the important means to replace manual grinding and polishing, robotic grinding and polishing has been widely used in the field of grinding and polishing. The prediction of the surface state after grinding and polishing is the guarantee of the quality of robot intelligent grinding and polishing. In the existing public patent: Robotic grinding and polishing removal prediction method and equipment based on prior knowledge model (patent number: CN 113159121 B), a method for predicting grinding and polishing removal is disclosed. This method uses training samples and material removal empirical models to train the prior model to obtain a practical prior knowledge model. The prior knowledge model can predict the robot grinding and polishing removal for the input grinding and polishing process parameters, and has good robustness and high grinding and polishing material removal prediction accuracy. Public patent: A robot grinding and polishing prediction control method based on deep belief network (patent number: CN 110315552 A) discloses a robot grinding and polishing prediction control method. This method establishes a dynamic model of the robot grinding and polishing system, and combines the deep belief network proposed in the application to obtain a complete robot grinding and polishing prediction control model. By analyzing the system information at the current moment, the model can predict the robot's grinding and polishing force at a certain moment in the future, and implement predictive control of the robot's grinding and polishing process by adjusting the control voltage. This method alleviates the real-time requirements of the grinding and polishing control system, can quickly control the establishment of the model, achieve more accurate grinding and polishing control, and control the grinding and polishing deformation amplitude to avoid grinding and polishing vibration.

[0003] Contour dimensions and surface roughness are key factors affecting the performance of parts such as fit stability, wear resistance, and fatigue strength. Grinding and polishing are usually required to ensure the final surface quality and surface state of the parts. The existing workpiece material properties, contour features, and grinding and polishing requirements are different, the material removal mechanism is complex, and the process parameters are numerous and interact with each other. When debugging the grinding and polishing process, it is necessary to manually select the grinding and polishing process parameters, including the selection of grinding and polishing tools and consumables, tool speed, grinding and polishing pressure, and grinding and polishing time, and move the workpiece to a special station, use three-coordinate measuring machines, roughness meters and other equipment or instruments to perform corresponding grinding and polishing quality inspections, analyze the grinding and polishing results, and adjust the process parameters repeatedly based on experience to achieve the required grinding and polishing quality. This method of exploring the process by trial and error and adjustment based on manual experience requires a large amount of work. In the grinding and polishing of batch parts, since the allowance to be processed for each part is different, the processing parameters and processing results of each part are different, but due to the limitation of cycle time, the processing quality of each part cannot be inspected, and unqualified parts after grinding and polishing often directly lead to product scrapping. Summary of the invention

[0004] The embodiments of the present application provide a method and system for predicting grinding and polishing quality digital twins based on machine learning. The method and system are trained according to a sample set consisting of part characteristics, grinding and polishing process parameters and corresponding grinding and polishing quality. The trained algorithm model can use the actual material properties, contour features and grinding and polishing process parameters of the workpiece as input, and output the contour dimensions and surface roughness of the polished parts. The method can predict and digitally display the contour dimensions and surface roughness of the workpiece after grinding and polishing in real time online.

[0005] The present application embodiment provides a grinding and polishing quality digital twin prediction method based on machine learning, including:

[0006] Obtain the surface parameters of the current workpiece before grinding and polishing, as well as the material characteristics of the part;

[0007] Using a transfer learning model to perform training learning based on a prior task associated with the acquired surface parameters and material properties;

[0008] Input the process parameters of the current workpiece grinding and polishing process into the trained transfer learning model to predict the grinding and polishing quality of the current workpiece under the grinding and polishing process parameters, and present it based on the pre-built digital twin model;

[0009] The grinding and polishing quality under the grinding and polishing process parameters is measured, and the transfer learning model is updated and trained.

[0010] Optionally, the transfer learning model is a machine learning model based on a deep deterministic policy gradient (DDPG) algorithm, and the transfer learning model specifically performs the following training steps:

[0011] Initialize the current Actor network and the current Critic network, and copy the parameters of the current Actor network and the current Critic network to the target Actor network and the target Critic network respectively;

[0012] Input the workpiece state into the current Actor network to use the previous Actor network to output the grinding and polishing process parameters, apply the output grinding and polishing process parameters to the workpiece, and feed back the workpiece state and reward at the next moment to construct a four-tuple (S, A, R, S') according to the workpiece state, grinding and polishing process parameters, the workpiece state at the next moment and the reward to generate a training sample set;

[0013] From the samples in the training samples, the workpiece state and polishing process parameters of the samples are input into the current Critic network to use the current Critic network to output Q(S,A), and the workpiece state S' at the next moment is input into the target Actor network to obtain the polishing process parameter A' at the next moment. The workpiece state S' at the next moment and the polishing process parameter A' at the next moment are input into the target Critic network to obtain Q(S', A'), then the target value Q':

[0014] Q'=R+γQ(S',A')

[0015] Use the gradient descent method to update all the parameters θ of the current Critic network to minimize L(θ) and complete the update of the current Critic network;

[0016] L(θ)=(Q(S,A)-Q') 2

[0017] Calculate the grinding and polishing process parameters A under the current state S of the workpiece through the Actor network new , given Q(S,A) in the current Critic network new ), and update the current Actor network parameter ω by the gradient ascent method, so that the output Q(S,A new ) to complete the current Actor network parameter update:

[0018]

[0019] In the formula

[0020]

[0021] Introduce the learning rate τ∈(0,1), take the weighted average of the old target network parameters and the new corresponding network parameters, and assign them to the target network:

[0022] Target Actor Network:

[0023] ω'=τω+(1-τ)ω'

[0024] Target Critic Network:

[0025] θ'=τθ+(1-τ)θ'

[0026] Repeatedly input the artifact state into the current Actor network to perform training.

[0027] Optionally, performing training learning using the transfer learning model according to a prior task associated with the acquired surface parameters and material properties further includes:

[0028] Based on policy transfer (PTF), before transfer learning, a set of source policy sets similar to the task set is pre-set s ={π1,π2,π3,…π n}, and initialize a set of options O = {ο1,ο2,ο3,…,ο n} and option-value network with random parameters;

[0029] During the operation of PTF, a set of corresponding workpiece change parameters is output under the input of the workpiece initial state, the reward r is obtained by interacting with the digital twin model and the next state is converted, and an option O is selected according to the option strategy and the termination function. i ;

[0030] In the case of PTF running updates, a supplementary loss function is introduced to mimic the source strategy π in the option i To transfer knowledge, and adaptively adjust the weighting factor f(β o , t), where the option-value network and option i The termination function is updated by PTF based on the knowledge, and PTF based on the option. i The termination function determines whether to reuse strategy π i , if option strategy π is not reused i Another strategy π will be chosen j For reuse.

[0031] Optionally, performing training learning using the transfer learning model according to a prior task associated with the acquired surface parameters and material properties further includes:

[0032] In the process of transfer learning, multiple source strategies are used to train new strategies. The source strategy π i Generate a trajectory x, which contains a state sequence (x t ) t≥0 , the new strategy is π s (θ), the loss function Loss corresponding to each state sequence at each time step t satisfies:

[0033] Loss = f(β o ,t)H(π t (A|x t )∥π s (A|x t ,θ))

[0034] Where: H(π t (A|x t )∥π s (A|x t ,θ)) is the cross entropy loss function;

[0035] An adaptive heuristic method is introduced to adjust the weighting factor of the cross entropy loss function:

[0036] f(β o ,t)=f(t)(1-β(S t ,O|θ β ))

[0037] Where f(t) is a decay function. When the value of the termination function of option 0 increases, it means that the performance of option 0 is not the best among all options based on existing experience, and the weighting factor f(β o ,t).

[0038] Optionally, the acquired surface parameters of the current workpiece before polishing include surface profile size and roughness.

[0039] Optionally, the process parameters input into the trained transfer learning model include: grinding and polishing tool rotation speed, grinding and polishing tool grinding ability, grinding and polishing contact force, grinding and polishing moving speed, grinding and polishing position, and grinding and polishing posture.

[0040] Optionally, it also includes pre-building a digital twin model to display the surface quality state parameters of the workpiece geometric model, wherein the surface quality state parameters include the current grinding and polishing processing area, contour size parameters, and surface roughness parameters.

[0041] An embodiment of the present application also proposes a digital twin prediction system for grinding and polishing quality based on machine learning, including a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the digital twin prediction method for grinding and polishing quality based on machine learning are implemented as described above.

[0042] The method of the embodiment of the present application is trained based on a sample set consisting of part characteristics, grinding and polishing process parameters and corresponding grinding and polishing qualities. The trained algorithm model can take the actual material properties, contour features and grinding and polishing process parameters of the workpiece as input, and output the contour size and surface roughness of the polished part. It can predict and digitally display the contour size and surface roughness of the workpiece after grinding and polishing in real time online.

[0043] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0045] Figure 1 The basic process diagram of the digital twin prediction method for grinding and polishing quality in the embodiment of the present application is shown;

[0046] Figure 2 The system architecture of the digital twin prediction method for grinding and polishing quality according to the embodiment of the present application is shown in FIG.

[0047] Figure 3 This is a schematic diagram of the deep deterministic policy gradient (DDPG) algorithm flow of the digital twin prediction method for grinding and polishing quality in the embodiment of the present application;

[0048] Figure 4 The strategy migration process of the digital twin prediction method for grinding and polishing quality in the embodiment of the present application is schematically shown;

[0049] Figure 5 This is a schematic diagram of the option module process of the digital twin prediction method for grinding and polishing quality in the embodiment of the present application. DETAILED DESCRIPTION

[0050] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0051] Digital twins are interactive mappings between digital models and physical entities built in virtual space, that is, digital expressions of real-world objects or systems, used to describe the operation trajectory of physical entities throughout their life cycle. They use data such as physical models, sensor updates, and operation history to integrate multi-disciplinary, multi-physical quantities, multi-scales, and multi-probabilities for simulation. Digital twin technology can predict system behavior and performance through historical data and simulation results, and monitor system status through real-time synchronization of sensor data and the real world. Therefore, it is widely used in the field of intelligent manufacturing. However, at this stage, the application of digital twin technology is mainly concentrated in production process simulation, equipment motion process simulation, and other aspects.

[0052] In order to solve the problem that the contour size and surface roughness of the workpiece surface cannot be predicted and digitally displayed during the grinding and polishing process, this application proposes a digital twin prediction method for grinding and polishing quality based on a machine learning algorithm, which is used to predict and digitally display the contour size and surface roughness of the workpiece after grinding and polishing in real time online. The system has a built-in deep reinforcement learning algorithm model, which can be trained based on a sample set consisting of part characteristics, grinding and polishing process parameters and corresponding grinding and polishing quality. The trained algorithm model can use the actual material properties, contour features and grinding and polishing process parameters of the workpiece as input, and output the contour size and surface roughness of the polished part. Specific embodiments of the present application provide a digital twin prediction method for grinding and polishing quality based on machine learning, such as Figure 1 As shown, the following steps are included:

[0053] In step S101, the surface parameters of the current workpiece before grinding and polishing, as well as the material properties of the part, are obtained. In some embodiments, the surface parameters of the current workpiece before grinding and polishing obtained include surface profile size and roughness. Specifically, for example, this can be achieved through a data acquisition subsystem, which uses the data acquisition subsystem to collect the surface profile size and roughness measurement of the workpiece before grinding and polishing, and collect the material properties of the part.

[0054] In step S102, training learning is performed using a transfer learning model according to a previous task associated with the acquired surface parameters and material properties.

[0055] In step S103, the process parameters of the current workpiece grinding and polishing process are input into the trained transfer learning model to predict the grinding and polishing quality of the current workpiece under the grinding and polishing process parameters, and are presented based on the pre-built digital twin model. In some examples, the process parameters input into the trained transfer learning model include: grinding and polishing tool rotation speed, grinding and polishing tool grinding ability, grinding and polishing contact force, grinding and polishing movement speed, grinding and polishing position, and grinding and polishing posture. The relevant parameters can be obtained through the grinding and polishing process parameter subsystem. In some embodiments, it also includes pre-building a digital twin model to display the surface quality state parameters of the workpiece geometric model, wherein the surface quality state parameters include the current grinding and polishing processing area, contour size parameters, and surface roughness parameters.

[0056] In step S104, the polishing quality under the polishing process parameters is measured, and the transfer learning model is updated and trained. Figure 2 As shown, the transfer learning model can be configured in the machine learning subsystem, and the machine learning subsystem is connected to the data acquisition subsystem, the grinding and polishing process parameter subsystem, and the digital twin subsystem respectively.

[0057] The data acquisition subsystem transmits the workpiece initial state data to the machine learning subsystem and the digital twin subsystem, and the grinding and polishing process parameter subsystem inputs the grinding and polishing process data to the machine learning subsystem. During the grinding and polishing process, the machine learning subsystem transmits the grinding and polishing quality prediction data to the digital twin subsystem in real time, and the digital twin subsystem performs digital display of the grinding and polishing quality.

[0058] The method of the embodiment of the present application is trained based on a sample set consisting of part characteristics, grinding and polishing process parameters and corresponding grinding and polishing qualities. The trained algorithm model can take the actual material properties, contour features and grinding and polishing process parameters of the workpiece as input, and output the contour size and surface roughness of the polished part. It can predict and digitally display the contour size and surface roughness of the workpiece after grinding and polishing in real time online.

[0059] This algorithm is suitable for problems with continuous state space and has three major advantages in continuous action space:

[0060] 1. DDPG uses a deterministic strategy to directly output continuous action values, which is easier to optimize than a probabilistic strategy; 2. It uses a target network to stabilize training, reduce oscillation and divergence problems during training, and converge faster; 3. It uses a policy gradient method to optimize more efficiently in dealing with highly discontinuous and highly perturbed spatial actions. In some embodiments, the transfer learning model is a machine learning model based on the deep deterministic policy gradient (DDPG) algorithm, such as Figure 3 As shown, the transfer learning model specifically performs the following training steps:

[0061] Initialize the current Actor network and the current Critic network, and copy the parameters of the current Actor network and the current Critic network to the target Actor network and the target Critic network respectively, completing the construction of the four networks;

[0062] The workpiece state S is input into the current Actor network to use the previous Actor network to output the grinding and polishing process parameters A, apply the output grinding and polishing process parameters A to the workpiece, and feed back the workpiece state S' and reward R at the next moment, so as to construct a four-tuple (S, A, R, S') according to the workpiece state, grinding and polishing process parameters, the workpiece state S' at the next moment and reward R, so as to generate a training sample set;

[0063] From the samples in the training samples, the workpiece state and polishing process parameters of the samples are input into the current Critic network to use the current Critic network to output Q(S,A), and the workpiece state S' at the next moment is input into the target Actor network to obtain the polishing process parameter A' at the next moment. The workpiece state S' at the next moment and the polishing process parameter A' at the next moment are input into the target Critic network to obtain Q(S', A'), then the target value Q':

[0064] Q'=R+γQ(S',A') (1)

[0065] Use the gradient descent method to update all the parameters θ of the current Critic network to minimize L(θ) and complete the update of the current Critic network;

[0066] L(θ)=(Q(S,A)-Q') 2 (2)

[0067] Calculate the grinding and polishing process parameters A under the current state S of the workpiece through the Actor network new , given Q(S,A) in the current Critic network new ), and update the current Actor network parameter ω by the gradient ascent method, so that the output Q(S,A new ) to complete the current Actor network parameter update:

[0068]

[0069] In the formula

[0070]

[0071] Using the soft update method, the learning rate τ∈(0, 1) is introduced to make a weighted average of the old target network parameters and the new corresponding network parameters and assign them to the target network:

[0072] Target Actor Network:

[0073] ω'=τω+(1-τ)ω' (5)

[0074] Target Critic Network:

[0075] θ'=τθ+(1-τ)θ' (6)

[0076] Repeatedly input the artifact state into the current Actor network, perform training, and loop until the termination condition is met.

[0077] Due to the limited training samples and computing power, there are many types of workpieces to be ground and polished, and there are still differences in size, features, materials, etc. between the workpieces. The deep reinforcement learning algorithm is trained and optimizes model parameters through a large amount of sample data, but it can only achieve good performance on specific tasks. The embodiment of the present application proposes a transfer learning algorithm based on policy transfer (PTF) combined with deep reinforcement learning (DDPG), which uses appropriate source policy knowledge to directly optimize the target policy, which can improve the efficiency of extracting effective knowledge from a small number of samples, avoid negative transfer, and quickly improve the adaptability and generalization ability of the algorithm. Figure 4As shown, the learning framework of the embodiment of the present application includes an agent module (based on DDPG) and an option module. The agent module learns the target strategy under the guidance of the option module, and the option module selects the source strategy closest to the agent module at the current moment. In some embodiments, using the transfer learning model to perform training learning based on the previous task associated with the acquired surface parameters and material properties also includes:

[0078] Based on policy transfer (PTF), before transfer learning, a set of source policy sets similar to the task set is pre-set s ={π1,π2,π3,…π n}, and initialize a set of options O = {ο1,ο2,ο3,…,ο n} and option-value networks with random parameters.

[0079] During the operation of PTF, a set of corresponding workpiece change parameters is output under the input of the workpiece initial state, the reward r is obtained by interacting with the digital twin model and the next state is converted, and an option O is selected according to the option strategy and the termination function. i .

[0080] In the case of PTF running updates, a supplementary loss function is introduced to mimic the source strategy π in the option i To transfer knowledge, and adaptively adjust the weighting factor f(β o , t), where the option-value network and option i The termination function is updated by PTF based on the knowledge, and PTF based on the option. i The termination function determines whether to reuse strategy π i , if option strategy π is not reused i Another strategy π will be chosen j The policy reuse option call-return execution mechanism of the embodiment of the present application effectively improves the utilization rate of useful information in the source policy and avoids negative migration.

[0081] In some embodiments, performing training learning based on a prior task associated with the acquired surface parameters and material properties using a transfer learning model further includes:

[0082] In the process of transfer learning, multiple source strategies are used to train new strategies. The source strategy π i Generate a trajectory x, which contains a state sequence (x t ) t≥0 , the new strategy is π s (θ), the loss function Loss corresponding to each state sequence at each time step t satisfies:

[0083] Loss = f(β o,t)H(π t (A|x t )∥π s (A|x t ,θ))

[0084] Where: H(π t (A|x t )∥π s (A|x t ,θ)) is the cross entropy loss function;

[0085] An adaptive heuristic method is introduced to adjust the weighting factor of the cross entropy loss function:

[0086] f(β o ,t)=f(t)(1-β(S t ,O|θ β ))

[0087] Where f(t) is a decay function. When the value of the termination function of option 0 increases, it means that the performance of option 0 is not the best among all options based on existing experience, and the weighting factor f(β o ,t), and vice versa. f(t) controls the slow decline of the new strategy generated from the source strategy. In the early stage, the source strategy is used to speed up the exploration process. The utilization value of the source strategy gradually decreases with the progress of learning, and resources are gradually transferred to its own strategy.

[0088] When debugging the grinding and polishing process, the present application requires manual repeated adjustment of the grinding and polishing process parameters based on experience to achieve the required grinding and polishing quality. This method of exploring the process by relying on manual experience, trial and error, and adjustment requires a large amount of work. In addition, due to the limitation of cycle time, it is impossible to test the processing quality of each part. Failure of parts inspection after grinding and polishing often directly leads to the problem of product scrapping. The method proposed in the embodiment of the present application can use the current state of the part and the real-time grinding and polishing process parameters as input, and online predict and digitally display the contour size and surface roughness of the workpiece after grinding and polishing, thereby reducing the time for exploring process parameters and testing grinding and polishing quality.

[0089] The method of this application combines deep reinforcement learning with transfer learning, and uses appropriate source strategy knowledge to directly optimize the target strategy, which can improve the efficiency of extracting effective knowledge from a small number of samples, avoid negative transfer, and quickly improve the adaptability and generalization ability of the algorithm. When facing parts with different appearance features and contour dimensions, it is only necessary to apply the trained deep reinforcement learning algorithm model to a small number of target data sets for parameter fine-tuning, which can quickly generalize and achieve better grinding and polishing quality prediction results.

[0090] An embodiment of the present application also proposes a digital twin prediction system for grinding and polishing quality based on machine learning, including a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the digital twin prediction method for grinding and polishing quality based on machine learning are implemented as described above.

[0091] It should be noted that in the various embodiments of the present application, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0092] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0093] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0094] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A digital twin prediction method for grinding and polishing quality based on machine learning, characterized in that: include: Obtain the surface parameters of the current workpiece before grinding and polishing, as well as the material characteristics of the part; Using a transfer learning model to perform training learning based on a prior task associated with the acquired surface parameters and material properties; Input the process parameters of the current workpiece grinding and polishing process into the trained transfer learning model to predict the grinding and polishing quality of the current workpiece under the grinding and polishing process parameters, and present it based on the pre-built digital twin model; The grinding and polishing quality under the grinding and polishing process parameters is measured, and the transfer learning model is updated and trained.

2. The method for predicting grinding and polishing quality digital twin based on machine learning according to claim 1, characterized in that: The transfer learning model is a machine learning model based on the deep deterministic policy gradient (DDPG) algorithm. The transfer learning model specifically performs the following training steps: Initialize the current Actor network and the current Critic network, and copy the parameters of the current Actor network and the current Critic network to the target Actor network and the target Critic network respectively; Input the workpiece state into the current Actor network to use the previous Actor network to output the grinding and polishing process parameters, apply the output grinding and polishing process parameters to the workpiece, and feed back the workpiece state and reward at the next moment to construct a four-tuple (S, A, R, S') according to the workpiece state, grinding and polishing process parameters, the workpiece state at the next moment and the reward to generate a training sample set; From the samples in the training samples, the workpiece state and polishing process parameters of the samples are input into the current Critic network to use the current Critic network to output Q(S,A), and the workpiece state S' at the next moment is input into the target Actor network to obtain the polishing process parameter A' at the next moment. The workpiece state S' at the next moment and the polishing process parameter A' at the next moment are input into the target Critic network to obtain Q(S', A'), then the target value Q': Q'=R+γQ(S',A') Use the gradient descent method to update all the parameters θ of the current Critic network to minimize L(θ) and complete the update of the current Critic network; L(θ)=(Q(S,A)-Q’) 2 Calculate the grinding and polishing process parameters A under the current state S of the workpiece through the Actor network new , given Q(S,A) in the current Critic network new ), and update the current Actor network parameter ω by the gradient ascent method, so that the output Q(S,A new ) to complete the current Actor network parameter update: ω←ω+α▽ ω Q(S,A),(0≤α≤1) In the formula Introduce the learning rate τ∈(0,1), take the weighted average of the old target network parameters and the new corresponding network parameters, and assign them to the target network: Target Actor Network: ω'=τω+(1-τ)ω' Target Critic Network: θ'=τθ+(1-τ)θ' Repeatedly input the artifact state into the current Actor network to perform training.

3. The method for predicting grinding and polishing quality digital twin based on machine learning according to claim 2, characterized in that: Performing training learning based on a prior task associated with the acquired surface parameters and material properties using a transfer learning model also includes: Based on policy transfer (PTF), before transfer learning, a set of source policy sets similar to the task set is pre-set s ={π1,π2,π3,…π n }, and initialize a set of options O = {ο1,ο2,ο3,…,ο n } and option-value network with random parameters; During the operation of PTF, a set of corresponding workpiece change parameters is output under the input of the workpiece initial state, the reward r is obtained by interacting with the digital twin model and the next state is converted, and an option O is selected according to the option strategy and the termination function. i ; In the case of PTF running updates, a supplementary loss function is introduced to mimic the source strategy π in the option i To transfer knowledge, and adaptively adjust the weighting factor f(β o , t), where the option-value network and option i The termination function is updated by PTF based on the knowledge, and PTF based on the option. i The termination function determines whether to reuse strategy π i , if option strategy π is not reused i Another strategy π will be chosen j For reuse.

4. The method for predicting grinding and polishing quality digital twin based on machine learning according to claim 3 is characterized in that: Performing training learning based on a prior task associated with the acquired surface parameters and material properties using a transfer learning model also includes: In the process of transfer learning, multiple source strategies are used to train new strategies. The source strategy π i Generate a trajectory x, which contains a state sequence (x t ) t≥0 , the new strategy is π s (θ), the loss function Loss corresponding to each state sequence at each time step t satisfies: Loss=f(β o ,t)H(π t (A|x t )∥π s (A|x t ,i)) Where: H(π t (A|x t )∥π s (A|x t ,θ)) is the cross entropy loss function; An adaptive heuristic method is introduced to adjust the weighting factor of the cross entropy loss function: f(b o ,t)=f(t)(1-β(S t ,O|θ β )) Where f(t) is a decay function. When the value of the termination function of option 0 increases, it means that the performance of option 0 is not the best among all options based on existing experience, and the weighting factor f(β o ,t).

5. The method for predicting grinding and polishing quality digital twin based on machine learning according to claim 1, characterized in that: The acquired surface parameters of the current workpiece before grinding and polishing include surface profile size and roughness.

6. The method for predicting grinding and polishing quality digital twin based on machine learning according to claim 1, characterized in that: The process parameters of the transfer learning model input for training include: grinding and polishing tool rotation speed, grinding and polishing tool grinding ability, grinding and polishing contact force, grinding and polishing moving speed, grinding and polishing position, and grinding and polishing posture.

7. The method for predicting grinding and polishing quality digital twin based on machine learning according to claim 1, characterized in that: It also includes pre-building a digital twin model to display the surface quality status parameters of the workpiece geometric model, wherein the surface quality status parameters include the current grinding and polishing processing area, contour size parameters, and surface roughness parameters.

8. A digital twin prediction system for grinding and polishing quality based on machine learning, characterized in that: It includes a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the digital twin prediction method for grinding and polishing quality based on machine learning are implemented as described in any one of claims 1 to 7.

Images