Digital twinning dynamic optimization method and system based on machine learning

Abstract

The invention provides a digital twinning dynamic optimization method and system based on machine learning, and relates to the field of data processing, and the method comprises the steps: obtaining a real-time data flow of a manufacturing environment; acquiring real-time performance data of the digital twin model; acquiring a real-time task requirement of the digital twin model; based on the real-time data flow of the manufacturing environment, the real-time performance data of the digital twinborn model or the real-time task requirement of the digital twinborn model, judging whether to adjust the network structure of the digital twinborn model; when it is judged that the network structure of the digital twin model is adjusted, the network structure of the digital twin model is optimized, parameter updating is conducted on the digital twin model with the optimized structure, and the method has the advantages that automatic optimization of the digital twin model structure is achieved so as to improve the performance of the digital twin model.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a digital twin dynamic optimization method and system based on machine learning. Background Art

[0002] Digital twin technology is a technology that closely connects the physical world with the virtual world based on the integration of physical models, sensor updates, historical and real-time data. Simply put, digital twin technology is to collect various data and build a virtual model of the entity, which can simulate, predict and optimize various situations in the real world in a virtual environment. The application of digital twin technology in the field of intelligent manufacturing includes: production process optimization, digital twin technology can simulate various links in the production process by building a virtual model of the production line, find bottlenecks and problems in the production process, optimize the production process, and improve production efficiency; product quality improvement, digital twin technology can simulate various performance and parameters of the product by building a virtual model of the product, predict possible quality problems of the product, make improvements and optimizations in advance, and improve product quality; equipment maintenance and fault prediction, digital twin technology can collect various data of the equipment, build a virtual model of the equipment, predict possible faults and problems of the equipment, repair and replace in advance, reduce equipment downtime, and reduce production costs; production plan optimization, digital twin technology can simulate various links in the production plan by building a virtual model of the production plan, predict possible problems and risks, optimize the production plan, and improve the accuracy and feasibility of the production plan.

[0003] In the development of digital twin technology, the traditional static network structure cannot effectively cope with real-time data flow and complex and changing task requirements, and has limitations in processing real-time data and making adaptive adjustments to models, such as being unable to accurately predict the future state of the system and requiring a lot of manual intervention in the model adjustment process. These problems limit the application scope and effectiveness of digital twin technology.

[0004] Therefore, it is necessary to provide a digital twin dynamic optimization method and system based on machine learning to realize automatic optimization of the digital twin model structure in order to improve the performance of the digital twin model. Summary of the invention

[0005] The present invention provides a digital twin dynamic optimization method based on machine learning, including: obtaining a real-time data stream of a manufacturing environment; obtaining real-time performance data of a digital twin model; obtaining real-time task requirements of the digital twin model; judging whether to adjust the network structure of the digital twin model based on the real-time data stream of the manufacturing environment, the real-time performance data of the digital twin model or the real-time task requirements of the digital twin model; optimizing the network structure of the digital twin model when it is determined to adjust the network structure of the digital twin model; and updating parameters of the digital twin model after structural optimization.

[0006] Furthermore, based on the real-time data flow of the manufacturing environment, the real-time performance data of the digital twin model or the real-time task requirements of the digital twin model, it is determined whether to adjust the network structure of the digital twin model, including: based on the real-time data flow of the manufacturing environment and the historical prediction data of the digital twin model, determining the prediction accuracy of the digital twin model, when the prediction accuracy of the digital twin model is less than a preset prediction accuracy threshold, determining to adjust the network structure of the digital twin model.

[0007] Furthermore, the network structure of the digital twin model is optimized, including: when the prediction accuracy of the digital twin model is less than a preset prediction accuracy threshold, optimizing the network structure of the digital twin model based on a Q learning algorithm.

[0008] Furthermore, the reward function of the Q learning algorithm is: Among them, Q(s,a) is the expected utility of taking action a in state s, Q(s′,a′) is the expected utility of taking action a′ in state s′, s′ is the new state after taking the action, a′ is the action of optimizing the network structure of the digital twin model in state s′, α is the learning rate, r is the immediate reward, γ is the discount factor of future rewards, and the action of optimizing the network structure of the digital twin model includes changing the number of layers of the neural network layer, changing the number of neurons in the neural network layer, changing the connection between neurons, and changing at least one of the activation functions in the neural network layer.

[0009] Furthermore, based on the real-time data flow of the manufacturing environment, the real-time performance data of the digital twin model or the real-time task requirements of the digital twin model, it is determined whether to adjust the network structure of the digital twin model, including: calculating the data flow similarity between the real-time data flow of the manufacturing environment and the historical data flow of the manufacturing environment; when the data flow similarity is less than a preset data flow similarity threshold, determining to adjust the network structure of the digital twin model.

[0010] Furthermore, the network structure of the digital twin model is optimized, including: when the data flow similarity is less than a preset data flow similarity threshold, the network structure of the digital twin model is optimized based on an evolutionary strategy and a reinforcement learning algorithm.

[0011] Furthermore, based on the real-time data flow of the manufacturing environment, the real-time performance data of the digital twin model or the real-time task requirements of the digital twin model, it is determined whether to adjust the network structure of the digital twin model, including: determining the real-time scores of the digital twin model in multiple target performance indicators based on the real-time performance data of the digital twin model; determining whether the real-time scores of the digital twin model in multiple target performance indicators meet the preset performance indicator requirements, and when it is determined that the real-time scores of the digital twin model in multiple target performance indicators do not meet the preset performance indicator requirements, determining to adjust the network structure of the digital twin model, or determining whether the real-time scores of the digital twin model in multiple target performance indicators meet the task performance requirements corresponding to the real-time task requirements of the digital twin model, and when it is determined that the real-time scores of the digital twin model in multiple target performance indicators do not meet the task performance requirements corresponding to the real-time task requirements of the digital twin model, determining to adjust the network structure of the digital twin model.

[0012] Furthermore, the network structure of the digital twin model is optimized, including: when the real-time scores of the digital twin model in multiple target performance indicators do not meet the preset performance indicator requirements, the network structure of the digital twin model is optimized based on a deep deterministic policy gradient algorithm.

[0013] Furthermore, the network structure of the digital twin model is optimized, including: when it is determined that the real-time scores of the digital twin model in multiple target performance indicators do not meet the task performance requirements corresponding to the real-time task requirements of the digital twin model, using an evolutionary strategy of a genetic algorithm or a particle swarm optimization, based on the real-time data flow of the manufacturing environment and the real-time task requirements of the digital twin model, to optimize the network structure of the digital twin model.

[0014] The present invention provides a digital twin dynamic optimization system based on machine learning, including: an environmental adaptability module, used to obtain the real-time data flow of the manufacturing environment; a performance monitoring module, used to obtain the real-time performance data of the digital twin model; a task analysis module, used to obtain the real-time task requirements of the digital twin model; an optimization judgment module, used to judge whether to adjust the network structure of the digital twin model based on the real-time data flow of the manufacturing environment, the real-time performance data of the digital twin model or the real-time task requirements of the digital twin model; a structure search and optimization module, used to optimize the network structure of the digital twin model when it is determined to adjust the network structure of the digital twin model; a model retraining module, used to update the parameters of the digital twin model after structural optimization.

[0015] The present invention provides a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, and when the instructions are executed by the processor, the processor executes the above-mentioned digital twin dynamic optimization method based on machine learning.

[0016] The present invention provides one or more non-volatile computer-readable storage media containing computer-executable instructions. When the computer-executable instructions are executed by one or more processors, the processors execute the above-mentioned digital twin dynamic optimization method based on machine learning.

[0017] Compared with the prior art, the digital twin dynamic optimization method and system based on machine learning provided by the present invention have at least the following beneficial effects:

[0018] By obtaining the real-time data stream of the manufacturing environment, the real-time performance data of the digital twin model, or the real-time task requirements of the digital twin model, it is possible to more comprehensively judge from multiple angles whether it is necessary to adjust the network structure of the digital twin model. Furthermore, the network structure of the digital twin model is optimized, which not only involves increasing or decreasing the number of network layers, that is, adjusting the depth of the network to adapt to task requirements of different complexities, but also includes fine-tuning the details within each layer, such as changing the activation function or adjusting the number of neurons, that is, adjusting the width of the network. This multi-level adjustment method allows the digital twin model to avoid unnecessary computing overhead while maintaining sufficient adaptability. The digital twin model structure is automatically optimized through machine learning algorithms to improve the performance of the digital twin model, such as real-time performance and prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] This specification will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:

[0020] Figure 1It is a flowchart of a digital twin dynamic optimization method based on machine learning according to some embodiments of this specification;

[0021] Figure 2 It is a schematic diagram of a process for optimizing the network structure of a digital twin model according to some embodiments of this specification;

[0022] Figure 3 It is a schematic diagram of a process for optimizing the network structure of a digital twin model according to other embodiments of this specification;

[0023] Figure 4 It is a schematic diagram of a process for optimizing the network structure of a digital twin model according to other embodiments of this specification;

[0024] Figure 5 It is a module schematic diagram of a digital twin dynamic optimization system based on machine learning according to some embodiments of this specification;

[0025] Figure 6 is a schematic diagram of a computer device according to some embodiments of the present specification. DETAILED DESCRIPTION

[0026] In order to more clearly illustrate the technical solutions of the embodiments of this specification, the following is a brief introduction to the drawings required for the description of the embodiments. Obviously, the drawings described below are only some examples or embodiments of this specification. For ordinary technicians in this field, this specification can also be applied to other similar scenarios based on these drawings without creative work. Unless it is obvious from the language environment or otherwise explained, the same reference numerals in the figures represent the same structure or operation.

[0027] Figure 1 is a flow chart of a digital twin dynamic optimization method based on machine learning according to some embodiments of this specification, such as Figure 1 As shown, the digital twin dynamic optimization method based on machine learning can include the following process.

[0028] Step 110, obtaining a real-time data stream of the manufacturing environment.

[0029] The real-time data stream of the manufacturing environment may include real-time data related to the production line, such as the real-time temperature of the production line equipment, the production line processing speed, etc.

[0030] Step 120, obtaining real-time performance data of the digital twin model.

[0031] The real-time performance data of the digital twin model may include prediction accuracy, response time, resource consumption, etc. The real-time performance data of the digital twin model may be used to evaluate the current operating status of the digital twin model.

[0032] Step 130, obtaining the real-time task requirements of the digital twin model.

[0033] The real-time task requirements of the digital twin model can include the performance requirements of different tasks for the digital twin model. For example, the performance requirements of the production line processing prediction task for the digital twin model include fast response speed, etc. For another example, the real-time temperature task of the production line equipment, the performance requirements of the digital twin model include high accuracy of the prediction information, etc.

[0034] Step 140, based on the real-time data flow of the manufacturing environment, the real-time performance data of the digital twin model, or the real-time task requirements of the digital twin model, determine whether to adjust the network structure of the digital twin model.

[0035] In some embodiments, determining whether to adjust the network structure of the digital twin model includes:

[0036] Based on the real-time data stream of the manufacturing environment and the historical prediction data of the digital twin model, the prediction accuracy of the digital twin model is determined. When the prediction accuracy of the digital twin model is less than the preset prediction accuracy threshold, it is determined to adjust the network structure of the digital twin model.

[0037] For example, for the temperature prediction task, the prediction accuracy of the digital twin model can be calculated based on the following formula:

[0038]

[0039] Among them, A is the prediction accuracy of the digital twin model, P1 is the preset parameter, and P1>0, T (n,current) is the real-time temperature of the production line equipment at the nth time point, T (n,forecast) is the real-time temperature of the production line equipment at the nth time point predicted by the digital twin model, and N is the total number of time points for temperature prediction by the digital twin model.

[0040] In some embodiments, determining whether to adjust the network structure of the digital twin model includes:

[0041] Calculate data flow similarity between the real-time data flow of the manufacturing environment and the historical data flow of the manufacturing environment;

[0042] When the data flow similarity is less than the preset data flow similarity threshold, it is determined to adjust the network structure of the digital twin model.

[0043] For example, if the difference between the data dimension of the real-time data stream of the manufacturing environment and the data dimension of the historical data stream of the manufacturing environment is greater than a preset dimension difference threshold, it is determined that the data stream similarity is less than the preset data stream similarity threshold.

[0044] For another example, if the data volume difference per unit time of the real-time data flow of the manufacturing environment and the data volume per unit time of the historical data flow of the manufacturing environment is greater than a preset data volume difference threshold, then the data flow similarity is determined to be less than the preset data flow similarity threshold.

[0045] In some embodiments, determining whether to adjust the network structure of the digital twin model includes:

[0046] Based on the real-time performance data of the digital twin model, determine the real-time scores of the digital twin model in multiple target performance indicators (e.g., prediction accuracy indicator, response time indicator, etc.);

[0047] Determine whether the real-time scores of the digital twin model in multiple target performance indicators meet the preset performance indicator requirements. When it is determined that the real-time scores of the digital twin model in multiple target performance indicators do not meet the preset performance indicator requirements, determine to adjust the network structure of the digital twin model, or,

[0048] Determine whether the real-time scores of the digital twin model in multiple target performance indicators meet the task performance requirements corresponding to the real-time task requirements of the digital twin model. When it is determined that the real-time scores of the digital twin model in multiple target performance indicators do not meet the task performance requirements corresponding to the real-time task requirements of the digital twin model, determine to adjust the network structure of the digital twin model.

[0049] As an example only, the higher the real-time accuracy of the digital twin model, the higher the score of the digital twin model in the prediction accuracy index, the shorter the time required for the real-time response of the digital twin model, and the higher the score of the digital twin model in the response time index. The preset performance index requirements can be pre-set by the operator, for example, the score of the prediction accuracy index is greater than 8, and the score of the response time index is greater than 7, etc. If the score of the digital twin model in the prediction accuracy index is less than 8 and / or the score of the response time index is less than 7, it is determined that the real-time scores of the digital twin model in multiple target performance indicators do not meet the preset performance index requirements.

[0050] As another example, the task performance requirements corresponding to the real-time task requirements of the digital twin model include a score of a prediction accuracy index greater than 9, and a score of a response time index greater than 8. If the digital twin model scores less than 9 in the prediction accuracy index and / or less than 8 in the response time index, it is determined that the real-time scores of the digital twin model in multiple target performance indicators do not meet the task performance requirements corresponding to the real-time task requirements of the digital twin model.

[0051] Step 150, when it is determined to adjust the network structure of the digital twin model, the network structure of the digital twin model is optimized.

[0052] For example, when the digital twin model performs poorly on a specific task, more neural network layers can be automatically introduced or the number of neurons in the existing layers can be increased to increase the complexity and learning ability of the model. Conversely, if the digital twin model shows signs of overfitting on a specific task, the number of network layers or neurons can be reduced to improve the generalization ability of the digital twin model. In addition, if real-time data shows that a specific type of activation function performs better in the current environment, the activation function can be automatically adjusted to optimize the performance of the digital twin model.

[0053] For example, when the complexity of the digital twin model needs to be increased to improve learning capabilities, this can be achieved by adding new network layers or increasing the number of neurons in existing layers. Conversely, when overfitting of the digital twin model is detected, the strategy will simplify the digital twin model by reducing the number of network layers or neurons to improve its generalization ability. Within the layer, the activation function may be adjusted based on real-time performance feedback, such as switching from the ReLU function to the Sigmoid function, to optimize the model's performance on a specific data set.

[0054] In some embodiments, the network structure of the digital twin model is optimized, including:

[0055] When the prediction accuracy of the digital twin model is less than the preset prediction accuracy threshold, the network structure of the digital twin model is optimized based on the Q learning algorithm, and the optimal network structure configuration of the digital twin model is found in the process of continuous exploration and utilization.

[0056] In some embodiments, the reward function of the Q-learning algorithm is:

[0057]

[0058] Among them, Q(s,a) is the expected utility of taking action a in state s, Q(s′,a′) is the expected utility of taking action a′ in state s′, s′ is the new state after taking the action, a′ is the action of optimizing the network structure of the digital twin model in state s′, α is the learning rate, r is the immediate reward, γ is the discount factor of future rewards, and the action of optimizing the network structure of the digital twin model includes changing the number of layers of the neural network layer, changing the number of neurons in the neural network layer, changing the connection between neurons, and changing at least one of the activation functions in the neural network layer. The activation function can be a function used in the neural network to add nonlinear factors to help the model learn complex functions. The reward function evaluates the quality of the network structure based on the performance of the digital twin model (for example, accuracy or loss function). In this way, it is possible to learn how to adjust the network structure in multiple iterations to maximize the reward function value.

[0059] Figure 2 is a schematic diagram of a process for optimizing the network structure of a digital twin model according to some embodiments of this specification, such as Figure 2 As shown in the figure, specifically, the state space and action space of the network structure are first defined. The state space includes all possible configurations of the current network, while the action space includes all possible structural adjustment operations. In each iteration, an action is selected according to the current network state, and the new state and the immediate reward obtained are observed after execution, and then the Q value is updated according to the above Q learning formula. In this way, the optimal network structure adjustment strategy is gradually learned. "Current network state" represents the current structural configuration of the digital twin model, "selecting an action" is a strategy selection process based on the Q table, "observing the new state and the immediate reward obtained after execution" is the process of model structure adjustment and evaluation, and "updating the Q value according to the above Q learning formula" is the core of the learning process, which is responsible for updating the strategy to improve future decisions.

[0060] Just as an example, suppose the digital twin model is performing a temperature prediction task. Based on the real-time data stream, the model's accuracy begins to decline. After detecting this performance decline, the reinforcement learning algorithm selects the action of adding a network layer. After executing this action, if the digital twin model's temperature prediction accuracy improves, then this action and the resulting state change will receive a positive reward, making it more likely to be selected in similar situations in the future. Over time, the network structure is continuously optimized to adapt to the changing data streams and task requirements.

[0061] For example, the digital twin model is processing a complex manufacturing process optimization task. During this process, the accuracy of the model begins to decrease, indicating that the complexity of the model needs to be increased to capture more process features. You can choose to add a new network layer or add more neurons to the existing layer. After this adjustment, if the performance of the digital twin model is improved, that is, the accuracy increases or the loss function value decreases, the system will receive positive feedback, making it more inclined to take similar actions when encountering similar situations in the future. In this way, the network structure configuration can be continuously optimized to adapt to changing task requirements and data environments.

[0062] In some embodiments, the network structure of the digital twin model is optimized, including:

[0063] When the data flow similarity is less than the preset data flow similarity threshold, the network structure of the digital twin model is optimized based on evolutionary strategy and reinforcement learning algorithm.

[0064] Figure 3 is a schematic diagram of a process for optimizing the network structure of a digital twin model according to other embodiments of this specification, such as Figure 3 As shown in the figure, specifically, the reinforcement learning part defines a set of environmental states, which represents the performance and network configuration of the current digital twin model. In each state, a variety of actions can be taken, such as adding or reducing network layers, changing activation functions, adjusting the number of neurons, etc. Each action will lead to a change in model performance, which is regarded as a reward or punishment given by the environment. The Q learning algorithm in reinforcement learning is used to update the value of the action to optimize future action selection. Q learning updates the action value function by iteration, and the formula is as follows:

[0065]

[0066] Among them, Q(s, a) is the expected utility of taking action a in state s, Q(s′, a′) is the expected utility of taking action a′ in state s′, s′ is the new state after taking the action, a′ is the action to optimize the network structure of the digital twin model in state s′, α is the learning rate, r is the immediate reward, and γ is the discount factor for future rewards.

[0067] In the evolution strategy part, the concept of population is used to maintain a variety of network structure candidates. Through the application of fitness evaluation and genetic operators (such as crossover and mutation), those network structures with good performance can be retained and enhanced in the population, while those with poor performance can be eliminated. Evolutionary strategies can provide global search capabilities to help reinforcement learning avoid falling into local optimal solutions.

[0068] As an example, the digital twin model is facing a new data set whose characteristics are quite different from the previous data set. First, the performance of the current network structure is evaluated through reinforcement learning, and actions are taken to adjust it if necessary. After the adjustment, the system will update the Q value according to the new performance of the model. Subsequently, the evolutionary strategy is applied to the population of network structures, and new candidate network structures are generated through crossover and mutation operations, and evaluated based on their performance on the new data set. Excellent network structures will be retained and used for the next round of optimization. In this way, the network structure can be continuously optimized to adapt to complex and changing data environments and performance requirements.

[0069] In some embodiments, the network structure of the digital twin model is optimized, including:

[0070] When the real-time scores of multiple target performance indicators of the digital twin model do not meet the preset performance indicator requirements, the network structure of the digital twin model is optimized based on the deep deterministic policy gradient algorithm.

[0071] The deep deterministic policy gradient algorithm can learn policies in a continuous action space. The formula is as follows:

[0072]

[0073] Among them, Δθ represents the change of policy parameters, α is the learning rate, Q(s,a|θ) is the action value function, and s t is the current state, μ(s t |θ μ ) is the action under the current strategy, where the strategy parameters may include at least one of the number of neural network layers, the number of neurons in the neural network layer, the connections between neurons, and the activation function in the neural network layer.

[0074] Figure 4 is a schematic diagram of a process for optimizing the network structure of a digital twin model according to other embodiments of this specification, such as Figure 4 As shown, for example only, assume that the performance of the digital twin model degrades when processing a new data stream. Performance monitoring is performed to capture performance indicators in real time and analyze the reasons for the performance degradation through feedback. A deep deterministic policy gradient algorithm is then used for decision making to determine the optimal adjustment strategy, such as increasing the number of layers to increase the complexity and learning ability of the model. The dynamic adjustment module executes the strategy, adds network layers, and retrains the model. The updated digital twin model is put into use again for performance monitoring, and its performance continues to be monitored to ensure that the adjustment has achieved the expected effect. Through this circular feedback process, the digital twin model can continuously optimize itself to adapt to the changing data environment and performance requirements.

[0075] In some embodiments, the network structure of the digital twin model is optimized, including:

[0076] Using genetic algorithms or particle swarm optimization evolutionary strategies, the network structure of the digital twin model is optimized based on the real-time data flow of the manufacturing environment and the real-time task requirements of the digital twin model.

[0077] Just as an example, if the digital twin model is deployed in a changing manufacturing environment, the environmental adaptability module may detect the increase in production line speed, which requires the digital twin model to make predictions quickly. At the same time, the task analysis module recognizes that the current task requires the model to improve the accuracy of the prediction. The structure search and optimization module will search for network structures that can improve accuracy while ensuring response speed, and may choose to add convolutional layers to extract more complex features. Finally, the model retraining module updates the model parameters using the new network structure and recently collected data to ensure that the model can perform optimally under new environmental and task requirements. In this way, it is ensured that the digital twin model can dynamically adjust its structure to adapt to changing conditions and requirements.

[0078] Step 160, updating the parameters of the digital twin model after structural optimization.

[0079] It is understandable that the dynamic optimization method of digital twins based on machine learning can dynamically optimize the performance of digital twin models. This strategy not only involves increasing or decreasing the number of network layers, that is, adjusting the depth of the network to adapt to the requirements of tasks of different complexity, but also includes fine-tuning the details within each layer, such as changing the activation function or adjusting the number of neurons, that is, adjusting the width of the network. This multi-level adjustment method allows the model to avoid unnecessary computational overhead while maintaining sufficient adaptability, and realizes automatic optimization of the digital twin model structure through machine learning algorithms to improve the real-time performance and prediction accuracy of the digital twin model.

[0080] Figure 5 is a module schematic diagram of a digital twin dynamic optimization system based on machine learning according to some embodiments of this specification, such as Figure 5 As shown, the digital twin dynamic optimization system based on machine learning may include an environmental adaptability module 510, a performance monitoring module 520, a task analysis module 530, an optimization judgment module 540, a structure search and optimization module 550 and a model retraining module 560.

[0081] The environmental adaptability module 510 may be used to obtain real-time data streams of the manufacturing environment.

[0082] Specifically, the environmental adaptability module 510 may be responsible for analyzing the characteristics of the real-time data stream input into the manufacturing environment and its changing trends to identify changes in the operating environment, and utilize machine learning techniques, such as cluster analysis or anomaly detection algorithms, to monitor and adapt to changes in the manufacturing environment in real time.

[0083] The performance monitoring module 520 can be used to obtain real-time performance data of the digital twin model.

[0084] The task parsing module 530 can be used to obtain the real-time task requirements of the digital twin model.

[0085] Specifically, the task analysis module 530 can be used to understand the current task requirements and convert high-level business goals into performance indicators that can be understood by the digital twin model. For example, if the task requirement emphasizes response speed, the task analysis module 530 will give priority to reducing the response time of the model.

[0086] The optimization judgment module 540 can be used to determine whether to adjust the network structure of the digital twin model based on the real-time data flow of the manufacturing environment, the real-time performance data of the digital twin model, or the real-time task requirements of the digital twin model.

[0087] The structure search and optimization module 550 can be used to optimize the network structure of the digital twin model when it is determined to adjust the network structure of the digital twin model.

[0088] Specifically, according to the output of the environment adaptability module 510 or the task analysis module 530, the structure search and optimization module 550 uses dynamic network structure search technology, such as evolutionary strategy based on genetic algorithm or particle swarm optimization, to explore the optimal network structure configuration. The structure search and optimization module 550 will evaluate the impact of different network structures on performance indicators and select the best solution.

[0089] The model retraining module 560 can be used to update the parameters of the digital twin model after structural optimization.

[0090] Specifically, the model retraining module 560 can be responsible for retraining the digital twin model after structural optimization using the latest data to ensure that the parameters of the digital twin model after structural optimization are adapted to the new structure and maximize performance improvement.

[0091] The digital twin dynamic optimization system based on machine learning can be used to execute the digital twin dynamic optimization method based on machine learning. For more descriptions of the digital twin dynamic optimization system based on machine learning, please refer to the relevant descriptions of the digital twin dynamic optimization method based on machine learning, which will not be repeated here.

[0092] Figure 6It is a schematic diagram of a computer device according to some embodiments of this specification, the computer device includes a memory and a processor, the memory stores computer-readable instructions, and when the instructions are executed by the processor, the processor executes the above-mentioned digital twin dynamic optimization method based on machine learning. It should be noted that Figure 6 The computer device 600 shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0093] like Figure 6 As shown, the processor includes a central processing unit (CPU) 601, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 602 or the program loaded from the storage part 608 to the random access memory (RAM) 603, such as executing the method in the above embodiment. In RAM 603, various programs and data required for system operation are also stored. CPU 601, ROM 602 and RAM 603 are connected to each other through bus 604. Input / output (I / O) interface 605 is also connected to bus 604.

[0094] The following components are connected to the I / O interface 605: an input section 606 including a keyboard, a mouse, etc.; an output section 607 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN (Local Area Network) card, a modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the I / O interface 605 as needed. A removable medium 611, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 610 as needed so that a computer program read therefrom is installed into the storage section 608 as needed.

[0095] In particular, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication section 609, and / or installed from a removable medium 611. When the computer program is executed by a central processing unit (CPU) 601, various functions defined in the system of the present application are executed.

[0096] It should be noted that the computer-readable medium shown in the embodiment of the present application can be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media can include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a flash memory, an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, wherein a computer-readable computer program is carried. This propagated data signal can take a variety of forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. A computer program contained on a computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wired, etc., or any suitable combination of the above.

[0097] Finally, it should be understood that the embodiments described in this specification are only used to illustrate the principles of the embodiments of this specification. Other variations may also fall within the scope of this specification. Therefore, as an example and not a limitation, alternative configurations of the embodiments of this specification may be considered consistent with the teachings of this specification. Accordingly, the embodiments of this specification are not limited to the embodiments explicitly introduced and described in this specification.

Claims

1. A digital twin dynamic optimization method based on machine learning, characterized in that: include: Get real-time data streams from your manufacturing environment; Obtain real-time performance data of the digital twin model; Obtaining real-time task requirements of the digital twin model; Based on the real-time data flow of the manufacturing environment, the real-time performance data of the digital twin model, or the real-time task requirements of the digital twin model, determining whether to adjust the network structure of the digital twin model; When it is determined to adjust the network structure of the digital twin model, optimizing the network structure of the digital twin model; Update the parameters of the digital twin model after structural optimization.

2. The digital twin dynamic optimization method based on machine learning according to claim 1 is characterized in that: Based on the real-time data flow of the manufacturing environment, the real-time performance data of the digital twin model, or the real-time task requirements of the digital twin model, determining whether to adjust the network structure of the digital twin model includes: Based on the real-time data stream of the manufacturing environment and the historical prediction data of the digital twin model, the prediction accuracy of the digital twin model is determined. When the prediction accuracy of the digital twin model is less than a preset prediction accuracy threshold, it is determined to adjust the network structure of the digital twin model.

3. The digital twin dynamic optimization method based on machine learning according to claim 2 is characterized in that: Optimizing the network structure of the digital twin model includes: When the prediction accuracy of the digital twin model is less than a preset prediction accuracy threshold, the network structure of the digital twin model is optimized based on a Q learning algorithm.

4. The digital twin dynamic optimization method based on machine learning according to claim 3 is characterized in that: The reward function of the Q-learning algorithm is: Among them, Q(s,a) is the expected utility of taking action a in state s, Q(s′,a′) is the expected utility of taking action a′ in state s′, s′ is the new state after taking the action, a′ is the action of optimizing the network structure of the digital twin model in state s′, α is the learning rate, r is the immediate reward, γ is the discount factor of future rewards, and the action of optimizing the network structure of the digital twin model includes changing the number of layers of the neural network layer, changing the number of neurons in the neural network layer, changing the connection between neurons, and changing at least one of the activation functions in the neural network layer.

5. The digital twin dynamic optimization method based on machine learning according to claim 1, characterized in that: Based on the real-time data flow of the manufacturing environment, the real-time performance data of the digital twin model, or the real-time task requirements of the digital twin model, determining whether to adjust the network structure of the digital twin model includes: Calculating data flow similarity between the real-time data flow of the manufacturing environment and the historical data flow of the manufacturing environment; When the data flow similarity is less than a preset data flow similarity threshold, it is determined to adjust the network structure of the digital twin model.

6. The digital twin dynamic optimization method based on machine learning according to claim 5 is characterized in that: Optimizing the network structure of the digital twin model includes: When the data flow similarity is less than a preset data flow similarity threshold, the network structure of the digital twin model is optimized based on evolutionary strategy and reinforcement learning algorithm.

7. The digital twin dynamic optimization method based on machine learning according to claim 1 is characterized in that: Based on the real-time data flow of the manufacturing environment, the real-time performance data of the digital twin model, or the real-time task requirements of the digital twin model, determining whether to adjust the network structure of the digital twin model includes: Based on the real-time performance data of the digital twin model, determining the real-time scores of the digital twin model in multiple target performance indicators; Determine whether the real-time scores of the digital twin model in multiple target performance indicators meet the preset performance indicator requirements, and when it is determined that the real-time scores of the digital twin model in multiple target performance indicators do not meet the preset performance indicator requirements, determine to adjust the network structure of the digital twin model, or, Determine whether the real-time scores of the digital twin model in multiple target performance indicators meet the task performance requirements corresponding to the real-time task requirements of the digital twin model. When it is determined that the real-time scores of the digital twin model in multiple target performance indicators do not meet the task performance requirements corresponding to the real-time task requirements of the digital twin model, determine to adjust the network structure of the digital twin model.

8. The digital twin dynamic optimization method based on machine learning according to claim 7 is characterized in that: Optimizing the network structure of the digital twin model includes: When the real-time scores of the digital twin model in multiple target performance indicators do not meet the preset performance indicator requirements, the network structure of the digital twin model is optimized based on a deep deterministic policy gradient algorithm.

9. The digital twin dynamic optimization method based on machine learning according to claim 7 is characterized in that: Optimizing the network structure of the digital twin model includes: When it is determined that the real-time scores of the digital twin model in multiple target performance indicators do not meet the task performance requirements corresponding to the real-time task requirements of the digital twin model, an evolutionary strategy of a genetic algorithm or a particle swarm optimization is used to optimize the network structure of the digital twin model based on the real-time data flow of the manufacturing environment and the real-time task requirements of the digital twin model.

10. A digital twin dynamic optimization system based on machine learning, characterized in that: include: Environmental adaptability module, used to obtain real-time data streams of the manufacturing environment; Performance monitoring module, used to obtain real-time performance data of the digital twin model; A task parsing module, used to obtain the real-time task requirements of the digital twin model; An optimization judgment module, used to judge whether to adjust the network structure of the digital twin model based on the real-time data flow of the manufacturing environment, the real-time performance data of the digital twin model or the real-time task requirements of the digital twin model; A structure search and optimization module, used to optimize the network structure of the digital twin model when it is determined to adjust the network structure of the digital twin model; The model retraining module is used to update the parameters of the digital twin model after structural optimization.

Images