Digital twin model incremental learning method and system and equipment state prediction method

Abstract

The invention provides a digital twin model incremental learning method and system and an equipment state prediction method.The digital twin model incremental learning method introduces an incremental learning mechanism based on elastic weight merging to train an industrial equipment state parameter prediction model; the current working condition samples are stored in an experience playback buffer area, the experience playback buffer area adopts a circular queue structure, the samples stored in the buffer area are used in a mixed mode for training when a new working condition is learned, and mixed loss optimization is executed on a current data batch; after batch training is completed, Fisher information matrix diagonal terms, about historical data, of current parameter configuration are calculated firstly, then global parameter importance is updated until Neural ODE incremental training converges, and finally a complete model state tuple fusing multi-working-condition knowledge is output; the incremental Neural ODE can continue to learn knowledge of new working conditions on the premise that knowledge of old working conditions is not forgotten, and the sustainable evolution ability of the digital twin under the limited resource condition is achieved.

Description

TECHNICAL FIELD

[0001] The application belongs to the cross field of artificial intelligence and industrial digitization, and particularly relates to a digital twin model incremental learning method, a system and an equipment state prediction method. BACKGROUND

[0002] Industrial equipment digital twin (DT) is a real-time mirror system of a physical entity in a virtual space established through multi-dimensional and multi-scale dynamic modeling. Its core components include: a physical entity layer, a virtual model layer, and a data interaction layer. In the physical entity layer, a sensor network distributed throughout various types of industrial equipment collects real-time state data of the equipment; the virtual model layer integrates mechanism models (such as multi-physical field coupled finite element models, behavior characteristic evolution models, etc.) based on physical laws and data-driven models (machine learning algorithms, deep neural networks, etc. artificial intelligence algorithms) to form a high-fidelity dynamic simulation environment; the data interaction layer relies on an industrial Internet of Things platform to achieve millisecond-level data synchronization, supporting more than ten thousand times of data interaction per second.

[0003] Currently, the digital twin system mainly supports industrial applications through three core capabilities. First, the adaptive dynamic modeling capability realizes the accurate digital mapping of physical objects through a multi-scale joint solving engine. The technical core of adaptive dynamic modeling includes a multi-physical field coupling mechanism, a dynamic evolution representation mechanism, and an uncertainty quantification mechanism. The multi-physical field coupling mechanism relies on physical laws such as comprehensive solid mechanics deformation field, electromagnetic field, thermodynamics field, and fluid field to establish a coupled partial differential equation system solving framework. The variational autoencoder (VAE) is used to embed the physical law constraints into the neural network architecture, realizing parameter sharing of mechanism models and data-driven models, and ensuring the robustness of model extrapolation under unobserved working conditions. The dynamic evolution representation mechanism is based on continuous time hidden space modeling, describing the random evolution process of system parameters through stochastic differential equations, capturing uncontrollable disturbances in the degradation process of industrial equipment, and realizing the quantitative representation of continuous time-varying characteristics such as wear and aging of industrial equipment in reality. The uncertainty quantification mechanism adopts Bayesian deep kernel learning to construct a probability prediction model, separates the observation noise and model error through a latent variable Gaussian process, and outputs the confidence interval of the prediction result, supporting risk-sensitive decision-making.

[0004] Although digital twin technology has achieved remarkable results in industrial applications, the current technology still faces severe challenges. In terms of dynamic adaptability, most existing models cannot effectively track the real-time changes in the running state of industrial equipment. The lack of continuous learning capability is also a deep-seated problem. When industrial equipment faces changes in working conditions or model switching, the accuracy of the original model will be greatly reduced. In addition, there are also great challenges in the real-time interaction between the digital twin model of industrial equipment and the physical entity. The patent with publication number CN120145285A provides a robot health state evaluation method based on artificial intelligence, which constructs an optimized feature set by fusing multi-source sensor data, realizes digital twin health modeling by using a dynamic double data flow neural differential equation combined with physical constraints and wear evolution, and outputs robot health degree parameters; introduce quantum annealing optimization and graph attention mechanism, build a multi-level predictive maintenance architecture, realize fault propagation prediction and maintenance optimization from the part level to the system level, and realize fault-tolerant control combined with model predictive control, which focuses on using a dynamic double data flow neural differential equation modeling method to perform digital twin health modeling, realize joint energy evolution modeling and wear trend prediction under physical consistency constraints, and improve modeling accuracy and actual scene fitting degree, but the model does not have continuous learning capability.

[0005] The existing digital twin technology lacks continuous description capability for dynamic evolution systems and lacks cross-life cycle knowledge fusion mechanism, which constitutes the core bottleneck restricting industrial landing. Traditional static modeling methods cannot capture the continuous parameter drift characteristics of the equipment degradation process, and mainstream incremental learning techniques have inherent problems such as catastrophic forgetting and model inflation when dealing with time-series dynamic systems. Especially in high real-time requirement scenarios, the computational complexity of regularization methods and the memory occupation of dynamic architecture schemes all form a serious conflict with industrial embedded platforms. SUMMARY

[0006] In order to solve the problems existing in the prior art, the present application provides a digital twin model incremental learning method, specifically a digital twin model incremental learning method and system based on neural ordinary differential equation (Neural ODE), which fuses an elastic weight consolidation (EWC) mechanism in Neural ODE, aiming to solve the problem of continuous model updating of digital twin systems in dynamic industrial environments, and to overcome the difficulty of interaction between the twin model and the physical entity in complex and variable industrial environments.

[0007] In order to achieve the above purpose, in a first aspect, the present application provides a digital twin model incremental learning method, comprising the following steps: Establish an industrial equipment state parameter prediction model based on Neural ODE; An incremental learning mechanism based on elastic weight merging is introduced to train the industrial equipment state parameter prediction model, specifically, a Neural ODE new working condition incremental training based on the fusion of the EWC mechanism, the following three steps are repeatedly executed until the Neural ODE incremental training converges, and finally the complete model state tuple fused with multi-working condition knowledge is output : When the new working condition data set is first acquired, the current working condition sample is stored in the experience replay buffer, the experience replay buffer adopts a circular queue structure, and the sample stored in the buffer is mixed and used for training when learning the new working condition, When the training task is not the first time, the historical model parameters are loaded, and the current data batch Perform hybrid loss optimization; After completing batch training, the current parameter configuration Regarding the Fisher information matrix diagonal item of the historical data, the global parameter importance is updated again.

[0008] Further, the industrial equipment state parameter prediction model based on Neural ODE includes: The running parameters and state parameters of the equipment are summarized as external control variables and target prediction variables of the Neural ODE model as needed, the time series of the external control variables and the target prediction variables at the corresponding time are input into the ODE function, the ODE function learns the mapping relationship between the first derivative of the target variable and the control variable, the first derivative of the target variable is projected into the hidden space, and the hidden state of the target variable is input into the ODE module as a new sequence, dopri5 with adaptive step size characteristics is used as a common differential equation solver, and the hidden state of the target variable is integrated step by step through dopri5, to realize the reconstruction of the numerical value of the target variable in the original data space.

[0009] Further, the basic form of Neural ODE is:

[0010] Among them, is the running state of the industrial equipment at a certain time; is the parameter to be learned in the network; is the time, The initial state of each parameter is represented as: The hidden state of the target variable is gradually integrated to solve each layer hidden state, the running parameters of the device are taken as the characteristic input of the Neural ODE, the to-be-predicted quantity is taken as the output, the derivative of the to-be-predicted quantity change is fitted by the ODE function, and then the derivative is integrated by the ODE module to solve the continuous value of the to-be-predicted quantity, after the solution of the to-be-predicted quantity value at the termination time is completed, the model validation loss is solved; the model gradient is updated by using the adjoint sensitivity method.

[0011] The model validation loss is solved according to the following formula:

[0012] Further, the model gradient update by using the adjoint sensitivity method includes: The adjoint state is the first derivative of the to-be-predicted quantity loss at each time with respect to the hidden state at the time, and is specifically expressed as follows:

[0013] In the formula, is the hidden state of the to-be-predicted quantity at the time the adjoint state of the network parameters under the time variable and the adjoint state of the time is as follows:

[0014]

[0015] According to , the and are solved by reverse iteration, and the global parameter gradient is obtained by the following formula:

[0016]

[0017] The network parameters are updated, the Neural ODE gradient update is completed, and the above process is repeatedly iterated to promote the model validation loss to converge, and the training of the Neural ODE state prediction model is completed.

[0018] Further, when the current working condition sample is stored in the experience replay buffer, the maximum threshold C of the experience replay buffer is set, if the current sample number of the experience replay buffer is greater than or equal to C, the old sample is removed by a pseudo-random elimination mechanism.

[0019] Further, the specific formula of the mixed loss optimization performed on the current data batch is as follows:​​

[0020] wherein the first term is the supervised loss of the new task, the second term is the EWC regularization term, and the third term is the knowledge distillation loss. 、 determined by hyperparameter tuning, is the historical task parameter importance matrix.

[0021] Further, the current The Fisher information matrix diagonal term about historical data:

[0022] Update the global parameter importance according to the following formula:

[0023] wherein, is the decay factor.

[0024] In a second aspect, the present application can provide a digital twin model incremental learning system, comprising a model construction module and a training module; The model construction module establishes an industrial equipment state parameter prediction model based on Neural ODE; The training module is used to introduce an incremental learning mechanism based on elastic weight merging to train the industrial equipment state parameter prediction model, specifically based on Neural ODE new condition incremental training of the fusion EWC mechanism, the following three steps are repeatedly executed until the Neural ODE incremental training converges, and finally a complete model state tuple of fused multi-condition knowledge is output

[0025] When the new condition data set is acquired for the first time, the current condition sample is stored in the experience replay buffer, the experience replay buffer adopts a circular queue structure, and the sample stored in the buffer is mixed and used for training when learning the new condition, When the training task is not the first time, the historical model parameters are loaded, and the current data batch Perform mixed loss optimization; After completing batch training, first calculate the current parameter configuration The Fisher information matrix diagonal term about historical data is updated again.

[0026] In a third aspect, the present application provides an industrial equipment state prediction method, which adopts the digital twin model incremental learning method as described above and comprises the following steps: The real-time operation state data generated by the industrial equipment test bench is taken as a training data set of the model, and the model is trained in sequence according to the working condition label, so as to obtain the digital twin model of the industrial equipment based on the incremental Neural ODE. The real-time operation state data generated by the industrial equipment test bench is taken as an external control variable of the digital twin model based on the incremental Neural ODE, the real-time operation state data of the industrial equipment is taken as an input of the digital twin model based on the incremental Neural ODE, the digital twin model based on the incremental Neural ODE outputs a prediction result, and the prediction result includes a monitoring target of the industrial equipment.

[0027] Further, when the model is trained in sequence according to the working condition label, the original Neural ODE model is pre-trained, the optimal parameter configuration is saved, the incremental training of the remaining working conditions is started one by one, and the training of all working conditions is completed, so as to obtain the digital twin model of the industrial equipment based on the incremental Neural ODE.

[0028] Compared with the prior art, the present application has at least the following beneficial effects: The present application firstly establishes a data-driven state prediction model based on a deep neural network for a target device, adopts a Neural ODE model with good modeling capability for a dynamic continuous system as a device state prediction model, dynamically models the target device state, realizes offline high-precision prediction for a to-be-predicted quantity, secondly introduces an incremental learning mechanism based on EWC on the basis of the Neural ODE model, endows the model with continuous learning capability for complex and variable working conditions, and can improve the online prediction capability of the model through small-batch training, constructs an incremental Neural ODE model for device variable working condition state prediction, which can meet the requirement that after the training of the last working condition is completed, the state prediction of the model for the original several working conditions still maintains high precision, and further makes the model have good cross-working condition generalization capability, effectively improves the performance defects of ordinary layered neural networks including the Neural ODE in the variable working condition state prediction scene. The device digital twin model is constructed based on the incremental Neural ODE, so that the digital twin model can predict the real-time state of the device online, meets the real-time requirement of the digital twin technology, forms a faithful mapping relationship between the twin and the physical entity, and establishes a reliable model basis for device state monitoring, intelligent operation and maintenance and health management. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 It is a Neural ODE modeling method schematic diagram.

[0030] Figure 2 It is a new working condition incremental training logic schematic diagram.

[0031] Figure 3 Incremental Neural ODE training method flowchart.

[0032] Figure 4 Application of incremental Neural ODE-based digital twin model in generator health state monitoring. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0034] The present application provides an industrial equipment digital twin model incremental learning method based on neural ordinary differential equation, proposes a breakthrough scheme of integrating EWC mechanism into Neural ODE framework, realizes the continuous evolution ability of digital twin under the condition of limited resources by establishing a dynamic parameter importance evaluation model and a lightweight regularization calculation paradigm, and provides a new generation of solution for manufacturing industry considering dynamic adaptability, knowledge continuity and deployment feasibility. The method first establishes an industrial equipment state parameter prediction model based on Neural ODE. Specifically, Neural ODE with good dynamic system modeling performance is used to dynamically model a single operating scene of industrial equipment with high real-time demand, and preliminary high-precision offline prediction of equipment state under single condition is realized. Secondly, the incremental learning mechanism based on EWC is introduced on the basis of the original Neural ODE model, so that the incremental Neural ODE can continue to learn the knowledge of new working conditions without forgetting the knowledge of old working conditions, maintain stable model prediction performance in the whole prediction process, endow the original Neural ODE model with continuous learning ability for complex and variable dynamic environmental conditions faced by various equipment, and then realize dynamic updating of the industrial equipment digital twin model based on incremental learning, and respond to the real-time interaction demand of the model and the physical entity.

[0035] Step 1, constructing a Neural ODE state prediction model Neural Ordinary Differential Equation (Neural ODE) is a model framework that integrates ordinary differential equation solving mechanism into neural network, which has excellent continuous modeling ability for dynamic system and can flexibly capture the characteristics of dynamic evolution of industrial equipment operating state. Based on ordinary differential equation, continuous dynamics modeling of equipment state is realized.

[0036] The basic form of Neural ODE is:

[0037] where, is the running state of the industrial equipment at a certain time (to be predicted quantity); is the to-be-learned parameter in the network (including the weight and bias of the linear layer in the ODE function); is the time.

[0038] The initial state of each parameter is represented as: The formula for solving the hidden state of each layer is:

[0039] The running parameters of the equipment are taken as the feature input of the Neural ODE, and the to-be-predicted quantity is taken as the output. The derivative of the to-be-predicted quantity change is fitted by the ODE function, and then the derivative is integrated by the ODE module to solve the continuous value of the to-be-predicted quantity. After completing the solution of the to-be-predicted quantity value at the termination time, the loss is solved according to the following formula:

[0040] The model gradient is updated by using the adjoint sensitivity method. The adjoint state is the first-order derivative of the to-be-predicted quantity loss with respect to the hidden state at the time, which is specifically represented as follows:

[0041] In the formula, is the hidden state of the to-be-predicted quantity at the time .

[0042] Similarly, the network parameters under the time variable and the adjoint state of the time are as follows:

[0043]

[0044] Since the state quantity at the termination time is only related to the network parameter state at the last time, it is set that , and is solved by backward iteration and . The global parameter gradient is obtained by the following formula:

[0045] ​​

[0046] updating the network parameters , completing the Neural ODE gradient update process, and through repeated iteration of the above process, the model verification loss is converged, and the training process of the Neural ODE state prediction model is completed.

[0047] The modeling method of Neural ODE is as shown in Figure 1 The Neural ODE network is composed of an ODE function and an ODE module, and its core is to realize time series regression prediction under multi-factor control conditions by using the state transition process in the continuous time domain. Specifically, according to the application scene and application requirements of a specific industrial equipment, the operating parameters and state parameters of the equipment are summarized as external control variables and target prediction variables of the Neural ODE model as needed, the time series of the external control variables and the target prediction variables at the corresponding time are input into the ODE function, and the ODE function is used to learn the mapping relationship between the first derivative of the target variable and the control variable. The derivative of the target variable is projected into the hidden space, and then the hidden state of the target variable is input into the ODE module as a new sequence. Dopri5 with adaptive step size characteristics is selected as the ordinary differential equation solver, and the hidden state of the target variable is integrated step by step by dopri5. Finally, the numerical reconstruction of the target variable in the original data space is realized, and the prediction process based on Neural ODE is performed.

[0048] Step 2, build and train the incremental Neural ODE model Elastic Weight Consolidation (EWC) is an algorithm for solving the catastrophic forgetting problem in neural network continual learning. The core idea of EWC is to dynamically constrain the update of key parameters, expecting to find a solution for a new task near the old task, and then balance the stability of new task learning and old task memory.

[0049] First, based on Neural ODE, the initial working condition of the industrial equipment is learned, and the optimal parameter configuration of the trained model is saved Prepare for subsequent incremental learning for new working conditions.

[0050] By introducing the EWC mechanism to the Neural ODE reduced-order model, continuous learning of complex and variable working conditions is realized, and the data-driven model is given the ability to dynamically update parameters with changes in working conditions, avoiding excessive forgetting of old working condition knowledge when learning new working condition knowledge.

[0051] The Fisher information matrix measures the sensitivity of the model parameters to the working condition loss function based on the data distribution, and the important parameters in the model correspond to a higher Fisher value. The Fisher information matrix is calculated as follows:

[0052] EWC calculates the importance of model parameters based on the Fisher information matrix, which can realize the regularization constraint of the key weights of the old working condition, and then limit the update range of important weights.

[0053] Before the incremental training of the new working condition, the Fisher information matrix needs to be calculated to quantify the importance of the parameters to the old working condition, and the importance of the parameters calculated after each incremental training is saved. The old working condition data is used as prior knowledge for the incremental training of the new working condition; the incremental training logic of the new working condition is as shown in Figure 2 .

[0054] The Neural ODE incremental training method based on the fusion EWC mechanism provided in this embodiment is as shown in Figure 3 .

[0055] The model initialization module constructs the Neural ODE model framework. The Neural ODE model establishes the state evolution relationship in the form of ordinary differential equations. When the system first acquires the new working condition data set , the historical parameter loading process is skipped, and the data management module is executed.

[0056] The current working condition sample is stored in the experience replay buffer, and the experience replay buffer uses a circular queue structure. When learning the new working condition, the samples stored in the buffer are mixed and used for training to help the model maintain the memory of the old working condition. Set the maximum threshold C of the experience replay buffer. If the current number of samples in the experience replay buffer is greater than or equal to C, remove the old samples through a pseudo-random elimination mechanism to control the memory occupation in the incremental training.

[0057] For the first task, the current data batch is directly executed for mixed loss optimization; For non-first tasks, the dynamic training module loads the historical model parameters, and the current data batch is executed for mixed loss optimization, and the specific formula is as follows:

[0058] The first term is the supervised loss of the new task, the second term is the EWC regularization term, and the third term is the knowledge distillation loss. The weight , is determined through hyperparameter tuning, is the historical task parameter importance matrix.

[0059] The knowledge solidification module performs the following after completing batch training: first, calculate the current parameter configuration The formula for the Fisher information matrix diagonal item about historical data is:

[0060] After completing the current Fisher information matrix diagonal item calculation, update the global parameter importance, and the formula is as follows:

[0061] Wherein is the decay factor, save the parameter snapshot and the Fisher information matrix for subsequent tasks.

[0062] The closed-loop control module repeatedly performs the foregoing steps until the Neural ODE incremental training converges, and finally outputs the complete model state tuple fused with the knowledge of multiple working conditions , completing the knowledge inheritance closed loop.

[0063] The pseudo code of the incremental Neural ODE algorithm is as follows:

[0064] Introducing the EWC mechanism to the Neural ODE framework can effectively solve the problem of catastrophic forgetting of neural networks. The EWC mechanism allows Neural ODE to protect the knowledge of previous working conditions in the new learning process, and realizes the continuous learning of the model by selectively reducing the plasticity of model weights and biases. By constructing the incremental Neural ODE in the above way, it can dynamically update and predict the running state of industrial equipment under complex and variable working condition conditions, endow the model with the ability of continuous learning, and thus meet the real-time interaction demand of industrial equipment digital twin model and physical entity, and improve the dynamic adaptability of industrial equipment digital twin model under complex environment or variable working condition conditions.

[0065] The digital twin model based on incremental Neural ODE can be applied to dynamic working condition adaptive control, equipment fault prediction and health management, multi-product line rapid switching production and other industrial scenes.

[0066] The present embodiment provides an industrial application case: applying the method of the present application to generator health state monitoring. The specific application process schematic diagram is shown in Figure 4 Figure 4 ​It can be known that when the incremental Neural ODE is used as the digital twin model of the generator for health state monitoring, the incremental Neural ODE model is first trained, and the specific mode is as follows: the running state data of the generator test bench under complex and changeable working conditions is collected by the sensor network as the training data set of the model, and then the model is trained according to the explicit working condition label in sequence.

[0067] During the training, the original Neural ODE model is first pre-trained, the optimal parameter configuration is saved, and then the incremental training of the subsequent working conditions is started one by one until the training of all working conditions is completed, and the obtained model is the generator digital twin model based on the incremental Neural ODE.

[0068] After the training is completed, the generator digital twin model is obtained, and the health state of the generator is monitored. First, the real-time running state data of the generator test bench is generated as the external control variable of the generator digital twin model, the real-time running state data is input into the generator digital twin model, the running state of the generator is predicted online by the twin model, the monitoring targets including the short circuit of the stator winding of the generator, the demagnetization state of the permanent magnet and the temperature rise of the generator are completed by analyzing the prediction results of the model, and the accuracy and real-time performance of the generator digital twin model are evaluated according to the actual situation.

[0069] The above content only illustrates the technical idea of the present application, and cannot limit the protection scope of the present application, and any modification made according to the technical idea of the present application on the basis of the technical scheme falls within the protection scope of the claims of the present application.

Claims

1. A method for incremental learning of a digital twin model, the method comprising: The method comprises the following steps: establishing an industrial equipment state parameter prediction model based on a Neural ODE; An incremental learning mechanism based on elastic weight combination is introduced to train the industrial equipment state parameter prediction model, specifically, Neural ODE new working condition incremental training based on the fusion of the EWC mechanism, the following three steps are repeatedly executed until the Neural ODE incremental training converges, and finally the complete model state tuple fused with multi-working condition knowledge is output : When a new working condition data set is obtained for the first time, the current working condition sample is stored in an experience replay buffer, the experience replay buffer adopts a circular queue structure, and the sample stored in the buffer is mixed and used for training when a new working condition is learned, Load historical model parameters for the current batch of data Perform mixed loss optimization; Compute current parameter configuration importance before completing batch training Update global parameter importance again with respect to Fisher information matrix diagonal of historical data. 2.The digital twin model incremental learning method of claim 1, wherein, The method for establishing an industrial equipment state parameter prediction model based on a Neural ODE comprises the following steps: The running parameters and state parameters of the equipment are summarized as external control variables and target prediction variables of the Neural ODE model as needed, the time sequence of the external control variables and the time sequence of the target prediction variables at the corresponding time are input into an ODE function, the ODE function learns the mapping relationship between the first-order derivative of the target variable and the control variable, the first-order derivative of the target variable is projected into a hidden space, the hidden state of the target variable is input into the ODE module as a new sequence, dopri5 with adaptive step characteristics is used as a common differential equation solver, the hidden state of the target variable is integrated step by step by the dopri5, and numerical reconstruction of the target variable in the original data space is realized. 3.The digital twin model incremental learning method of claim 1, wherein, The basic form of the Neural ODE is as follows: wherein, is the running state of the industrial equipment at a certain time; is the to-be-learned parameter in the network; is the time, The initial state of each variable is represented as: The hidden state of the target variable is gradually integrated to solve each layer hidden state, the running parameters of the device are taken as the characteristic input of the Neural ODE, the to-be-predicted quantity is taken as the output, the derivative of the to-be-predicted quantity change is fitted by the ODE function, the integral of the derivative is carried out by the ODE module, the continuous numerical value of the to-be-predicted quantity is solved, after the solution of the to-be-predicted quantity numerical value at the termination time is completed, the model verification loss is solved; the model gradient is updated by using the adjoint sensitivity method.

4. The digital twin model incremental learning method of claim 1, wherein, The model gradient update is performed by using the adjoint sensitivity method, which comprises the following steps: The adjoint state is the first-order derivative of the loss of the to-be-predicted variable with respect to the hidden state at the time, and is specifically expressed as follows: In the formula, is the hidden state of the to-be-predicted quantity at the moment the accompanying state; Time variable Network parameters under the time variable And time The accompanying state is as follows: According to Backward iteration And The global parameter gradient is obtained by the following formula: Updating network parameters , the Neural ODE gradient update is completed, the model validation loss is promoted to convergence through repeated iteration of the above process, and the training of the Neural ODE state prediction model is completed.

5. The digital twin model incremental learning method of claim 1, wherein, When the current working condition sample is stored in the experience replay buffer, a maximum threshold C of the experience replay buffer is set, if the number of current samples in the experience replay buffer is greater than or equal to C, old samples are removed through a pseudo-random elimination mechanism.

6. The digital twin model incremental learning method of claim 1, wherein, for the current batch of data The specific formula for performing the hybrid loss optimization is as follows: where the first term is the supervised loss of the new task, the second term is the EWC regularization term, the third term is the knowledge distillation loss, and the weight , is determined by hyperparameter tuning, is the parameter importance matrix of the historical task.

7. The digital twin model incremental learning method of claim 1, wherein, The current Fisher information matrix is calculated as Fisher information matrix diagonal terms for historical data: The importance of the global parameters is updated according to the following formula: wherein is an attenuation factor.

8. A digital twin model incremental learning system, comprising: The method comprises a model construction module and a training module; The model construction module establishes an industrial equipment state parameter prediction model based on a Neural ODE; The training module is configured to introduce an incremental learning mechanism based on elastic weight consolidation to train the industrial equipment state parameter prediction model, specifically, to perform the following three steps repeatedly until the Neural ODE incremental training converges, and finally output a complete model state tuple that integrates knowledge of multiple working conditions When a new working condition data set is obtained for the first time, the current working condition sample is stored in an experience replay buffer, the experience replay buffer adopts a circular queue structure, and the sample stored in the buffer is mixed and used for training when a new working condition is learned, At a non-first training task, load history model parameters, to current data batch Perform mixed loss optimization; Compute current parameter configuration importance before completing batch training Update global parameter importance again with respect to Fisher information matrix diagonal of historical data.

9. An industrial equipment state prediction method characterized by, The method for establishing an industrial equipment state parameter prediction model based on a Neural ODE comprises the following steps: Real-time running state data generated by an industrial equipment test bench is used as a training data set of the model, the model is trained according to the working condition labels in sequence, and an industrial equipment digital twin model based on an incremental Neural ODE is obtained; Real-time running state data generated by an industrial equipment test bench is used as an external control variable of the digital twin model based on the incremental Neural ODE, real-time running state data of the industrial equipment is used as an input of the digital twin model based on the incremental Neural ODE, the digital twin model based on the incremental Neural ODE outputs a prediction result, and the prediction result comprises a monitoring target of the industrial equipment.

10. The industrial equipment condition prediction method according to claim 9, characterized by, When the model is trained according to the working condition labels in sequence, the original Neural ODE model is pre-trained, the optimal parameter configuration is saved, incremental training of the remaining working conditions is started one by one, and the training of all working conditions is completed, so that an industrial equipment digital twin model based on an incremental Neural ODE is obtained.

Images