An industrial system mechanism modeling and simulation platform

Abstract

The application discloses an industrial system mechanism modeling simulation platform, a unified system modeling simulation code system is constructed based on a Python language, a system identification model function library is constructed based on AI technology, and a dynamic model of a complex industrial system or equipment is identified through industrial big data; a general system modeling simulation environment is provided, a linear dynamic system and an artificial intelligence nonlinear dynamic model are incorporated into a unified modeling simulation system, visual combination modeling and industrial system simulation are carried out, and identification modeling simulation integration is supported; and the application is used for industrial system identification modeling simulation technology required by industrial internet and industrial intelligent construction, and provides a reliable system model for design of a high-end controller.

Description

Technical Field

[0001] This invention relates to the fields of industrial automation and intelligent technology, and in particular to a general system modeling and simulation platform for the dynamic mechanisms of complex industrial systems or equipment. Background Technology

[0002] Currently, the construction of the Industrial Internet and industrial intelligence is booming, creating an urgent need for an independent and controllable set of industrial system mechanism modeling and simulation software and technologies to improve product design and achieve lean quality control, intelligent and unmanned control in production. Currently, mainstream industrial system modeling and simulation software includes MATLAB / SIMULINK, ProModel, and SIMIT. These software programs excel at modeling classic control systems but have limited support for emerging technologies such as industrial big data and artificial intelligence.

[0003] Chinese invention patent CN201711318215 discloses an industrial process simulation system and simulation method, which can be used to model by engineering parameters input on the terminal device, or by collecting the output control signals of the actual control system and the response signals of the industrial process system, and then using the Model Parameter Recognition Model (ARX) in the terminal device for identification, modeling and process simulation.

[0004] Chinese invention patent application CN 114297946A discloses an industrial internet platform for reducing the order of multidisciplinary simulation models. It can collect simulation result data and train models according to business needs, create a model reduction modeling process for multidisciplinary simulation, adapt to more disciplinary simulation scenarios, and provide a model reduction modeling solution for complex industrial simulation scenarios.

[0005] Chinese invention patent application CN111263327 discloses a modeling method and simulation platform for a single-loop model with multiple inputs and multiple outputs, used to establish a signal propagation model when a moving car passes a base station. This patent is based on existing software and has built a modeling and simulation platform for a specific problem.

[0006] The aforementioned patent applications either only allow modeling and simulation under conditions of production shutdown and disconnection of the original closed-loop circuit, making them difficult for manufacturers to accept; or the identified process models are limited to simple ARX linear models, making it difficult to handle complex process modeling; or they focus on order reduction and co-simulation of existing multidisciplinary models; none of them have universality. Summary of the Invention

[0007] To address the aforementioned existing problems, this invention fully utilizes the big data generated during normal industrial production processes, integrates artificial intelligence technology with traditional control theory, and constructs an industrial system mechanism modeling and simulation platform. It provides a general-purpose industrial modeling and simulation environment that incorporates continuous linear dynamic systems, discrete linear dynamic systems, continuous nonlinear state-space equations based on artificial intelligence technology, and combinatorial discrete nonlinear dynamic neural network models into a unified system for visual combined modeling and system simulation. It also provides an integrated working mode for system identification, modeling, and simulation.

[0008] The industrial system mechanism modeling and simulation platform proposed in this invention includes: a unified modeling and simulation code system based on Python, an AI-based system identification model function library, and a general system modeling and simulation environment.

[0009] The unified system modeling and simulation code system based on the Python language includes a scientific computing library and a control function library, which are used to support the construction of system identification model library and modeling and simulation functions. The control function library supports model expression of linear / nonlinear dynamic systems, supports state space and transfer function calculations, supports basic functions such as solving differential / partial differential equations, nonlinear optimization, and stability analysis.

[0010] The AI-based system identification model function library is built on the aforementioned unified system modeling and simulation code system based on Python, and supports the identification of linear dynamic systems and nonlinear dynamic systems. Linear dynamic systems include continuous and discrete forms, specifically including continuous transfer functions with multiple inputs and single outputs, discrete transfer functions, and state-space equations.

[0011] The linear dynamic system identification model function library constructs various offline and online adaptive identification functions through a reusable computational unit layer. The specific construction steps are as follows: S201: For continuous transfer functions, the reusable computational unit layer includes spectrum analysis, state variable filtering, and the SK iterative algorithm. Based on this, frequency response estimation, transfer function identification, and noise spectrum estimation identification functions are constructed. S202: For state-space equations, reusable computational unit layers can be used to abstract computational units including orthogonal projection operations, Hankel matrix operations, and generalized observable matrix operations. Then, N4SID, MOESP, and CVA subspace identification functions, as well as PARSIM and IVM improved subspace identification functions, are constructed. S203: For discrete transfer functions, the computational unit layer can be reused to extract computational units for z-polynomial operations, diophantine equation solving, and abstract recursive information, and construct identification functions such as ARMAX, OE, auxiliary variables, and generalized least squares. S204: Online adaptive identification algorithms for various forms of discrete transfer functions and state-space equations are constructed using abstract recursive innovation units.

[0012] The state-space equations and discrete transfer function models also support the iterative identification method of PEM series automatic differentiation. The gradient descent method is used to optimize the identification parameters, and PyTorch is selected as the automatic differentiation optimization solver.

[0013] The state-space equations identify optimization parameters according to the model definition, supporting batch data identification. The specific steps of the gradient optimization method for discrete transfer function models are as follows: S211: For the discrete transfer function form of z (taking a two-input example): (1) in, For output, , All are inputs; z is the discrete transfer function operator. All are z-polynomial expressions with known order but unknown coefficients; , , , , ; , , , They are polynomials The order of; , , These are the polynomial coefficients.

[0014] After z-polynomial operations, it can be simplified to: (2) S212: As described above The expression for any discrete transfer function at time t can be expressed as an algebraic weighted sum of the inputs and outputs at past times, with the weighting coefficients derived from a polynomial. Uniquely certain; S213: Using the above weighting coefficients as an unknown parameter vector, calculate the predicted y value of the model by taking the inner product with the vector formed by the input and output data at the corresponding time. S214: Use PyTorch to minimize the average error between the prediction and the actual value, and identify and solve for the unknown parameters; S215: During the optimization process, the unknown parameters of the discrete transfer function model will have their gradients automatically calculated and iterative optimization performed.

[0015] The nonlinear dynamic system identification model function library utilizes artificial intelligence technology to model the dynamic nonlinear dynamic systems of complex industrial processes. It provides parallel accelerated identification functions for continuous nonlinear state-space equations, as well as composable discrete nonlinear dynamic neural network models and training methods.

[0016] For the parallel accelerated identification function of continuous nonlinear state-space equations, we avoid stepwise cyclic numerical integration and use Tensor vector matrix operations to directly calculate the adjacent state difference matrix for all time steps in one go. The specific steps are as follows: S2211: Given the known structure of the differential equation model, construct the state derivative vector containing unknown parameters. In the case of an unknown differential equation structure, a neural network is constructed to approximate the state derivative vector, so as to... If a state matrix has two dimensions, time and state order, then the definition of the difference matrix between two adjacent states is: Using the fourth-order Runge-Kutta formula, the state difference matrix between adjacent steps at all time points is calculated in one step through the derivative function model. ;in, Input to the model, These are model parameters; S2212. Given the structure of the differential equation, construct an output function with unknown parameters. Otherwise, construct a neural network and calculate the model output. ,in, Input to the model, For model parameters, measure the actual output values. Then the objective function is , , For adjustable weights, Mean square error; S2213. Determine the number of iteration steps, automatically calculate the partial derivatives of the unknown parameters with respect to the target using PyTorch, and update the parameters or network weights.

[0017] This invention provides a composable discrete nonlinear dynamic neural network model and a training method. The composable discrete nonlinear dynamic neural network model includes static neural network primitives, dynamic neural network primitives, impulse response convolutional network primitives, and temporal interface components. It can construct a process dynamic neural network model by flexibly combining discrete transfer function models with gradients (described in S215) and provides a unified training method.

[0018] Dynamic neural network models for discrete systems that can be constructed include classic Hammerstein-Wiener nonlinear models and NARMAX models, including but not limited to indirect dynamic nonlinear neural networks. Their characteristic is that the input signal does not directly have a dynamic relationship with the output; instead, an intermediate quantity calculated through an unknown function forms a nonlinear dynamic correlation with the output. Such models are more common in process modeling. Because gradients can propagate at all stages, these models are trained within the PyTorch deep learning framework.

[0019] The aforementioned general-purpose system modeling and simulation environment integrates continuous linear dynamic systems, discrete linear dynamic systems, and continuous nonlinear state-space equations and composable discrete nonlinear dynamic neural network models based on artificial intelligence technology into a unified system for visual combined modeling and system simulation. The common component library within the modeling environment includes components such as signal sources, dataset field selectors, adders, delayers, filters, and user-defined functions; the models to be identified are also encapsulated as components, and after determining the model parameters, industrial systems are modeled using a visual drag-and-drop approach.

[0020] The system simulation mainly includes the following steps: S301: Analyze the signal flow and graph network of the modeling system, obtain the connection relationships between models, and calculate the input, output, and adjacency matrices; S302: Analyze linear and nonlinear dynamic system models and convert them into a unified state-space equation or direct-form II transfer function form. Estimate the initial state of each model using input and output data and align the delay. S303: The system simulation is performed using sequential modular simulation technology oriented towards graph structure. It uses numerical integration to solve differential equations, difference equations, and differential algebraic equations (DAE), calculates the state matrix, buffers the output response, and handles algebraic ring problems.

[0021] The system modeling and simulation environment described is a cloud-based platform integrating identification, modeling, and simulation. The system identification function is encapsulated as model components, which are combined and modeled via drag-and-drop. Its inputs and outputs are connected to field selectors of other model components or datasets. Automatic forward and reverse data extraction calculates the input and output data required for the identification function and completes the identification process. After completing model identification for all stages, the field selectors are ignored, and the dynamics of the entire system are simulated, calculating the simulation outputs of all models.

[0022] In the aforementioned automatic forward and reverse data extraction function, the input and output data required for identification components are automatically extracted from the system graph network model through forward and reverse recursive operations. The main steps of its forward recursive operation are as follows: S311. If the current node is a field selector, return the time series data of the specified field and range in the dataset. S312. Find all input nodes adjacent to the current node, and call the forward recursive operation function on each input node to obtain its output sequence data; S313. If the current node is the link to be identified, return the dynamic output sequence data of all input nodes; otherwise, concatenate the dynamic output sequence data of all input nodes and calculate the forward dynamic output of the current node. The main steps of its reverse recursive operation are as follows: S314. If the current node is a field selector, return the time series data of the specified field and range in the dataset. S315. Find all output nodes adjacent to the current node, and recursively call the reverse operation function on each output node to obtain the sequence data of its entry branch (from the current node to a certain output node); S316. Find all input nodes adjacent to the current node, and call the forward recursive operation function on each input node to obtain its output sequence data; S317. If the current node is the link to be identified, return all output sequence data; otherwise, concatenate all output and input sequence data except for the entry branch, calculate the reverse dynamic output of the current node, and obtain the time series data of the entry branch.

[0023] The platform supports batch data identification. Multiple batches of data can be selected as training and test sets for identification. Then, the automatic forward and reverse data extraction function is activated to obtain the required input and output data.

[0024] The integrated identification, modeling, and simulation also includes process closed-loop identification during system simulation, with the specific steps as follows: S321: Identify the models of all control elements in a closed-loop system; S322: Use an open-loop method to identify the process model as the initial value for the model; S323: Perform a simulation of the closed-loop system, that is, calculate the output of the process model at the current moment, then feed it back to all control links, calculate the control output, and then calculate the model output at the next moment, until the simulation of all time points is completed; S324: Compare the input and output of the simulation of the process steps with the actual sampled process input and output, use the average error as the objective function, and use PyTorch to automatically calculate the partial derivatives of the objective function with respect to the process parameters; S325: Update the process parameter of one step, return to step S323, until the process parameter does not change significantly.

[0025] The closed-loop system identification method proposed in this patent examines the actual effect of the process model on feedback control in closed-loop system simulation and iteratively updates the process parameters in global system optimization. The process identified in this way is more consistent with global dynamics and physical characteristics, rather than just numerical fitting of local process links.

[0026] The beneficial effects of this invention are as follows: By integrating artificial intelligence technology with traditional control theory, this invention constructs an industrial system mechanism modeling and simulation platform, which solves the problems of industrial system identification required for the construction of industrial Internet and industrial intelligence, dynamic model construction required for the localization of high-end equipment, and provides a reliable system model for the design of high-end controllers.

[0027] Compared with existing technologies, this invention provides a complete set of dynamic modeling technologies through a platform. In particular, for complex nonlinear dynamic systems, it makes full use of artificial intelligence modeling and optimization technology to construct more flexible nonlinear models, and provides integrated identification, modeling and simulation technology and parallel acceleration methods to adapt to complex and ever-changing physical process modeling scenarios. Attached Figure Description

[0028] Figure 1 This is a diagram showing the overall architecture of the industrial system mechanism modeling and simulation platform proposed in this invention. Figure 2 This is a schematic diagram of the linear dynamic system identification model function library in this invention; Figure 3 This is a schematic diagram of online adaptive recursive identification. Figure 4 An example of the accuracy of the CVA subspace method for state-space equations; Figure 5 An example of the accuracy of the iterative optimization identification method for state-space equations; Figure 6 This illustrates parallel accelerated training of nonlinear state-space equations. Figure 7 This is a schematic diagram of the basic elements of a static neural network. Figure 8 This is a schematic diagram of the basic elements of a dynamic neural network. Figure 9 This is a schematic diagram of the basic elements of an impulse response convolutional network. Figure 10 This is a schematic diagram of a timing interface component; Figure 11 A schematic diagram of the Hammerstein-Wiener and deformed nonlinear models; Figure 12 This is a schematic diagram of an indirect dynamic nonlinear neural network. Figure 13 One of the schematic diagrams of the system modeling and simulation environment interface; Figure 14The second illustration of the system modeling and simulation environment interface; Figure 15 This is a schematic diagram of the system simulation interface; Figure 16 This is a diagram of the interface for selecting and configuring training parameters for multiple batches of datasets. Detailed Implementation

[0029] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0030] The present invention proposes an industrial system mechanism modeling and simulation platform, comprising the following steps: I. Establishing a unified system modeling and simulation code system based on the Python language To perform complex modeling, simulation, and control calculations, a complete and usable scientific computing environment is required. Using the Python Sciences library (including NumPy, ATLAS, and BLAS), basic functions such as calculus, matrix operations, signal processing, and machine learning can be performed. Simultaneously, the Python language can fully leverage artificial intelligence technologies for industrial mechanism modeling and simulation.

[0031] Furthermore, this platform provides a Python-based control function library that supports fundamental functions including model representation of linear / nonlinear dynamic systems, state-space and transfer function calculations, solving differential / partial differential equations, nonlinear optimization, and stability analysis. Subsequent system identification and simulation are based on this codebase.

[0032] II. Constructing a System Identification Function Library To meet the system modeling requirements of complex industrial systems, it is necessary to establish a rich identification function library to support the identification of differential equations for various forms of linear and nonlinear dynamic systems.

[0033] The system identification function library is built on the Python unified system modeling and simulation code system, such as Figure 2 Linear dynamic systems are mainly used for identifying differential equation models in the form of continuous transfer functions, state spaces, and discrete forms with multiple inputs and single outputs.

[0034] The linear dynamic system identification model function library has a reusable computational unit layer, enabling the construction of various offline and online adaptive identification functions. The specific construction steps are as follows: S201: For continuous transfer functions, the computational unit layer can be reused to complete spectrum analysis, state variable filtering, and SK iterative algorithm. On this basis, frequency response estimation, transfer function identification, and noise spectrum estimation identification functions are constructed. S202: For the state-space equations, reusable computational unit layers can be used to abstract computational units such as orthogonal projection operations, Hankel matrix operations, and observable matrix operations. Then, N4SID, MOESP, and CVA subspace identification functions, as well as PARSIM and IVM improved subspace identification functions, are constructed. S203: For discrete transfer functions, the reusable computational unit layer can extract computational units such as z-polynomial operations, diophantine equation solving, and abstract recursive information, and construct identification functions such as ARMAX, OE, auxiliary variables, and generalized least squares. S204: Online adaptive identification algorithms for various forms of discrete transfer functions and state-space equations are constructed using abstract recursive innovation units.

[0035] The basic form of the diophantine equation in S203 is: ,in, Given a discrete transfer function polynomial, solve for the unknown polynomial in the equation. It can be used to construct the optimal prediction model form of discrete transfer functions.

[0036] The abstract recursive innovation unit in S203 is constructed using the following steps, where the input and output data vectors are: Forgetting factor is The estimated model parameters are The covariance matrix is ​​P: S2041: Calculate the gain matrix at the current time k. : ; S2042: Update model parameters using innovation and gain. : ; S2043: Record one-step forecast and error : , ; This is the actual output; S2044: Update the covariance matrix at the current time step. : I is the identity matrix; S2045: Return to step 2041 to perform parameter identification at the next time step.

[0037] The abstract recursive innovation unit does not limit the specific form of the discrete transfer function, but rather represents a general abstraction of the online adaptive identification process. During identification, data vectors are extracted based on different transfer function forms. .

[0038] For example, for the ARX model ARMAX model, Finite impulse response model And so on. Where y and u are the output and input, respectively. This is the error.

[0039] The main steps of the offline subspace identification method for the state-space equations in S202 are as follows: S2021: Calculate the Hankel matrix of the input and output data. , Output the Hankel matrix of Y Data representing future predictions can be derived from past Y and future Y. Linear representation; S2022: Through calculation Towards The orthogonal projection is performed, and different weighting matrices are selected for left and right multiplication based on the different algorithms of CVA, N4SID, and MOESP. S2023: Perform SVD or QR decomposition on the matrix to obtain the state sequence matrix. and generalized observability matrix The coefficient matrices A, B, C, D, and K of the state-space equations are obtained by solving the overdetermined system of equations.

[0040] In S204, the adaptive recursive identification of the state-space equations is generally used for predictive control, employing a multi-step predictive model. in, It is the prediction vector that moves forward q steps continuously at time k. = It is the output and input data vectors of the past p time steps, and u is the input data vector for the next q time steps. The corresponding coefficient matrices are related to the coefficient matrices of the state-space equations.

[0041] Let the parameter matrix after merging the coefficient matrices be: The data vector is .

[0042] Due to its linearity, the construction steps of the abstract recursive innovation unit in S203 are substituted into the online adaptive recursive identification of the state-space equation prediction model. During identification, the model prediction value... Compared with the actual output data vector Alignment time, In reality, the input data has a time range of [kq,k], and the prediction will be performed using the future input of [k,k+q].

[0043] Online adaptive parameter identification of discrete transfer function in S204, such as Figure 3 As shown, the platform performed continuous online parameter identification on multiple batches of data. The multiple batches of data are separated by arrows, and the parameter identification and convergence status are displayed at the top.

[0044] Furthermore, the state-space equations and discrete transfer function models also support the iterative method identification of PEM series automatic differentiation, and the identification parameters are optimized using the gradient descent method, with PyTorch selected as the automatic differentiation optimization solver.

[0045] The state-space equations directly identify and optimize parameters according to the model definition, and also support multi-batch training. The specific steps of the gradient optimization method for discrete transfer function models are as follows: S211: For the discrete transfer function form of z (taking a two-input example): (1) in, For output, , All are inputs; z is the discrete transfer function operator. All are z-polynomial expressions with known order but unknown coefficients; , , , , ; , , , They are polynomials The order of , , Let be the polynomial coefficients; after z-polynomial operations, it can be simplified to: (2) S212: As described above The expression for any discrete transfer function at time t can be expressed as an algebraic weighted sum of the inputs and outputs at past times, with the weighting coefficients derived from a polynomial. Uniquely certain; S213: Using the above weighting coefficients as an unknown parameter vector, calculate the predicted y value of the model by taking the inner product with the vector formed by the input and output data at the corresponding time. S214: Use PyTorch to minimize the average error between the prediction and the actual value, and identify and solve for the unknown parameters; S215: During the optimization process, the unknown parameters of the discrete transfer function model will have their gradients automatically calculated and iterative optimization performed.

[0046] Platform execution of state-space equations CVA subspace method ( Figure 4 ) and iterative optimization identification method ( Figure 5 Case accuracy comparison, such as Figure 4 and Figure 5 This demonstrates the performance of the identified model across different batches of data, which helps in finding models with stronger generalization capabilities in industrial big data. Note: Figure 4 and Figure 5 In the diagram, the horizontal axis represents the sampling points, the vertical axis represents the variable values ​​(depending on the specific scheme), the dashed line is the fitted curve, and the solid line is the true curve.

[0047] Modeling the dynamic systems of complex industrial processes often employs nonlinear dynamic system identification methods. This platform fully leverages the advantages of artificial intelligence in modeling nonlinear dynamic systems, providing a parallel accelerated identification function for continuous nonlinear state-space equations (Equation (3)): S221: If the structure of the process differential equation model is known, construct a state derivative vector containing unknown parameters. , See formula (3), where, Let U be the state, U be the input, and Y be the output. The derivative function representing the state. It is the output function. For parameters, a parallel acceleration method is used to identify unknown parameters; S222: When the process differential equation model is too complex or unknown, construct a neural network to approximate the unknown derivative function equation and parameters, and use parallelization methods to accelerate it. ; (3) S223: When the process differential equation model is too complex or unknown, discrete system modeling can also be used. The system identification library provides static neural network primitives, dynamic neural network primitives, impulse response convolutional network primitives, and time-series interface components. Complex process dynamic neural network models can be constructed by combining the discrete transfer function model with gradient in S215, and a unified training method is provided.

[0048] The parallelization acceleration method is proposed for the following reasons: For the identification of continuous nonlinear dynamic systems, the usual approach is to identify unknown parameters by numerical integration, which is very slow. This is because solving continuous differential equations is a step-by-step serial recursion, which requires iterative calculation of all states through loops.

[0049] The parallel accelerated identification method for continuous nonlinear dynamic systems proposed in this invention avoids stepwise cyclic integration and uses Tensor vector matrix operations to directly calculate the adjacent state difference matrix for all time steps in one go.

[0050] Taking the 4th-order Runge-Kutta as an example, using the state vector matrix at all times... (The two dimensions are time and state order, respectively) and the input matrix The state difference matrix between adjacent steps at all times can be calculated directly. Then, fit the definition of the state difference. That's it. The specific calculation formula is as follows: , , , , , (4) In the above formula It can be a vector of derivative functions with known structure but unknown parameters, or it can represent a neural network with unknown weights. The derivative vector calculated for interpolation. For parameters.

[0051] This platform selects PyTorch as the automatic differential solver to iteratively solve for network weights or unknown parameters. If the state... Even though it is unmeasurable, the automatic differentiation framework can still use it as an intermediate vector with gradients (Tensor) to estimate the unknown state. In summary, the parallel algorithm steps for identifying continuous nonlinear state space equations are as follows: S2211: Given the known structure of the differential equation model, construct the state derivative vector containing unknown parameters. When the structure of the differential equation model is unknown, a neural network is constructed to approximate the state derivative function vector, and the state difference matrix between adjacent steps of the derivative model output is calculated using formula (4). The definition of the state difference matrix is ; S2212: Construct an output function with unknown parameters given the structure of the differential equation. Otherwise, construct a neural network and calculate the model output. Measure the actual output value Then the objective function is , , For adjustable weights, Mean square error; S2213: Determine the number of iteration steps, automatically calculate the partial derivatives of unknown parameters with respect to the target, and update the parameters or network weights.

[0052] Example of the platform executing the identification function to model the temperature change process in the reactor. Figure 6 As shown in the figure, the temperature change process of the reactor is fitted by approximating the state derivative function.

[0053] This invention provides a composable discrete nonlinear dynamic neural network model and a training method. The composable discrete nonlinear dynamic neural network model includes static neural network primitives, dynamic neural network primitives, impulse response convolutional network primitives, and temporal interface components. It can construct a process dynamic neural network model by flexibly combining discrete transfer function models with gradients (described in S215) and provides a unified training method.

[0054] Static neural network primitives, such as Figure 7 As shown, this is a neural network where the input, output, and number of hidden layer units are all configurable, modeling the nonlinear relationship between input and output at any given time. Dynamic neural network primitives, such as... Figure 8 As shown, this is also a neural network where the number of input, output, and hidden layer units are configurable. It models the nonlinear dynamic relationship between input and output at adjacent time steps. When noise e is added as input, it can characterize the NARMAX model. The impulse response convolutional network primitives, such as... Figure 9 As shown, impulse response sequences are identified using 1D convolutional kernels, and nonlinearity can be increased using nonlinear activation functions. Timing interface components, such as... Figure 10 As shown, the output sequence of the cached model serves as the input for the sampling of neighboring time steps in subsequent models. It is an intermediate Tensor vector that can backpropagate gradients during the identification process.

[0055] By combining neural network primitives and temporal interface components, along with discrete transfer function models with accompanying gradients, complex process dynamic neural network models can be constructed in a visual drag-and-drop manner within the modeling environment.

[0056] like Figure 11 It can construct classic Hammerstein-Wiener nonlinear models, where the discrete transfer function with gradient can be any linear model such as ARX, ARMAX, or OE; it can also construct more flexible dynamic neural network models through connections, such as... Figure 11The dashed line represents a feedthrough between a certain input and the dynamic output. It can be used to construct indirect dynamic nonlinear neural networks, such as... Figure 12 The characteristic is that the input signal does not directly have a dynamic relationship with the output. Instead, the intermediate quantity obtained by calculating through a complex and unknown function has a nonlinear dynamic relationship with the output. This type of model is more common in process modeling.

[0057] The training of the nonlinear dynamic neural network in the combination process aims to fit the target time series data y(k). Since gradients can be propagated in all stages, this is done in the PyTorch deep learning framework.

[0058] III. System Modeling and Simulation The general system modeling and simulation environment provided by this invention incorporates continuous linear dynamic systems, discrete linear dynamic systems, continuous nonlinear state-space equations based on artificial intelligence technology, and combinatorial discrete nonlinear dynamic neural network models into a unified system for visual combinatorial modeling and system simulation.

[0059] The common component library in the modeling environment includes components such as signal sources, field selectors for datasets, adders, delayers, filters, and user-defined functions; the model to be identified is also encapsulated as a component, and after the model parameters are determined, the industrial system is modeled in a visual drag-and-drop manner.

[0060] System simulation mainly includes the following steps: S301: Analyze the signal flow and graph network of the modeling system, obtain the connection relationships between models, and calculate the input, output, and adjacency matrices; S302: For linear and nonlinear dynamic system models, convert them into a unified state-space equation or direct-form II transfer function form, estimate the initial state of each model using input and output data, and align the delays; S303: The system simulation is performed using sequential modular simulation technology oriented towards graph structure. It uses numerical integration to solve differential / difference equations and differential-algebraic equations (DAE), calculates the state matrix, buffers the output response, and handles algebraic ring problems.

[0061] System modeling and simulation environment such as Figure 13 and Figure 14 As shown, the component with the circle and gear icon is the field selector for the dataset (such as the target temperature), and the component with the stacked parallelogram icon is the encapsulated recognition model function. Figure 14 The pop-up on the right displays the platform's identification function component library, which can be dragged and dropped into the environment for composite system modeling. It should be noted that... Figure 13 and Figure 14This is a schematic diagram of the industrial system mechanism modeling and simulation platform environment proposed in this invention. The icon shapes can be clearly displayed to illustrate the environment. Other parts can be defined and designed by those skilled in the art based on the textual disclosure of this invention and are not protected by this invention.

[0062] Once all steps have been identified, the platform simulates the system, such as... Figure 15 As shown.

[0063] User-defined functions can be used in Python to express complex algebraic computation logic and differential equations.

[0064] The system identification model is encapsulated as components, which are combined and modeled by dragging and dropping. Its inputs and outputs are connected to field selectors of other model components or datasets. Automatic forward and reverse data extraction calculates the input and output data required by the identification function and executes the identification function. After completing model identification for all stages, the field selectors are ignored, and the dynamics of the entire system are simulated, calculating the simulation outputs of all models.

[0065] For the automatic forward and reverse data extraction function, the input and output data required for component identification are automatically extracted from the system graph network model through forward and reverse recursive operations. The main steps of its forward recursive operation are as follows: S311. If the current node is a field selector, return the time series data of the specified field and range in the dataset. S312. Find all input nodes adjacent to the current node, and call the forward recursive operation function on each input node to obtain its output sequence data; S313. If the current node is the link to be identified, return the dynamic output sequence data of all input nodes; otherwise, concatenate the dynamic output sequence data of all input nodes and calculate the forward dynamic output of the current node; the main steps of its reverse recursive operation are as follows: S314. If the current node is a field selector, return the time series data of the specified field and range in the dataset. S315. Find all output nodes adjacent to the current node, and recursively call the reverse operation function on each output node to obtain the sequence data of its entry branch (from the current node to a certain output node); S316. Find all input nodes adjacent to the current node, and call the forward recursive operation function on each input node to obtain its output sequence data; S317. If the current node is the link to be identified, return all output sequence data; otherwise, concatenate all output and input sequence data except for the entry branch, calculate the reverse dynamic output of the current node, and obtain the time series data of the entry branch.

[0066] The platform supports batch data identification, allowing users to select multiple batches of data as training and test sets for identification, such as... Figure 16 As shown, the automatic forward and reverse data extraction function is then activated to obtain the required input and output data, and the identification results are as follows. Figure 4 and Figure 5 As shown.

[0067] The integration of identification, modeling, and simulation is also reflected in the identification of closed-loop systems during system simulation. Industrial sites primarily generate data from closed-loop systems; using only open-loop methods to identify complex processes within closed loops can lead to problems such as insufficient modal excitation, bias, and overfitting. This invention proposes an iterative method for process closed-loop identification in system simulation, with the following specific steps: S321: Identify the models of all control elements in a closed-loop system; S322: Use an open-loop method to identify the process model as the initial value for the model; S323: Perform a simulation of the closed-loop system, that is, calculate the output of the process model at the current moment, then feed it back to all control links, calculate the control output, and then calculate the model output at the next moment, until the simulation of all time points is completed; S324: Compare the input and output of the simulation of the process steps with the actual sampled process input and output, use the average error as the objective function, and use PyTorch to automatically calculate the partial derivatives of the objective function with respect to the process parameters; S325: Update the process parameter of one step, return to step S323, until the process parameter does not change significantly.

[0068] It is evident that this method verifies the actual effect of the process model on feedback control in closed-loop system simulation and iteratively updates the process parameters in global system optimization. The process identified in this way is more consistent with global dynamics and physical characteristics, rather than just numerical fitting of local process links.

[0069] The implementation effect of this patent: Using this platform, a mechanism model of an imported tobacco drying machine for a certain tobacco company was created. Industrial big data was used to identify control links and process models, and a system model of the equipment was constructed, such as... Figure 4 , 5 As shown in Figures 14 and 15, simulation verification was performed. Figure 16 Based on this, a corresponding intelligent control system was developed, enabling the average cooling water content of each batch of products to meet the average control accuracy requirements by 99%, with an average deviation of 0.036, exceeding the manual average control level of 0.042. This achieved unmanned operation, giving the company a leading position in the industry, significantly improving economic efficiency, and promoting the transformation of industrial automation to industrial intelligence.

Claims

1. An industrial system mechanism modeling and simulation platform, characterized in that, The platform includes: a unified modeling and simulation code system based on Python, a system identification model function library based on AI, and a general system modeling and simulation environment; The unified system modeling and simulation code system based on Python is used to support the construction of system identification model library and modeling and simulation functions, including scientific computing library and control function library; The AI-based system identification model function library identifies the dynamic models of complex industrial systems or equipment through industrial big data, including identification models and functions for continuous linear dynamic systems and discrete linear dynamic systems; it includes parallel accelerated identification functions for continuous nonlinear state-space equations based on artificial intelligence technology; and it includes composable discrete nonlinear dynamic neural network models and training methods. The general system modeling and simulation environment integrates continuous linear dynamic systems, discrete linear dynamic systems, and continuous nonlinear state-space equations and combinatorial discrete nonlinear dynamic neural network models based on artificial intelligence technology into a unified system for visual combined modeling and system simulation; and integrates identification, modeling, and simulation, performing local component identification in the modeling environment and closed-loop process identification in the system simulation. The parallel accelerated identification function for the continuous nonlinear state-space equations avoids stepwise cyclic numerical integration and utilizes Tensor vector matrix operations to calculate the adjacent state difference matrix for all time steps in one go. The specific steps are as follows: S2211. Given the structure of the differential equation, construct a state derivative vector containing unknown parameters. In the case of an unknown differential equation structure, a neural network is constructed to approximate the state derivative vector, so as to... Let the state matrix represent a state matrix with two dimensions: time and state order. Then, the definition of the state difference matrix between two adjacent steps is: Using the fourth-order Runge-Kutta formula, the state difference matrix between adjacent steps at all time points is calculated in one step through the derivative function model. ;in, Input to the model, These are model parameters; S2212. Given the structure of the differential equation, construct an output function with unknown parameters. Otherwise, construct a neural network and calculate the model output. ,in, Input to the model, For model parameters, measure the actual output values. Then the objective function is , , For adjustable weights, Mean square error; S2213. Determine the number of iteration steps, automatically calculate the partial derivatives of the unknown parameters with respect to the target using PyTorch, and update the parameters or network weights. The general system modeling and simulation environment described above mainly includes the following steps for system simulation: S301. Analyze the signal flow and graph network of the modeling system, obtain the connection relationships between models, and calculate the input, output and adjacency matrix; S302. Analyze the linear and nonlinear dynamic system models and convert them into a unified state-space equation or direct-form II transfer function form. Estimate the initial state of each model using input and output data and align the delay. S303. System simulation is performed using sequential modular simulation technology oriented towards graph structure. Numerical integration is used to solve differential equations, difference equations, and differential-algebraic equations, calculate the state matrix, buffer the output response, and handle algebraic ring problems.

2. The industrial system mechanism modeling and simulation platform according to claim 1, wherein in the AI-based system identification model function library, continuous linear dynamic systems and discrete linear dynamic systems constitute a linear dynamic system, specifically including the identification of multi-input single-output continuous transfer functions, discrete transfer functions, and state-space equation models, characterized in that, The platform constructs offline and online adaptive identification functions through a reusable computing unit layer. The construction steps include: S201. For continuous transfer functions, the reusable computational unit layer includes spectrum analysis, state variable filtering, and SK iterative algorithm. Based on this, frequency response estimation, transfer function identification, and noise spectrum estimation identification functions are constructed. S202. For state-space equations, reusable computational unit layers can be used to abstract computational units including orthogonal projection operations, Hankel matrix operations, and generalized observable matrix operations; N4SID, MOESP, and CVA subspace identification functions, as well as PARSIM and IVM improved subspace identification functions can be constructed. S203. For discrete transfer functions, reusable computational unit layers can be used to extract computational units including Z-polynomial operations, diophantine equation solving, and abstract recursive information; ARMAX, OE, auxiliary variables, and generalized least squares identification functions can be constructed. S204. Online adaptive identification algorithms for various forms of discrete transfer functions constructed using abstract recursive information units, and online adaptive identification algorithms for state-space equations.

3. The industrial system mechanism modeling and simulation platform according to claim 2, characterized in that, In the AI-based system identification model function library, the identification and training are completed based on the PyTorch automatic differentiation deep learning framework, which models the dynamic nonlinear system of industrial processes using artificial intelligence technology, provides parallel accelerated identification functions for continuous nonlinear state-space equations, and provides composable discrete nonlinear dynamic neural network models and training methods.

4. The industrial system mechanism modeling and simulation platform according to claim 3, characterized in that, The composable discrete nonlinear dynamic neural network model includes static neural network primitives, dynamic neural network primitives, impulse response convolutional network primitives, and temporal interface components. It constructs a dynamic neural network model by combining discrete transfer function models with gradients and provides a unified training method.

5. The industrial system mechanism modeling and simulation platform according to claim 1, characterized in that, The integration of identification, modeling, and simulation, and the identification of local components within the modeling environment, mainly includes the following steps: The system identification model is encapsulated as a model component, which is combined and modeled by dragging and dropping. Its input and output are connected to the field selectors of other model components or datasets. The input and output data required by the identification function are calculated through the automatic forward and reverse data extraction function to complete the identification. After completing the model identification for all stages, the field selector is ignored, and the dynamics of the entire system are simulated to calculate the simulation output of all models.

6. The industrial system mechanism modeling and simulation platform according to claim 5, characterized in that, In the automatic forward and reverse data extraction function, the input and output data required for identification components are automatically extracted from the system graph network model through forward and reverse recursive operations. The main steps of its forward recursive operation are as follows: S311. If the current node is a field selector, return the time series data of the specified field and range in the dataset. S312. Find all input nodes adjacent to the current node, and call the forward recursive operation function on each input node to obtain its output sequence data; S313. If the current node is the link to be identified, return the dynamic output sequence data of all input nodes; otherwise, concatenate the dynamic output sequence data of all input nodes and calculate the forward dynamic output of the current node. The main steps of its reverse recursive operation are as follows: S314. If the current node is a field selector, return the time series data of the specified field and range in the dataset. S315. Find all output nodes adjacent to the current node, and recursively call the reverse operation function on each output node to obtain the sequence data of its entering branch; S316. Find all input nodes adjacent to the current node, and call the forward recursive operation function on each input node to obtain its output sequence data; S317. If the current node is the link to be identified, return all output sequence data; Otherwise, concatenate all output and input sequence data except for the incoming branch, calculate the reverse dynamic output of the current node, and obtain the time series data of the incoming branch.

7. The industrial system mechanism modeling and simulation platform according to claim 5, characterized in that, The integrated identification, modeling, and simulation approach performs closed-loop process identification within system simulation. The specific steps are as follows: S321. Identify the models of all control elements in the closed-loop system; S322. Use the open-loop method to identify the process model as the initial value of the model; S323. Perform one round of simulation on the closed-loop system, that is, calculate the output of the process model at the current moment, then feed it back to all control links, calculate the control output, and then calculate the model output at the next moment, until the simulation of all time points is completed. S324. Compare the input and output of the simulation process with the actual sampled process input and output, use the average error as the objective function, and use PyTorch to automatically calculate the partial derivatives of the objective function with respect to the process parameters. S325. Update the process parameters of one step and return to step S323 until the process parameters no longer change significantly.

Images