Industrial Manufacturing Process and Production Operation and Maintenance Optimization Method and System Based on Digital Twin

Abstract

The present application relates to the technical field of data processing, and discloses an industrial manufacturing process and production operation and maintenance optimization method and system based on digital twin. The method includes: performing multi-physical field coupling modeling on pre-collected industrial production process data to obtain an industrial digital twin model; performing dynamic sensitivity analysis on the industrial digital twin model to obtain a time-series parameter sensitivity map; performing hybrid optimization on the time-series parameter sensitivity map to obtain a multi-objective process strategy; performing Monte Carlo simulation on the multi-objective process strategy to obtain a strategy risk assessment matrix; performing adaptive control analysis according to the strategy risk assessment matrix to obtain real-time optimization effect data; inputting the real-time optimization effect data into a pre-set graph neural network for multi-source data correlation analysis to obtain an industrial manufacturing knowledge graph, and generating a target optimization strategy according to the industrial manufacturing knowledge graph. The present application improves the efficiency of industrial manufacturing process and production operation and maintenance optimization based on digital twin.

Description

Technical Field

[0001] This application relates to the field of data processing, and in particular, to an industrial manufacturing process and production operation and maintenance optimization method and system based on digital twin. Background Art

[0002] The optimization of industrial manufacturing processes has always been the key to improving production efficiency, product quality, and resource utilization. Traditional optimization methods mainly rely on static models and empirical rules, and improve production processes through the analysis of historical data and the application of expert knowledge. With the development of digital technologies, advanced technologies such as digital twin, artificial intelligence, and big data analysis have been introduced into the field of industrial manufacturing, providing new ideas and tools for optimization.

[0003] However, there are still some deficiencies in existing optimization methods. First, most methods are difficult to comprehensively consider the complex coupling relationships between multiple physical fields, resulting in insufficient model accuracy. Second, traditional methods often ignore the dynamic change characteristics of parameter sensitivity and cannot adapt to time-varying factors in the production process. In addition, when dealing with multi-objective optimization problems, existing methods often have difficulty finding a reasonable balance among multiple objectives such as efficiency, quality, and cost. Finally, many methods lack the risk assessment and adaptive adjustment capabilities of optimization strategies and are difficult to cope with complex and changeable production environments. Summary of the Invention

[0004] This application provides an industrial manufacturing process and production operation and maintenance optimization method and system based on digital twin, which is used to improve the efficiency of industrial manufacturing process and production operation and maintenance optimization based on digital twin.

[0005] In a first aspect, this application provides an industrial manufacturing process and production operation and maintenance optimization method based on digital twin. The industrial manufacturing process and production operation and maintenance optimization method based on digital twin includes: performing multi-physical field coupling modeling on pre-collected industrial production process data to obtain an industrial digital twin model; performing dynamic sensitivity analysis on the industrial digital twin model to obtain a time-series parameter sensitivity map; performing hybrid optimization on the time-series parameter sensitivity map to obtain a multi-objective process strategy; performing Monte Carlo simulation on the multi-objective process strategy to obtain a strategy risk assessment matrix; performing adaptive control analysis based on the strategy risk assessment matrix to obtain real-time optimization effect data; inputting the real-time optimization effect data into a preset graph neural network for multi-source data association analysis to obtain an industrial manufacturing knowledge graph, and generating a target optimization strategy based on the industrial manufacturing knowledge graph.

[0006] Combined with the first aspect, in the first implementation manner of the first aspect of the present application, the multi-physical field coupling modeling of the pre-collected industrial production process data to obtain an industrial digital twin model includes: performing data cleaning and preprocessing on the pre-collected industrial production process data to obtain a standardized multi-source data set; performing principal component analysis and dimensionality reduction processing on the standardized multi-source data set to obtain a dimensionality-reduced feature data set; extracting features from the dimensionality-reduced feature data set through a pre-set deep autoencoder to obtain a latent feature vector; inputting the latent feature vector into a pre-trained long short-term memory network for time series modeling to obtain a time series feature representation; extracting spatial features from the time series feature representation through a convolutional neural network to obtain spatio-temporal feature mapping data; inputting the spatio-temporal feature mapping data into a multi-head attention mechanism for feature fusion to obtain multi-physical field coupling features; estimating the probability distribution of the multi-physical field coupling features through a variational inference algorithm to obtain parameter distribution data; inputting the parameter distribution data into a Monte Carlo sampling algorithm to generate multiple groups of parameter samples to obtain a parameter sampling set; performing numerical simulation on the parameter sampling set through a finite element analysis algorithm to obtain a multi-physical field simulation result; and performing error backpropagation optimization on the multi-physical field simulation result and the industrial production process data to obtain the industrial digital twin model.

[0007] Combined with the first aspect, in the second implementation manner of the first aspect of the present application, the dynamic sensitivity analysis of the industrial digital twin model to obtain a time series parameter sensitivity map includes: decomposing the parameters of the industrial digital twin model to obtain a set of key parameters; combining the key parameters through an orthogonal experimental algorithm to obtain a parameter combination scheme; performing multiple simulations on the industrial digital twin model according to the parameter combination scheme to obtain an initial simulation result set; performing simulation variance analysis on the initial simulation result set to obtain parameter main effects and interaction effects; inputting the parameter main effects and the interaction effects into a time series decomposition algorithm to obtain time-varying sensitivity indicators; performing wavelet transform processing on the time-varying sensitivity indicators to obtain multi-scale sensitivity features; inputting the multi-scale sensitivity features into an adaptive boosting algorithm to obtain a parameter importance ranking; performing hierarchical analysis on the parameter importance ranking to obtain a parameter weight matrix; inputting the parameter weight matrix into a non-linear regression algorithm to obtain a parameter response surface; and performing time series slicing and interpolation processing on the parameter response surface to obtain a time series parameter sensitivity map.

[0008] Combined with the first aspect, in the third implementation manner of the first aspect of the present application, the hybrid optimization of the timing parameter sensitivity map to obtain a multi-objective process strategy includes: dividing the timing parameter sensitivity map by time window to obtain sensitivity sub-maps for multiple time periods; performing principal component analysis on the sensitivity sub-maps for the multiple time periods to obtain a dimensionality-reduced feature vector; inputting the dimensionality-reduced feature vector into a genetic algorithm for initial population generation to obtain a candidate solution set; performing non-dominated sorting on the candidate solution set to obtain Pareto front solutions; inputting the Pareto front solutions into a particle swarm optimization algorithm for local search to obtain an optimized solution set; performing fuzzy C-means clustering analysis on the optimized solution set to obtain a process strategy cluster; inputting the process strategy cluster into a deep reinforcement learning network for strategy evaluation to obtain strategy value estimation data; performing strategy selection on the strategy value estimation data to obtain an optimal strategy subset; inputting the optimal strategy subset into a Bayesian optimization algorithm for parameter fine-tuning to obtain a target strategy plan; and correcting the target strategy plan through a multi-objective decision tree algorithm to obtain the multi-objective process strategy.

[0009] Combined with the first aspect, in the fourth implementation manner of the first aspect of the present application, the Monte Carlo simulation of the multi-objective process strategy to obtain a strategy risk assessment matrix includes: extracting parameters from the multi-objective process strategy to obtain a strategy parameter set; performing Latin hypercube sampling on the strategy parameter set to obtain a parameter sample space; inputting the parameter sample space into an industrial digital twin model for parallel simulation calculation to obtain a parallel simulation result set; performing probability distribution analysis on the parallel simulation result set to obtain performance distribution characteristics; performing smoothing processing on the performance distribution characteristics through a kernel density estimation method to obtain continuous probability density data; performing index confidence interval analysis on the continuous probability density data to obtain risk limit values; performing multi-dimensional scoring on the risk limit values to obtain a quantitative strategy risk index; performing dimensionality reduction processing on the quantitative strategy risk index through a principal component analysis method to obtain a dimensionality-reduced risk feature; performing clustering analysis on the dimensionality-reduced risk feature to obtain a risk level classification result; and associating and fusing the risk level classification result with the multi-objective process strategy to construct a strategy risk assessment matrix.

[0010] Combined with the first aspect, in the fifth implementation manner of the first aspect of the present application, the adaptive control analysis is performed according to the policy risk assessment matrix to obtain real-time optimization effect data, including: performing eigen-decomposition on the policy risk assessment matrix to obtain risk principal components; calculating the weights of the prime risk principal components through the fuzzy analytic hierarchy process to obtain a risk weight vector; constructing a risk-weighted decision tree according to the risk weight vector and the multi-objective process strategy to obtain an initial control strategy; performing robustness analysis on the initial control strategy to obtain a strategy robustness score; matching an adaptive learning rate through the strategy robustness score, and constructing a dynamic adjustment mechanism according to the adaptive learning rate to obtain an adaptive controller; integrating the adaptive controller with the industrial digital twin model and performing closed-loop simulation to obtain a control response curve; performing fast Fourier transform on the control response curve to obtain system dynamic characteristic indexes; online optimizing the parameters of the adaptive controller through the dynamic characteristic indexes to obtain optimized control parameters; performing industrial manufacturing simulation operation on the industrial digital twin model according to the optimized control parameters, and collecting simulated real-time process data; performing visualization processing on the simulated real-time process data to obtain the real-time optimization effect data.

[0011] Combined with the first aspect, in the sixth implementation manner of the first aspect of the present application, inputting the real-time optimization effect data into a pre-set graph neural network for multi-source data correlation analysis to obtain an industrial manufacturing knowledge graph, and generating a target optimization strategy according to the industrial manufacturing knowledge graph, including: performing time window segmentation on the real-time optimization effect data to obtain data subsets of multiple time periods; performing feature extraction and dimensionality reduction on the data subsets of the multiple time periods to obtain effect feature vectors; inputting the effect feature vectors into a graph convolutional neural network for spatial correlation analysis to obtain node embedding representation data; performing hierarchical clustering on the node embedding representation data and constructing a multi-level knowledge structure to obtain an initial knowledge graph; calculating the importance of the initial knowledge graph to obtain a weighted knowledge graph; performing path mining on the weighted knowledge graph to obtain a causal relationship network; fusing the causal relationship network with a preset expert rule to obtain an enhanced knowledge graph; performing graph reasoning operation on the enhanced knowledge graph to obtain a candidate optimization strategy set; evaluating and sorting the candidate optimization strategy set through a multi-objective optimization algorithm to obtain a strategy priority list; performing parameter sensitivity analysis and strategy optimization according to the strategy priority list to obtain a target optimization strategy.

[0012] In the second aspect, the present application provides an industrial manufacturing process and production operation and maintenance optimization system based on digital twin. The industrial manufacturing process and production operation and maintenance optimization system based on digital twin includes:

[0013] A modeling module for performing multi - physical - field coupling modeling on pre - collected industrial production process data to obtain an industrial digital twin model;

[0014] An analysis module for performing dynamic sensitivity analysis on the industrial digital twin model to obtain a time - series parameter sensitivity map;

[0015] An optimization module for performing hybrid optimization on the time - series parameter sensitivity map to obtain a multi - objective process strategy;

[0016] A simulation module for performing Monte Carlo simulation on the multi - objective process strategy to obtain a strategy risk assessment matrix;

[0017] A control module for performing adaptive control analysis based on the strategy risk assessment matrix to obtain real - time optimization effect data;

[0018] A generation module for inputting the real - time optimization effect data into a pre - set graph neural network for multi - source data association analysis to obtain an industrial manufacturing knowledge graph, and generating a target optimization strategy according to the industrial manufacturing knowledge graph.

[0019] In the technical solution provided by this application, a high - precision digital twin model is constructed through multi - physical - field coupling modeling. Dynamic sensitivity analysis is used to capture the time - series changes of parameter effects, and a multi - objective balance is achieved by combining a hybrid optimization strategy. Monte Carlo simulation and risk assessment in the method improve decision - making reliability, and adaptive control technology ensures real - time optimization effects. Innovatively, a graph neural network is introduced to construct an industrial manufacturing knowledge graph, realizing knowledge accumulation and intelligent decision - making. This method integrates a comprehensive data - processing process from data cleaning, dimensionality reduction, feature extraction to time - series modeling, and combines multiple advanced algorithms such as deep learning, reinforcement learning, and genetic algorithms, achieving multi - scale analysis and interpretable optimization. Through adaptive control and online tuning, the method has real - time performance and high adaptability, forming a highly integrated and automated closed - loop optimization process. This comprehensive, precise, dynamic, and intelligent optimization method can significantly improve production efficiency, product quality, and resource utilization rate, while effectively reducing operation risks and costs. Brief Description of the Drawings

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for description in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0021] Figure 1 It is a schematic diagram of an embodiment of the industrial manufacturing process and production operation and maintenance optimization method based on digital twin in the embodiments of this application;

[0022] Figure 2 This is a schematic diagram of an embodiment of the industrial manufacturing process and production operation and maintenance optimization system based on digital twin in the embodiments of the present application. Specific implementation manners

[0023] The embodiments of the present application provide a method and system for optimizing an industrial manufacturing process and production operation and maintenance based on digital twin. Terms such as "first", "second", "third", "fourth", etc. (if any) in the specification, claims and above-mentioned drawings of the present application are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "include" or "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0024] For ease of understanding, the specific process of the embodiments of the present application is described below. Please refer to Figure 1 An embodiment of the method for optimizing an industrial manufacturing process and production operation and maintenance based on digital twin in the embodiments of the present application includes:

[0025] Step S101: Perform multi-physical field coupling modeling on pre-collected industrial production process data to obtain an industrial digital twin model;

[0026] Step S102: Perform dynamic sensitivity analysis on the industrial digital twin model to obtain a time-series parameter sensitivity map;

[0027] Step S103: Perform hybrid optimization on the time-series parameter sensitivity map to obtain a multi-objective process strategy;

[0028] Step S104: Perform Monte Carlo simulation on the multi-objective process strategy to obtain a strategy risk assessment matrix;

[0029] Step S105: Perform adaptive control analysis based on the strategy risk assessment matrix to obtain real-time optimization effect data;

[0030] Step S106: Input the real-time optimization effect data into a preset graph neural network for multi-source data association analysis to obtain an industrial manufacturing knowledge graph, and generate a target optimization strategy according to the industrial manufacturing knowledge graph.

[0031] It can be understood that the execution entity of this application can be an industrial manufacturing process and production operation and maintenance optimization system based on digital twins, or it can also be a terminal or a server. Specifically, no limitation is made here. In the embodiments of this application, the server is taken as the execution entity for illustration.

[0032] Specifically, multi-physical field coupling modeling is performed on the pre-collected industrial production process data to obtain an industrial digital twin model, providing a basis for production operation and maintenance. This process includes data cleaning and preprocessing to obtain a standardized multi-source data set, followed by principal component analysis and dimensionality reduction processing. The deep autoencoder extracts features from the dimensionality-reduced feature data set to obtain latent feature vectors. The long short-term memory network performs temporal modeling on these features to generate temporal feature representations. The convolutional neural network further extracts spatial features to obtain spatio-temporal feature mapping data. The multi-head attention mechanism fuses these features to obtain multi-physical field coupling features. The variational inference algorithm estimates the probability distribution of the coupling features, and the Monte Carlo sampling algorithm generates multiple groups of parameter samples. The finite element analysis algorithm performs numerical simulations on the parameter samples, and finally optimizes through error backpropagation to obtain the industrial digital twin model. The UniTwin digital twin industrial software plays a key role in this process, achieving high-precision digital twin modeling and providing a reliable basis for production operation and maintenance decisions.

[0033] Performing dynamic sensitivity analysis on the industrial digital twin model to obtain a temporal parameter sensitivity map, which is helpful for predictive maintenance in production operation and maintenance. This step first performs parameter decomposition to obtain a set of key parameters. The orthogonal experiment algorithm is used for parameter combination to generate parameter combination schemes. Multiple simulations are performed according to the schemes to obtain an initial simulation result set. Analysis of variance of the results set is performed to obtain the main effects and interaction effects of the parameters. The time series decomposition algorithm processes these effects to obtain time-varying sensitivity indicators. Wavelet transform processes these indicators to obtain multi-scale sensitivity features. The adaptive boosting algorithm is used to obtain the parameter importance ranking, and the analytic hierarchy process obtains the parameter weight matrix. The non-linear regression algorithm generates a parameter response surface according to the weight matrix, and after temporal slicing and interpolation processing, a temporal parameter sensitivity map is formed.

[0034] Performing hybrid optimization on the temporal parameter sensitivity map to obtain a multi-objective process strategy and optimize the production operation and maintenance process. This process includes time window segmentation, principal component analysis, genetic algorithm initial population generation, non-dominated sorting, particle swarm optimization local search, fuzzy C-means clustering analysis, deep reinforcement learning network strategy evaluation, Bayesian optimization algorithm parameter fine-tuning, and multi-objective decision tree algorithm strategy correction. The UniTwin digital twin industrial software provides powerful algorithm support and data processing capabilities in this optimization process, effectively improving the production operation and maintenance efficiency.

[0035] Perform Monte Carlo simulation on the multi-objective process strategy to obtain the strategy risk assessment matrix, providing support for production operation and maintenance risk management. This step includes parameter extraction, Latin hypercube sampling, parallel simulation calculation, probability distribution analysis, kernel density estimation, index confidence interval analysis, multi-dimensional scoring, principal component analysis for dimensionality reduction, and clustering analysis, and finally constructs the strategy risk assessment matrix. Perform adaptive control analysis based on the strategy risk assessment matrix to obtain real-time optimization effect data and achieve intelligent control of production operation and maintenance. This process involves eigenvalue decomposition, fuzzy analytic hierarchy process, construction of risk-weighted decision tree, robustness analysis, adaptive learning rate matching, closed-loop simulation, fast Fourier transform, online optimization, etc. UniTwin digital twin industrial software achieves efficient adaptive control and optimization in this stage, greatly improving the efficiency and quality of production operation and maintenance.

[0036] Finally, input the real-time optimization effect data into the pre-set graph neural network for multi-source data correlation analysis to obtain the industrial manufacturing knowledge graph, and generate the target optimization strategy based on the knowledge graph, providing comprehensive knowledge support for production operation and maintenance. This step includes time window segmentation, feature extraction and dimensionality reduction, spatial correlation analysis of graph convolutional neural network, hierarchical clustering, importance calculation, path mining, expert rule fusion, graph reasoning operation, multi-objective optimization algorithm evaluation and ranking, and finally obtains the target optimization strategy.

[0037] For example: A manufacturing enterprise uses UniTwin digital twin industrial software to optimize the operation and maintenance management of its production line. Through multi-physical field coupling modeling, the software constructs an accurate digital twin model. Dynamic sensitivity analysis finds that temperature and pressure are the key parameters affecting product quality and equipment maintenance. Hybrid optimization generates a series of multi-objective process strategies, such as adjusting the temperature range to 185 - 195 °C and the pressure range to 5.5 - 6.5 MPa, and at the same time formulates corresponding predictive maintenance plans. Monte Carlo simulation evaluates the risks of these strategies and concludes that under the given conditions, the product qualification rate can be increased by about 8%, and the equipment failure rate can be reduced by about 15%. Adaptive control analysis further optimizes the parameters, and real-time data shows that the energy consumption is reduced by about 5%, and the equipment service life is extended by about 10%. Finally, through knowledge graph analysis, the software generates a set of comprehensive optimization strategies, which while ensuring product quality, increase the production efficiency by about 12%, reduce the energy consumption by about 7%, and reduce the equipment maintenance cost by about 20%. These data are obtained by comparing the production records, energy consumption monitoring, quality inspection results, and equipment maintenance records before and after optimization, fully demonstrating the powerful functions of UniTwin digital twin industrial software in industrial manufacturing process optimization and production operation and maintenance management.

[0038] In the embodiments of the present application, a high-precision digital twin model is constructed through multi-physical field coupling modeling, the temporal changes of parameter impacts are captured by dynamic sensitivity analysis, and multi-objective balance is achieved by combining a hybrid optimization strategy. Monte Carlo simulation and risk assessment in the method improve decision-making reliability, and adaptive control technology ensures real-time optimization effects. A graph neural network is innovatively introduced to construct an industrial manufacturing knowledge graph, realizing knowledge accumulation and intelligent decision-making. This method integrates a comprehensive data processing process from data cleaning, dimensionality reduction, feature extraction to temporal modeling, and combines various advanced algorithms such as deep learning, reinforcement learning, and genetic algorithms to achieve multi-scale analysis and interpretable optimization. Through adaptive control and online tuning, the method has real-time performance and high adaptability, forming a highly integrated and automated closed-loop optimization process. This comprehensive, accurate, dynamic, and intelligent optimization method can significantly improve production efficiency, product quality, and resource utilization rate, while effectively reducing operation risks and costs.

[0039] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0040] (1) Clean and preprocess the industrially produced process data collected in advance to obtain a standardized multi-source data set;

[0041] (2) Perform principal component analysis and dimensionality reduction processing on the standardized multi-source data set to obtain a dimensionality-reduced feature data set;

[0042] (3) Extract features from the dimensionality-reduced feature data set through a pre-set deep autoencoder to obtain latent feature vectors;

[0043] (4) Input the latent feature vectors into a pre-trained long short-term memory network for temporal modeling to obtain temporal feature representations;

[0044] (5) Extract spatial features from the temporal feature representations through a convolutional neural network to obtain spatio-temporal feature mapping data;

[0045] (6) Input the spatio-temporal feature mapping data into a multi-head attention mechanism for feature fusion to obtain multi-physical field coupling features;

[0046] (7) Estimate the probability distribution of the multi-physical field coupling features through a variational inference algorithm to obtain parameter distribution data;

[0047] (8) Input the parameter distribution data into a Monte Carlo sampling algorithm to generate multiple groups of parameter samples to obtain a parameter sampling set;

[0048] (9) Perform numerical simulation on the parameter sampling set through a finite element analysis algorithm to obtain multi-physical field simulation results;

[0049] (10) Optimize the multi - physical - field simulation results and industrial production process data through error backpropagation to obtain an industrial digital twin model.

[0050] Specifically, clean and preprocess the pre - collected industrial production process data to obtain a standardized multi - source data set. Data cleaning includes removing outliers, handling missing data, and eliminating noise, while preprocessing involves data normalization and standardization to ensure that data from different sources and scales can be uniformly processed. Subsequently, perform principal component analysis and dimensionality reduction on the standardized multi - source data set to obtain a dimensionality - reduced feature data set. Principal component analysis projects high - dimensional data into a low - dimensional space through linear transformation, retaining the main information of the data, thereby reducing computational complexity and eliminating redundancy. Extract features from the dimensionality - reduced feature data set through a pre - configured deep auto - encoder to obtain latent feature vectors. A deep auto - encoder is an unsupervised learning algorithm that learns the internal representation of data by compressing the input data into a low - dimensional latent space and then reconstructing the original input. This step can capture the non - linear features and latent structures of the data. The UniTwin digital twin industrial software utilizes advanced deep - learning techniques at this stage to achieve efficient feature extraction.

[0051] Input the latent feature vectors into a pre - trained long short - term memory network for time - series modeling to obtain a time - series feature representation. A long short - term memory network is a special type of recurrent neural network that can learn long - term dependencies and is particularly suitable for processing time - series data. Through this step, the model can capture the time - dynamic characteristics in the industrial production process. Extract spatial features from the time - series feature representation through a convolutional neural network to obtain spatio - temporal feature mapping data. A convolutional neural network can effectively extract spatial features through local receptive fields and weight - sharing mechanisms, and reduce the number of parameters while maintaining the spatial structure. This step enables the model to consider information in both time and space dimensions simultaneously.

[0052] Input the spatio-temporal feature mapping data into the multi-head attention mechanism for feature fusion to obtain the multi-physical field coupling features. The multi-head attention mechanism allows the model to simultaneously focus on information in different subspaces, thus capturing the complex relationships between features more comprehensively. This step plays a crucial role in the UniTwin digital twin industrial software, achieving the effective fusion of multi-physical field information. Estimate the probability distribution of the multi-physical field coupling features through the variational inference algorithm to obtain the parameter distribution data. Variational inference is an approximate Bayesian inference method that can estimate the posterior probability distribution in complex models. This step enables the model to quantify the uncertainty of the parameters, providing a basis for subsequent risk assessment. Input the parameter distribution data into the Monte Carlo sampling algorithm to generate multiple groups of parameter samples, obtaining the parameter sampling set. Monte Carlo sampling is a numerical calculation method based on random sampling that can generate samples from complex probability distributions. This step provides diverse inputs for subsequent numerical simulations, enhancing the robustness of the model.

[0053] Perform numerical simulations on the parameter sampling set through the finite element analysis algorithm to obtain the multi-physical field simulation results. Finite element analysis is a numerical method that solves partial differential equations by decomposing a complex system into simple elements. This step can simulate the physical field behavior under different parameter conditions, providing accurate simulation results for the digital twin model. Finally, optimize the multi-physical field simulation results and the industrial production process data through error backpropagation to obtain the industrial digital twin model. Error backpropagation is a gradient descent algorithm that optimizes model parameters by minimizing the error between the predicted value and the actual value. This step ensures that the digital twin model can accurately reflect the actual production process.

[0054] For example, a manufacturing enterprise uses UniTwin digital twin industrial software to optimize its production line. First, it collects multi-source production data including temperature, pressure, flow rate, etc. After data cleaning and preprocessing, about 5% of the outliers are removed, and all data is standardized to the range of [-1, 1]. Principal component analysis reduces the original 50 features to 15 principal components, retaining 95% of the data variance. The deep autoencoder further extracts an 8-dimensional latent feature vector. The long short-term memory network models the time series of these features, capturing the long-term dependencies in the production process, such as the impact of equipment aging on product quality. The convolutional neural network extracts spatial features from the time series data, identifying the interactions at different positions on the production line. The multi-head attention mechanism fuses multi-physical field information such as the temperature field and pressure field. Variational inference estimates the probability distribution of key parameters, such as the temperature control accuracy of ±2°C (95% confidence interval). Monte Carlo sampling generates 1000 groups of parameter samples, and finite element analysis simulates the production process under these samples. Finally, through error backpropagation, the prediction error of the digital twin model is reduced to less than 3%, successfully constructing a high-precision industrial digital twin model. This example demonstrates how UniTwin digital twin industrial software achieves a high-precision digital twin of the industrial production process through complex data processing and modeling processes.

[0055] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0056] (1) Decompose the parameters of the industrial digital twin model to obtain a set of key parameters;

[0057] (2) Use the orthogonal experiment algorithm to combine the key parameters to obtain a parameter combination scheme;

[0058] (3) Perform multiple simulations on the industrial digital twin model according to the parameter combination scheme to obtain an initial simulation result set;

[0059] (4) Perform simulation variance analysis on the initial simulation result set to obtain the main effects and interaction effects of the parameters;

[0060] (5) Input the main effects and interaction effects of the parameters into the time series decomposition algorithm to obtain time-varying sensitivity indicators;

[0061] (6) Perform wavelet transform processing on the time-varying sensitivity indicators to obtain multi-scale sensitivity features;

[0062] (7) Input the multi-scale sensitivity features into the adaptive boosting algorithm to obtain the parameter importance ranking;

[0063] (8) Perform hierarchical analysis on the parameter importance ranking to obtain the parameter weight matrix;

[0064] (9) Input the parameter weight matrix into the non - linear regression algorithm to obtain the parameter response surface;

[0065] (10) Perform time - series slicing and interpolation processing on the parameter response surface to obtain the time - series parameter sensitivity map.

[0066] Specifically, decompose the parameters of the industrial digital twin model to obtain the key parameter set. Parameter decomposition is the process of identifying the key factors affecting system performance. By analyzing the model structure and physical meaning, a complex system is simplified into several key parameters. Subsequently, use the orthogonal experiment algorithm to perform parameter combination on the key parameter set to obtain the parameter combination scheme. The orthogonal experiment algorithm is an efficient experimental design method that can investigate the influence of multiple factors with fewer experimental times, greatly reducing the experimental workload. Perform multiple simulations on the industrial digital twin model according to the parameter combination scheme to obtain the initial simulation result set. In this process, the UniTwin digital twin industrial software plays a key role. Through its efficient simulation calculation ability, a large amount of simulation data is quickly generated. Perform simulation variance analysis on the initial simulation result set to obtain the main effects and interaction effects of the parameters. Variance analysis is a statistical method used to evaluate the contribution degree of different factors to the result variation. The main effect reflects the influence of a single parameter, while the interaction effect reflects the interaction between parameters.

[0067] Input the main effects and interaction effects of the parameters into the time - series decomposition algorithm to obtain the time - varying sensitivity index. The time - series decomposition algorithm decomposes time - series data into trends, seasonality, and random components, which helps to identify the variation law of parameter sensitivity over time. Perform wavelet transform processing on the time - varying sensitivity index to obtain the multi - scale sensitivity features. Wavelet transform is a time - frequency analysis method that can analyze signal features at different time scales, which helps to capture the variation of parameter sensitivity at different time scales. Input the multi - scale sensitivity features into the adaptive boosting algorithm to obtain the parameter importance ranking. The adaptive boosting algorithm is an ensemble learning method that, through iterative training of multiple weak classifiers and assigning different weights, finally obtains a strong classifier. Here, it is used to evaluate and rank the importance of parameters. Perform hierarchical analysis on the parameter importance ranking to obtain the parameter weight matrix. The analytic hierarchy process is a multi - criterion decision - making method that, by constructing a hierarchical structure and making pairwise comparisons, finally obtains the relative importance weights of each parameter.

[0068] Input the parameter weight matrix into the non - linear regression algorithm to obtain the parameter response surface. The non - linear regression algorithm describes the relationship between input parameters and output responses by fitting non - linear functions and can capture complex non - linear effects. Finally, perform time - series slicing and interpolation on the parameter response surface to obtain the time - series parameter sensitivity map. Time - series slicing slices the response surface at different time points, and interpolation is used to fill the data between discrete time points, ultimately forming a continuous time - series parameter sensitivity map.

[0069] For example: A manufacturing enterprise uses UniTwin digital twin industrial software to optimize its production line. First, 10 key parameters such as temperature, pressure, and flow rate are identified through parameter decomposition. The orthogonal experiment algorithm designs 32 sets of parameter combination schemes, greatly reducing the required number of simulations. The UniTwin software efficiently executes these 32 sets of simulations to generate an initial simulation result set. Analysis of variance shows that the main effect of temperature is the most significant, contributing 40% of the result variation, while the interaction effect of temperature and pressure contributes 15% of the variation. Time - series decomposition finds that the temperature sensitivity shows an obvious periodic change with a period of about 8 hours, which is highly correlated with the production shift. Wavelet transform further reveals the characteristics of temperature sensitivity at different time scales, such as short - term fluctuations (10 - minute scale) and long - term trends (24 - hour scale). The adaptive boosting algorithm comprehensively considers these characteristics and ranks temperature first in terms of parameter importance, followed by pressure and flow rate.

[0070] The analytic hierarchy process quantifies the parameter weights. The temperature weight is 0.4, the pressure is 0.3, the flow rate is 0.2, and the other parameters together account for 0.1. Non - linear regression establishes the parameter response surface and finds that the product quality has a quadratic function relationship with temperature, and the optimal temperature range is 185 - 195 °C. Time - series slicing and interpolation processing generate a 24 - hour time - series parameter sensitivity map, clearly showing the variation of each parameter sensitivity over time. This time - series parameter sensitivity map provides valuable decision - making basis for the enterprise. For example, during periods with high temperature sensitivity (such as 2 - 4 am), strengthen temperature control; during periods with high pressure sensitivity (such as 9 - 11 am), optimize the pressure parameter. Through this refined parameter regulation, the enterprise has successfully increased the product qualification rate by 5% while reducing energy consumption by 3%.

[0071] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0072] (1) Divide the time - series parameter sensitivity map into time windows to obtain sensitivity sub - maps for multiple time periods;

[0073] (2) Perform principal component analysis on the sensitivity sub - maps for multiple time periods to obtain the dimensionality - reduced eigenvectors;

[0074] (3) Input the dimensionality-reduced feature vectors into a genetic algorithm to generate an initial population, and obtain a set of candidate solutions;

[0075] (4) Perform non-dominated sorting on the set of candidate solutions to obtain Pareto front solutions;

[0076] (5) Input the Pareto front solutions into a particle swarm optimization algorithm for local search to obtain an optimized solution set;

[0077] (6) Perform fuzzy C-means clustering analysis on the optimized solution set to obtain process strategy clusters;

[0078] (7) Input the process strategy clusters into a deep reinforcement learning network for policy evaluation to obtain policy value estimation data;

[0079] (8) Perform policy selection on the policy value estimation data to obtain an optimal policy subset;

[0080] (9) Input the optimal policy subset into a Bayesian optimization algorithm for parameter fine-tuning to obtain a target policy solution;

[0081] (10) Perform policy correction on the target policy solution through a multi-objective decision tree algorithm to obtain a multi-objective process strategy.

[0082] Specifically, perform time window segmentation on the time series parameter sensitivity map to obtain sensitivity sub-maps for multiple time periods. Time window segmentation divides continuous time series data into several discrete time periods, and the data features within each time period are relatively stable. This segmentation method helps to capture the changing characteristics of parameter sensitivity in different time periods. Then, perform principal component analysis on the sensitivity sub-maps for multiple time periods to obtain dimensionality-reduced feature vectors. Principal component analysis is a commonly used dimensionality reduction technique that projects high-dimensional data into a low-dimensional space through linear transformation, retains the main information of the data, thereby reducing the computational complexity and eliminating redundancy. Input the dimensionality-reduced feature vectors into a genetic algorithm to generate an initial population, and obtain a set of candidate solutions. The genetic algorithm is an optimization method that simulates the biological evolution process and continuously iterates through selection, crossover, and mutation operations to search for the optimal solution. Here, the genetic algorithm is used to generate a diverse set of initial candidate solutions. Perform non-dominated sorting on the set of candidate solutions to obtain Pareto front solutions. Non-dominated sorting is a method in multi-objective optimization used to identify the solution set that is not dominated by other solutions under multiple objectives, i.e., the Pareto front solutions.

[0083] The Pareto front solutions are input into the particle swarm optimization algorithm for local search to obtain an optimized solution set. The particle swarm optimization algorithm is a swarm intelligence algorithm that searches for the optimal solution by simulating the foraging behavior of bird flocks. Here, it is used to perform fine-grained local search near the Pareto front solutions to further improve the quality of the solutions. Fuzzy C-means clustering analysis is performed on the optimized solution set to obtain process strategy clusters. Fuzzy C-means clustering is a soft clustering method that allows a data point to belong to multiple clusters to varying degrees, thus being able to better handle data with fuzzy boundaries. The process strategy clusters are input into a deep reinforcement learning network for policy evaluation to obtain policy value estimation data. Deep reinforcement learning combines the advantages of deep learning and reinforcement learning and can learn the optimal policy in a complex environment. Here, it is used to evaluate the long-term value of different process strategies. Policy selection is performed on the policy value estimation data to obtain an optimal policy subset. Policy selection is the process of screening out the most potential policy subset based on the evaluation results.

[0084] The optimal policy subset is input into the Bayesian optimization algorithm for parameter fine-tuning to obtain the target policy scheme. Bayesian optimization is a global optimization algorithm that is particularly suitable for optimizing computationally expensive black-box functions. It guides the search process by constructing surrogate models and acquisition functions and can find solutions close to the global optimum with fewer evaluation times. Finally, the multi-objective decision tree algorithm is used to correct the target policy scheme to obtain the multi-objective process strategy. The multi-objective decision tree algorithm is an extension of the decision tree in multi-objective optimization problems, which can consider multiple objectives simultaneously and provide intuitive decision rules. In this process, the UniTwin digital twin industrial software plays a key role, providing powerful data processing and algorithm implementation capabilities. For example: A manufacturing enterprise uses the UniTwin digital twin industrial software to optimize its production line. First, the time window segmentation is performed on the 24-hour time-series parameter sensitivity map to obtain 8 sensitivity sub-maps of 3 hours each. Principal component analysis reduces the 50 features of each sub-map to 10 principal components, retaining 95% of the information. The genetic algorithm generates 1000 initial candidate solutions, and non-dominated sorting selects 100 Pareto front solutions from them.

[0085] The particle swarm optimization algorithm performs local search around these 100 solutions to obtain 50 optimized solutions. Fuzzy C-means clustering divides these 50 solutions into 5 process strategy clusters. The deep reinforcement learning network evaluates the long-term value of these 5 strategy clusters by simulating 10,000 production cycles. According to the evaluation results, the 2 strategy clusters with the highest value are selected as the optimal policy subset. The Bayesian optimization algorithm performs 100 iterations of parameter fine-tuning on these 2 strategy clusters to obtain the final target policy scheme. The multi-objective decision tree algorithm considers three objectives: product quality, production efficiency, and energy consumption, and corrects the policy scheme to finally obtain the multi-objective process strategy that balances these three objectives.

[0086] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0087] (1) Extract parameters from the multi-objective process strategy to obtain a set of strategy parameters;

[0088] (2) Perform Latin hypercube sampling on the set of strategy parameters to obtain a parameter sample space;

[0089] (3) Input the parameter sample space into the industrial digital twin model for parallel simulation calculation to obtain a set of parallel simulation results;

[0090] (4) Perform probability distribution analysis on the set of parallel simulation results to obtain performance distribution characteristics;

[0091] (5) Smooth the performance distribution characteristics by the kernel density estimation method to obtain continuous probability density data;

[0092] (6) Perform index confidence interval analysis on the continuous probability density data to obtain risk boundary values;

[0093] (7) Perform multi-dimensional scoring on the risk boundary values to obtain a quantitative index of strategy risk;

[0094] (8) Perform dimensionality reduction processing on the quantitative index of strategy risk by the principal component analysis method to obtain dimensionality reduction risk characteristics;

[0095] (9) Perform clustering analysis on the dimensionality reduction risk characteristics to obtain a result of risk level classification;

[0096] (10) Associate and fuse the result of risk level classification with the multi-objective process strategy to construct a strategy risk assessment matrix.

[0097] Specifically, parameter extraction is performed on the multi-objective process strategy to obtain a set of strategy parameters. Parameter extraction is the process of identifying and extracting key control parameters from complex process strategies, and these parameters directly affect the performance and results of the production process. Subsequently, Latin hypercube sampling is performed on the set of strategy parameters to obtain a parameter sample space. Latin hypercube sampling is an efficient statistical sampling method that can better cover the multi-dimensional parameter space with a smaller number of samples, ensuring the representativeness and uniformity of sampling. The parameter sample space is input into the industrial digital twin model for parallel simulation calculation to obtain a set of parallel simulation results. In this step, the UniTwin digital twin industrial software plays a key role. By using its powerful parallel computing ability, the simulation efficiency is greatly improved. Parallel simulation allows multiple sets of parameter combinations to be simulated simultaneously, quickly generating a large number of simulation results. Probability distribution analysis is performed on the set of parallel simulation results to obtain the performance distribution characteristics. Probability distribution analysis reveals the distribution laws of different performance indicators under various parameter combinations, providing a basis for subsequent risk assessment. The performance distribution characteristics are smoothed through the kernel density estimation method to obtain continuous probability density data. Kernel density estimation is a non-parametric statistical method used to estimate the probability density function of a random variable, which can effectively process discrete data points and generate a smooth continuous distribution. Index confidence interval analysis is performed on the continuous probability density data to obtain the risk threshold value. Confidence interval analysis determines the credible range of performance indicators, and values outside this range are regarded as potential risks.

[0098] Multi-dimensional scoring is performed on the risk threshold value to obtain a quantitative index of strategy risk. Multi-dimensional scoring takes into account the weights and interrelationships of different risk factors, transforming qualitative risk assessment into a quantitative risk index. The dimensionality reduction processing is performed on the quantitative index of strategy risk through the principal component analysis method to obtain the dimensionality-reduced risk characteristics. Principal component analysis reduces the data dimensionality while retaining the main information, facilitating subsequent analysis and visualization. Cluster analysis is performed on the dimensionality-reduced risk characteristics to obtain the results of risk level classification. Cluster analysis groups similar risk characteristics to form different risk levels, facilitating risk management and decision-making. Finally, the results of risk level classification are associated and integrated with the multi-objective process strategy to construct a strategy risk assessment matrix. This matrix intuitively shows the performance of different process strategies and the corresponding risk levels, providing a comprehensive reference for decision-making.

[0099] For example: A manufacturing enterprise uses the UniTwin digital twin industrial software to optimize its production line. First, 10 key parameters are extracted from the multi-objective process strategy, including temperature, pressure, flow rate, etc. Through Latin hypercube sampling, 1000 sets of parameter combinations are generated. The UniTwin software uses a high-performance computing cluster to complete the parallel simulation of these 1000 sets of parameters within 2 hours, obtaining a set of simulation results including indicators such as product quality, production efficiency, and energy consumption.

[0100] Probability distribution analysis shows that the product quality follows a normal distribution, with an average qualification rate of 98.5% and a standard deviation of 1.2%. Kernel density estimation further smooths these distributions, generating a continuous probability density function. Confidence interval analysis determines the risk boundary values at a 95% confidence level. For example, the lower limit of the product qualification rate is 96.1%. Multidimensional scoring comprehensively considers indicators such as quality, efficiency, and energy consumption, obtaining a risk quantitative indicator ranging from 0 to 100 points. Principal component analysis reduces the 10-dimensional risk indicators to 3 principal components, cumulatively explaining 92% of the variance.

[0101] The K-means clustering algorithm classifies risk characteristics into three levels: low, medium, and high. The final strategic risk assessment matrix shows that among 1000 strategies, 58% belong to low risk, 35% belong to medium risk, and 7% belong to high risk. Among them, a typical low-risk strategy improves the production efficiency by 7% and reduces the energy consumption by 5% while maintaining a product qualification rate of 98.7%. These data are obtained by comparing the actual production data of one week before and after optimization.

[0102] The powerful functions of UniTwin digital twin industrial software enable enterprises to quickly conduct large-scale simulations and risk assessments, providing comprehensive and intuitive risk information for decision-makers. By implementing low-risk optimization strategies, enterprises have successfully improved production efficiency and product quality while effectively controlling production risks.

[0103] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0104] (1) Perform eigenvalue decomposition on the strategic risk assessment matrix to obtain the risk principal components;

[0105] (2) Calculate the weights of the prime risk principal components through the fuzzy analytic hierarchy process to obtain the risk weight vector;

[0106] (3) Construct a risk-weighted decision tree based on the risk weight vector and multi-objective process strategies to obtain the initial control strategy;

[0107] (4) Conduct robustness analysis on the initial control strategy to obtain the strategy robustness score;

[0108] (5) Match the adaptive learning rate through the strategy robustness score and construct a dynamic adjustment mechanism based on the adaptive learning rate to obtain the adaptive controller;

[0109] (6) Integrate the adaptive controller with the industrial digital twin model and conduct closed-loop simulation to obtain the control response curve;

[0110] (7) Perform a fast Fourier transform on the control response curve to obtain the system dynamic characteristic indicators;

[0111] (8) Online optimize the parameters of the adaptive controller based on the dynamic characteristic indexes to obtain the optimized control parameters;

[0112] (9) Carry out industrial manufacturing simulation operation on the industrial digital twin model according to the optimized control parameters, and collect the simulated real-time process data;

[0113] (10) Visualize the simulated real-time process data to obtain the real-time optimization effect data.

[0114] Specifically, perform eigen-decomposition on the policy risk assessment matrix to obtain the risk principal components, providing a data basis for production operation and maintenance decision-making. Eigen-decomposition is a mathematical operation that decomposes a matrix into eigenvalues and eigenvectors, which helps to identify the main structures and change patterns in the data and is crucial for risk identification in production operation and maintenance. Calculate the weights of these risk principal components through the fuzzy analytic hierarchy process to obtain the risk weight vector, providing guidance for formulating production operation and maintenance strategies. The fuzzy analytic hierarchy process is a multi-criteria decision-making method that combines fuzzy theory and the analytic hierarchy process, capable of handling the uncertainty and ambiguity in the decision-making process and is particularly suitable for complex production operation and maintenance environments. Construct a risk-weighted decision tree based on the risk weight vector and the multi-objective process strategy to obtain the initial control strategy, providing decision support for production operation and maintenance. A decision tree is an intuitive decision support tool. By introducing risk weights, it can better balance the impacts of different risk factors on production operation and maintenance decisions. Conduct a robustness analysis on the initial control strategy to obtain the strategy robustness score and evaluate the reliability of the production operation and maintenance strategy. Robustness analysis evaluates the stability and reliability of the strategy under different conditions and is an important indicator for measuring the quality of production operation and maintenance strategies.

[0115] Adaptive learning rate is matched through policy robustness scoring, and a dynamic adjustment mechanism is constructed based on the adaptive learning rate to obtain an adaptive controller, realizing intelligent regulation of production operation and maintenance. The adaptive learning rate can dynamically adjust the update speed of control parameters according to the robustness of the policy, improving the adaptability and stability of the controller, which is crucial for the continuous optimization of production operation and maintenance. Integrate the adaptive controller with the industrial digital twin model and conduct closed-loop simulation to obtain the control response curve, simulating the production operation and maintenance process. In this process, the UniTwin digital twin industrial software plays a key role, providing a high-precision digital twin model and efficient simulation capabilities, providing strong support for the optimization of production operation and maintenance. Perform fast Fourier transform on the control response curve to obtain system dynamic characteristic indicators, and deeply analyze the performance of the production operation and maintenance system. Fast Fourier transform is an efficient signal processing algorithm that can convert time-domain signals into frequency-domain representations, revealing the frequency characteristics and dynamic behavior of the system, which helps to optimize production operation and maintenance strategies. Online optimize the parameters of the adaptive controller through the dynamic characteristic indicators to obtain optimized control parameters, realizing real-time optimization of production operation and maintenance. The online optimization process continuously optimizes the controller performance using real-time feedback information, improving the response speed and stability of the system, ensuring the efficiency of production operation and maintenance.

[0116] Conduct industrial manufacturing simulation operation on the industrial digital twin model according to the optimized control parameters, collect simulated real-time process data, and provide data support for production operation and maintenance decision-making. The UniTwin digital twin industrial software provides a highly simulated environment in this stage, enabling the simulation operation to accurately reflect the actual production process, providing a reliable platform for the verification and optimization of production operation and maintenance strategies. Visualize the simulated real-time process data to obtain real-time optimization effect data, intuitively showing the production operation and maintenance optimization results. Visualization processing converts complex data into intuitive charts and indicators, facilitating decision-makers to quickly understand and evaluate the production operation and maintenance optimization effect.

[0117] For example: A manufacturing enterprise uses the UniTwin digital twin industrial software to optimize the operation and maintenance management of its production line. First, perform eigenvalue decomposition on the 10×10 policy risk assessment matrix to identify 3 main risk components, representing quality risk, efficiency risk, and energy consumption risk respectively, laying a foundation for formulating a comprehensive production operation and maintenance strategy. The risk weight vector calculated by the fuzzy analytic hierarchy process is [0.5, 0.3, 0.2], reflecting the enterprise's high emphasis on quality, while also taking into account the production operation and maintenance goals of efficiency and energy consumption.

[0118] The decision tree constructed based on the risk weight vector generates an initial control strategy, including specific parameters such as temperature control at 190±5°C and pressure control at 6±0.5 MPa, and formulates a corresponding predictive maintenance plan. Robustness analysis shows that the strategy can still maintain a performance of more than 90% under the conditions of temperature fluctuation of ±10°C and pressure fluctuation of ±1 MPa, and obtains a robustness score of 85 (out of 100), proving the reliability of the production operation and maintenance strategy.

[0119] According to the robustness score, an initial adaptive learning rate of 0.01 is matched, providing flexibility for the continuous optimization of production operation and maintenance. The UniTwin software integrates the adaptive controller with the digital twin model, conducts 1000 closed-loop simulations, obtains the control response curve, and comprehensively simulates the production operation and maintenance process. Fast Fourier transform analysis shows that the main oscillation frequency of the system is 0.1 Hz. Based on this, the controller parameters are optimized, and the response time is shortened from the original 10 seconds to 5 seconds, greatly improving the efficiency of production operation and maintenance. Using the optimized control parameters, the UniTwin software simulates the production process for 24 hours, collects data every 10 seconds, and generates a total of 8640 data points, providing rich data support for production operation and maintenance decisions. These data not only reflect the optimization effect of the production process but also provide an important basis for predictive maintenance, helping to formulate a more accurate equipment maintenance plan, reduce the risk of unexpected downtime, and improve the overall production operation and maintenance efficiency.

[0120] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0121] (1) Perform time window segmentation on the real-time optimization effect data to obtain data subsets for multiple time periods;

[0122] (2) Perform feature extraction and dimensionality reduction on the data subsets for multiple time periods to obtain the effect feature vector;

[0123] (3) Input the effect feature vector into the graph convolutional neural network for spatial correlation analysis to obtain the node embedding representation data;

[0124] (4) Perform hierarchical clustering on the node embedding representation data and construct a multi-level knowledge structure to obtain the initial knowledge graph;

[0125] (5) Calculate the importance of the initial knowledge graph to obtain the weighted knowledge graph;

[0126] (6) Mine the paths of the weighted knowledge graph to obtain the causal relationship network;

[0127] (7) Integrate the causal relationship network with the preset expert rules to obtain the enhanced knowledge graph;

[0128] (8) Perform graph reasoning operations on the enhanced knowledge graph to obtain a candidate set of optimization strategies;

[0129] (9) Evaluate and rank the candidate set of optimization strategies through a multi-objective optimization algorithm to obtain a list of strategy priorities;

[0130] (10) Conduct parameter sensitivity analysis and strategy optimization based on the list of strategy priorities to obtain the target optimization strategy.

[0131] Specifically, perform time window segmentation on the real-time optimization effect data to obtain data subsets for multiple time periods. Time window segmentation divides continuous time series data into fixed-length or overlapping time periods, which helps capture data characteristics at different time scales. Subsequently, perform feature extraction and dimensionality reduction on the data subsets for multiple time periods to obtain effect feature vectors. Feature extraction involves extracting meaningful features from the original data, while dimensionality reduction reduces the data dimension through methods such as principal component analysis (PCA) to retain key information. Input the effect feature vectors into a graph convolutional neural network for spatial correlation analysis to obtain node embedding representation data. A graph convolutional neural network is a deep learning model specialized for processing graph-structured data, which can effectively capture the spatial relationships and topological structures between data. In this process, the UniTwin digital twin industrial software provides powerful data processing and model training capabilities, greatly improving the analysis efficiency. Perform hierarchical clustering on the node embedding representation data and construct a multi-level knowledge structure to obtain an initial knowledge graph. Hierarchical clustering organizes similar nodes into a tree-like structure to form a multi-level knowledge representation.

[0132] Calculate the importance of the initial knowledge graph to obtain a weighted knowledge graph. Importance calculation takes into account factors such as node connectivity and centrality to highlight the role of key knowledge points. Perform path mining on the weighted knowledge graph to obtain a causal relationship network. Path mining is the process of identifying important paths in a graph, which helps reveal the causal relationships and dependency structures between knowledge. Integrate the causal relationship network with preset expert rules to obtain an enhanced knowledge graph. This step combines data-driven knowledge discovery and the experience of domain experts to improve the accuracy and interpretability of the knowledge graph. Perform graph reasoning operations on the enhanced knowledge graph to obtain a candidate set of optimization strategies. Graph reasoning operations include rule-based reasoning and probability-based reasoning, which can derive new knowledge and strategies from known knowledge. Evaluate and rank the candidate set of optimization strategies through a multi-objective optimization algorithm to obtain a list of strategy priorities. Multi-objective optimization considers multiple potentially conflicting objectives, such as product quality, production efficiency, and energy consumption, to find the optimal solution that balances these objectives. Finally, conduct parameter sensitivity analysis and strategy optimization based on the list of strategy priorities to obtain the target optimization strategy. Parameter sensitivity analysis evaluates the impact of different parameters on the optimization objective to guide the fine-tuning of the final strategy.

[0133] For example, a manufacturing enterprise uses the UniTwin digital twin industrial software to optimize its production line. First, the real-time optimization effect data for 30 consecutive days is segmented by time window, with each 24-hour period as a time window, resulting in 30 data subsets. Feature extraction and PCA dimensionality reduction are performed on each subset, compressing the original 100 features to 10 main features to form an effect feature vector. The graph convolutional neural network module of the UniTwin software analyzes these feature vectors to generate a spatial correlation network with 500 nodes. Hierarchical clustering organizes these nodes into a 5-layer structure to construct an initial knowledge graph. Importance calculation is based on the PageRank algorithm to identify 20 key nodes in the graph to form a weighted knowledge graph. Path mining discovers 50 important causal paths, which are combined with 30 expert rules provided by the enterprise to generate an enhanced knowledge graph.

[0134] Graph reasoning operations generate 100 candidate optimization strategies based on the enhanced knowledge graph. The multi-objective optimization algorithm considers three objectives: product quality, production efficiency, and energy consumption, and evaluates and ranks these strategies. Parameter sensitivity analysis shows that temperature and pressure are the two most influential parameters. The final target optimization strategy controls the temperature within the range of 188 - 192 °C and the pressure within the range of 5.8 - 6.2 MPa. The powerful functions of the UniTwin digital twin industrial software enable the enterprise to quickly extract valuable knowledge and strategies from massive data, greatly shortening the optimization cycle and improving the scientificity and accuracy of decision-making. This example fully demonstrates the great value of the digital twin-based industrial manufacturing process and production operation and maintenance optimization method in practical applications, providing an intelligent and efficient optimization approach for manufacturing enterprises.

[0135] The above describes the digital twin-based industrial manufacturing process and production operation and maintenance optimization method in the embodiments of the present application. Next, the digital twin-based industrial manufacturing process and production operation and maintenance optimization system in the embodiments of the present application will be described. Please refer to Figure 2 One embodiment of the digital twin-based industrial manufacturing process and production operation and maintenance optimization system in the embodiments of the present application includes:

[0136] A modeling module 201 for performing multi-physical field coupling modeling on pre-collected industrial production process data to obtain an industrial digital twin model;

[0137] An analysis module 202 for performing dynamic sensitivity analysis on the industrial digital twin model to obtain a time-series parameter sensitivity map;

[0138] An optimization module 203 for performing hybrid optimization on the time-series parameter sensitivity map to obtain a multi-objective process strategy;

[0139] The simulation module 204 is configured to perform Monte Carlo simulation on the multi-objective process strategy to obtain a strategy risk assessment matrix;

[0140] The control module 205 is configured to perform adaptive control analysis based on the strategy risk assessment matrix to obtain real-time optimization effect data;

[0141] The generation module 206 is configured to input the real-time optimization effect data into a pre-set graph neural network for multi-source data association analysis to obtain an industrial manufacturing knowledge graph, and generate a target optimization strategy according to the industrial manufacturing knowledge graph.

[0142] Through the collaborative cooperation of the above-mentioned various components, a high-precision digital twin model is constructed through multi-physical field coupling modeling, the temporal changes of parameter impacts are captured by dynamic sensitivity analysis, and multi-objective balance is achieved by combining a hybrid optimization strategy. The Monte Carlo simulation and risk assessment in the method improve the decision-making reliability, and the adaptive control technology ensures the real-time optimization effect. The graph neural network is innovatively introduced to construct an industrial manufacturing knowledge graph, realizing knowledge accumulation and intelligent decision-making. The method integrates a comprehensive data processing process from data cleaning, dimensionality reduction, feature extraction to temporal modeling, and combines a variety of advanced algorithms such as deep learning, reinforcement learning, and genetic algorithms to achieve multi-scale analysis and interpretable optimization. Through adaptive control and online tuning, the method has real-time performance and high adaptability, forming a highly integrated and automated closed-loop optimization process. This comprehensive, accurate, dynamic, and intelligent optimization method can significantly improve production efficiency, product quality, and resource utilization rate, while effectively reducing operation risks and costs.

[0143] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the above-described system, system, and unit can refer to the corresponding processes in the foregoing method embodiments, and will not be described herein again.

[0144] The above is the case. The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for optimizing industrial manufacturing process and production operation and maintenance based on digital twin, characterized in that: The industrial manufacturing process and production operation and maintenance optimization method based on digital twins includes: Conduct multi-physics field coupling modeling on pre-collected industrial production process data to obtain an industrial digital twin model; Performing dynamic sensitivity analysis on the industrial digital twin model to obtain a time series parameter sensitivity map; Performing hybrid optimization on the timing parameter sensitivity map to obtain a multi-objective process strategy; Performing Monte Carlo simulation on the multi-objective process strategy to obtain a strategy risk assessment matrix; Perform adaptive control analysis based on the strategy risk assessment matrix to obtain real-time optimization effect data; Inputting the real-time optimization effect data into a preset graph neural network to perform multi-source data association analysis to obtain an industrial manufacturing knowledge graph, and generating a target optimization strategy based on the industrial manufacturing knowledge graph; The Monte Carlo simulation is performed on the multi-objective process strategy to obtain a strategy risk assessment matrix, including: Extracting parameters of the multi-objective process strategy to obtain a strategy parameter set; Perform Latin hypercube sampling on the strategy parameter set to obtain the parameter sample space; Input the parameter sample space into the industrial digital twin model for parallel simulation calculation to obtain a parallel simulation result set; Performing probability distribution analysis on the parallel simulation result set to obtain performance distribution characteristics; Smoothing the performance distribution characteristics by a kernel density estimation method to obtain continuous probability density data; Performing indicator confidence interval analysis on the continuous probability density data to obtain a risk limit value; Perform multi-dimensional scoring on the risk threshold value to obtain a quantitative indicator of strategic risk; The quantitative risk index of the strategy is subjected to dimensionality reduction processing by principal component analysis to obtain dimensionality reduction risk characteristics; Performing cluster analysis on the dimensionality reduction risk features to obtain risk level classification results; The risk level classification results are associated and integrated with the multi-objective process strategy to construct a strategy risk assessment matrix.

2. The industrial manufacturing process and production operation and maintenance optimization method based on digital twin according to claim 1 is characterized in that: The multi-physics field coupling modeling is performed on the pre-collected industrial production process data to obtain the industrial digital twin model, including: Clean and preprocess the pre-collected industrial production process data to obtain a standardized multi-source data set; Performing principal component analysis and dimensionality reduction processing on the standardized multi-source data set to obtain a dimensionality-reduced feature data set; Extracting features from the dimension-reduced feature data set by using a preset deep autoencoder to obtain a potential feature vector; Inputting the potential feature vector into a pre-trained long short-term memory network for time series modeling to obtain a time series feature representation; Performing spatial feature extraction on the temporal feature representation through a convolutional neural network to obtain spatiotemporal feature mapping data; Inputting the spatiotemporal feature mapping data into a multi-head attention mechanism for feature fusion to obtain multi-physical field coupling features; Probability distribution estimation of the multi-physical field coupling characteristics is performed by using a variational inference algorithm to obtain parameter distribution data; Inputting the parameter distribution data into a Monte Carlo sampling algorithm to generate multiple groups of parameter samples to obtain a parameter sampling set; Performing numerical simulation on the parameter sampling set by a finite element analysis algorithm to obtain a multi-physics field simulation result; The multi-physics field simulation results and the industrial production process data are subjected to error back-propagation optimization to obtain the industrial digital twin model.

3. The industrial manufacturing process and production operation and maintenance optimization method based on digital twin according to claim 1 is characterized in that: The dynamic sensitivity analysis of the industrial digital twin model is performed to obtain a time series parameter sensitivity map, including: Performing parameter decomposition on the industrial digital twin model to obtain a set of key parameters; Performing parameter combination on the key parameter set by using an orthogonal experimental algorithm to obtain a parameter combination scheme; Simulating the industrial digital twin model multiple times according to the parameter combination scheme to obtain an initial simulation result set; Performing simulation variance analysis on the initial simulation result set to obtain parameter main effects and interaction effects; Inputting the parameter main effect and the interaction effect into a time series decomposition algorithm to obtain a time-varying sensitivity index; Performing wavelet transform processing on the time-varying sensitivity index to obtain multi-scale sensitivity characteristics; Inputting the multi-scale sensitivity features into an adaptive boosting algorithm to obtain a parameter importance ranking; Performing a hierarchical analysis on the importance ranking of the parameters to obtain a parameter weight matrix; Inputting the parameter weight matrix into a nonlinear regression algorithm to obtain a parameter response surface; The parameter response surface is subjected to time series slicing and interpolation processing to obtain a time series parameter sensitivity map.

4. The industrial manufacturing process and production operation and maintenance optimization method based on digital twin according to claim 1 is characterized in that: The hybrid optimization of the timing parameter sensitivity map to obtain a multi-objective process strategy includes: Performing time window segmentation on the time series parameter sensitivity map to obtain sensitivity sub-maps of multiple time periods; Performing principal component analysis on the sensitivity sub-maps of the multiple time periods to obtain a dimension-reduced feature vector; Inputting the dimension-reduced feature vector into a genetic algorithm to generate an initial population to obtain a set of candidate solutions; Performing non-dominated sorting on the candidate solution set to obtain a Pareto front solution; Inputting the Pareto front solution into a particle swarm optimization algorithm for local search to obtain an optimized solution set; Performing fuzzy C-means clustering analysis on the optimization solution set to obtain a process strategy cluster; Inputting the process strategy cluster into a deep reinforcement learning network for strategy evaluation to obtain strategy value estimation data; Performing strategy selection on the strategy value estimation data to obtain an optimal strategy subset; Inputting the optimal strategy subset into the Bayesian optimization algorithm to fine-tune parameters and obtain the target strategy solution; The target strategy scheme is modified by a multi-objective decision tree algorithm to obtain the multi-objective process strategy.

5. The industrial manufacturing process and production operation and maintenance optimization method based on digital twin according to claim 1 is characterized in that: The adaptive control analysis is performed according to the strategy risk assessment matrix to obtain real-time optimization effect data, including: Performing eigendecomposition on the strategy risk assessment matrix to obtain risk principal components; The risk weight vector is obtained by weighting the prime risk principal components through fuzzy analytic hierarchy process. Constructing a risk-weighted decision tree according to the risk weight vector and the multi-objective process strategy to obtain an initial control strategy; Performing a robustness analysis on the initial control strategy to obtain a strategy robustness score; Matching an adaptive learning rate through the strategy robustness score, and constructing a dynamic adjustment mechanism according to the adaptive learning rate to obtain an adaptive controller; Integrating the adaptive controller with the industrial digital twin model and performing closed-loop simulation to obtain a control response curve; Performing a fast Fourier transform on the control response curve to obtain a system dynamic characteristic index; Online tuning of the adaptive controller parameters is performed through dynamic characteristic indicators to obtain optimized control parameters; Performing industrial manufacturing simulation operation on the industrial digital twin model according to the optimized control parameters, and collecting simulated real-time process data; The simulated real-time process data is visualized to obtain the real-time optimization effect data.

6. The industrial manufacturing process and production operation and maintenance optimization method based on digital twin according to claim 1 is characterized in that: The real-time optimization effect data is input into a preset graph neural network for multi-source data association analysis to obtain an industrial manufacturing knowledge graph, and a target optimization strategy is generated according to the industrial manufacturing knowledge graph, including: Performing time window segmentation on the real-time optimization effect data to obtain data subsets for multiple time periods; Performing feature extraction and dimensionality reduction on the data subsets of the multiple time periods to obtain effect feature vectors; Inputting the effect feature vector into a graph convolutional neural network for spatial correlation analysis to obtain node embedding representation data; Performing hierarchical clustering on the node embedding representation data and constructing a multi-level knowledge structure to obtain an initial knowledge graph; Calculating the importance of the initial knowledge graph to obtain a weighted knowledge graph; Performing path mining on the weighted knowledge graph to obtain a causal relationship network; The causal relationship network is integrated with preset expert rules to obtain an enhanced knowledge graph; Performing graph reasoning operations on the enhanced knowledge graph to obtain a set of candidate optimization strategies; Evaluate and sort the candidate optimization strategy set using a multi-objective optimization algorithm to obtain a strategy priority list; Parameter sensitivity analysis and strategy optimization are performed according to the strategy priority list to obtain the target optimization strategy.

7. An industrial manufacturing process and production operation and maintenance optimization system based on digital twin, used to implement the industrial manufacturing process and production operation and maintenance optimization method based on digital twin as described in any one of claims 1 to 6, characterized in that: The industrial manufacturing process and production operation and maintenance optimization system based on digital twins includes: The modeling module is used to perform multi-physics field coupling modeling on the pre-collected industrial production process data to obtain an industrial digital twin model; An analysis module, used to perform dynamic sensitivity analysis on the industrial digital twin model to obtain a time series parameter sensitivity map; An optimization module, used for performing hybrid optimization on the timing parameter sensitivity map to obtain a multi-objective process strategy; A simulation module, used for performing Monte Carlo simulation on the multi-objective process strategy to obtain a strategy risk assessment matrix; A control module, used to perform adaptive control analysis according to the strategy risk assessment matrix to obtain real-time optimization effect data; A generation module is used to input the real-time optimization effect data into a preset graph neural network for multi-source data association analysis to obtain an industrial manufacturing knowledge graph, and generate a target optimization strategy based on the industrial manufacturing knowledge graph.

Images