Simulation Method for Enterprise Digital Development Path Integrating Temporal Generative Adversarial Networks

By constructing a digital test case set and integrating a temporal adversarial generative network, the problem of insufficient representation of uncertainty and randomness in traditional methods is solved, and efficient evaluation and optimization decision-making for digital systems are achieved.

CN123070818APending Publication Date: 2026-11-10AI APP BIG DATA TECHNOLOGY (YANTAI) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611343426.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-01
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Traditional performance evaluation methods based on analytical models are difficult to accurately depict the dynamic interaction between digital systems and physical production systems. Relying solely on historical operating data faces difficulties such as scarce samples, insufficient coverage of operating conditions, and the impracticality of counterfactual reasoning, resulting in insufficient representation of uncertainty and randomness.

Method used

A digital test case set containing various typical working conditions is constructed. A temporal generative adversarial network is integrated. By combining the mechanism simulation model with the pre-trained temporal generative adversarial network, multiple simulation trajectories are generated. Statistical analysis is performed to obtain evaluation indicators and achieve a comprehensive evaluation of the digital solution.

Benefits of technology

It improves the coverage and representativeness of operating conditions in simulation verification, retains the physical consistency constraints of the mechanism model, learns the high-order temporal dependence of residuals and the heterogeneous distribution characteristics of operating conditions through adversarial training, improves the accuracy of uncertainty, and provides multi-level evaluation indicators to support the iterative search and optimization decision of digital parameters.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070818A_ABST
    Figure CN123070818A_ABST
Patent Text Reader

Abstract

This invention discloses a simulation method for enterprise digital development paths that integrates temporal adversarial generative networks, belonging to the field of digital simulation technology. The method includes: constructing a digital test case set; acquiring a digital scheme to be evaluated, initializing a pre-constructed mechanism simulation model, performing simulation analysis on the digital test case set, and obtaining a deterministic benchmark prediction sequence; using each test case and digital scheme as input conditions, generating a residual prediction sequence through a conditional temporal generator; using the test cases as indices, performing time-step superposition of the deterministic benchmark prediction sequence and each residual prediction sequence to generate multiple simulation trajectories; performing statistical analysis based on the simulation trajectories to obtain corresponding extracted evaluation indicators, and associating the evaluation indicators with the digital scheme to output the simulation results. This solves the technical problem of insufficient representation of uncertainty and randomness caused by relying solely on mechanism simulation models for deterministic prediction in existing technologies.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of digital simulation technology, specifically to a simulation method for enterprise digital development path that integrates temporal adversarial generative networks. Background Technology

[0002] In enterprise production scenarios such as production lines, the choice of digital development path directly determines the efficiency conversion rate of transformation investment. However, enterprise digital systems have complex characteristics such as multi-subsystem coupling, significant long-term delay effects, and numerous random disturbance sources. The impact of digital solutions on production operation status exhibits highly nonlinear, strong time-series dependence, and condition sensitivity.

[0003] However, traditional performance evaluation methods based on analytical models are difficult to accurately depict the dynamic interaction between digital systems and physical production systems. Relying solely on historical operating data for experience review also faces fundamental difficulties such as scarce samples, insufficient coverage of operating conditions, and the impracticality of counterfactual reasoning. Summary of the Invention

[0004] This invention provides a simulation method for enterprise digital development path that integrates temporal adversarial generative networks, which solves the technical problem of insufficient representation of uncertainty and randomness caused by relying solely on mechanism simulation models for deterministic prediction in existing technologies.

[0005] This invention provides a simulation method for enterprise digital development path that integrates temporal adversarial generative networks, the method comprising: Based on the target production scenario, construct a digital test case set containing multiple typical working conditions; Obtain the digital scheme to be evaluated, initialize the pre-built mechanism simulation model according to the digital scheme, and run the mechanism simulation model to perform simulation analysis on the digital test case set to obtain the deterministic benchmark prediction sequence under the corresponding test case. Using each test condition and the aforementioned digitization scheme as input conditions, a conditional time series generator of a pre-trained temporal generative adversarial network is used to generate the residual prediction sequence for the corresponding test condition, combined with a random noise sampling strategy. Using the test conditions as an index, the deterministic benchmark prediction sequence and each residual prediction sequence are superimposed step-by-step to generate multiple simulation trajectories. Statistical analysis is performed on the simulation trajectory based on all test conditions to obtain the statistical distribution and extract the corresponding evaluation indicators. The evaluation indicators are then associated with the digital scheme and output as simulation results.

[0006] Preferably, a digital test case set, constructed according to the target production scenario, includes multiple typical working conditions, including: Acquire historical operating data of the target production scenario, and extract and statistically analyze the operating condition parameters of the historical operating data to determine the original operating condition set; Acquire historical operational data from multiple production scenarios that share the same equipment type and similar processes as the target production scenario; Statistical analysis was performed on the reproducibility rate of each working condition in the historical operating data of the same production scenario. Multiple working conditions with the highest reproducibility were selected to determine the set of same working conditions. The original working condition set and the same working condition set are combined into a digital test working condition set, wherein the number of working conditions in the same working condition set is a preset ratio of the number of working conditions in the original working condition set.

[0007] Preferably, the pre-construction of the mechanism simulation model includes: Construct sub-models for equipment performance degradation, energy consumption prediction, failure rate prediction, data acquisition and communication, production scheduling, and maintenance strategy. Define the mapping relationship between the digitization parameters and the running parameters of each sub-model, where: Sensor configuration parameters are mapped to the data acquisition and communication sub-model and the energy consumption prediction sub-model; Edge computing configuration parameters and data acquisition configuration parameters are mapped to the data acquisition and communication sub-model; Production management configuration parameters are mapped to the production scheduling sub-model; The maintenance strategy configuration parameters are mapped to the maintenance strategy sub-model and the failure rate prediction sub-model; Multiple sub-models are assembled according to the process logic to obtain the mechanism simulation model, wherein the output of each sub-model serves as the input condition for the downstream sub-model.

[0008] Preferably, the process involves obtaining a digital scheme to be evaluated, initializing a pre-built mechanistic simulation model based on the digital scheme, and running the mechanistic simulation model to perform simulation analysis on the digital test case set to obtain a deterministic benchmark prediction sequence under the corresponding test case, including: Obtain the digitalization solution to be evaluated, and extract digitalization parameters from the digitalization solution, wherein the digitalization parameters include sensor configuration parameters, edge computing configuration parameters, data acquisition configuration parameters, production management configuration parameters, and maintenance strategy configuration parameters; Based on the digitized parameters, the operating parameters of each sub-model in the mechanism simulation model are initialized and set through a predefined parameter mapping relationship; The test conditions in the digital test condition set are traversed, and the condition parameters corresponding to each test condition are input into the initialized mechanism simulation model to perform simulation analysis. The system state prediction sequence is extracted from the simulation analysis results of each test condition, and the corresponding deterministic benchmark prediction sequence is constructed. The association marker between each deterministic benchmark prediction sequence and the corresponding test condition is established.

[0009] Preferably, the pre-training of the temporal generative adversarial network includes: Acquire historical operating status data, operating condition parameters, and digital parameters of multiple digitally configured production scenarios from the same source, and construct a training sample set; The historical operating state data in the training sample set is subtracted step by step from the deterministic prediction sequence output by the mechanism simulation model under the corresponding operating condition parameters and digital parameters to obtain the residual sequence. A conditional time series generator is constructed, wherein the conditional time series generator takes a random sampled noise vector as random input and operating condition parameters and digital parameters as conditional input, and generates residual prediction sequences step by step in an autoregressive manner based on a recurrent neural network; Construct a time series discriminant, which is used to distinguish between the real residual sequence and the predicted residual sequence; Using the residual sequence as the training target, the network parameters of the conditional time series generator and the time series discriminator are optimized through adversarial training. The loss function of the conditional time series generator consists of adversarial loss and physical constraint regularization loss.

[0010] Preferably, taking each test condition and the digitization scheme as input conditions, a conditional time series generator of a pre-trained time series generative adversarial network is used to generate a residual prediction sequence for the corresponding test condition, combined with a random noise sampling strategy, including: Based on the historical operating status data of the target production scenario, calculate the operating condition stability metric corresponding to each test condition, wherein the operating condition stability metric is used to characterize the degree of fluctuation of the test condition in the historical operating status data. Based on the stability metric of each test condition, the number of samplings corresponding to each test condition is determined, wherein the lower the stability metric of the test condition, the greater the number of samplings. For each test condition, the condition parameters of the test condition and the digital parameters extracted from the digital scheme are used as conditional inputs, and the random sampling noise vector is used as random input. The inputs are then fed into the conditional time series generator of the time series generative adversarial network to obtain the corresponding single residual prediction sequence. Based on the number of samplings determined for each test condition, the generation process is repeated to obtain multiple residual prediction sequences for each test condition.

[0011] Preferably, statistical analysis is performed on the simulation trajectory based on all test conditions to obtain the statistical distribution and extract corresponding evaluation indicators. The evaluation indicators are then correlated with the digital scheme and output as simulation results, including: For each test condition, the expected value and variance of the system state index are calculated based on the corresponding multiple simulation trajectories. For each test condition, based on the corresponding multiple simulation trajectories, the confidence interval of each system state index under the preset confidence level and the risk probability that the observed value of each system state index is lower than the corresponding preset target threshold are calculated, and the output is a single evaluation index. Across all test conditions, the average value of each system status indicator across all test conditions, the fluctuation of each system status indicator across all test conditions, and the maximum risk probability of each system status indicator across all test conditions are analyzed and obtained as comprehensive evaluation indicators across all test conditions. The individual evaluation indicators and the comprehensive evaluation indicators across test conditions are associated and labeled with the digital scheme, and the results are output as simulation results.

[0012] Preferably, the pre-training of the temporal generative adversarial network further includes: Acquire historical operational status data, operating condition parameters, and digital parameters from multiple production scenarios with the same origin; Using the operating condition parameters of each production scenario as input, adversarial training is performed on all historical operating status data to complete source domain pre-training and obtain the source domain pre-trained model. Historical operational status data of the target production scenario is obtained, and the working condition parameters of the target production scenario are used as input conditions. The conditional temporal generator in the source domain pre-trained model is fine-tuned with a learning rate lower than that of the source domain pre-training model to obtain the target domain fine-tuned model as the pre-trained temporal generative adversarial network.

[0013] Preferably, the enterprise digital development path simulation method integrating temporal adversarial generative networks further includes: A comprehensive evaluation objective function is constructed, wherein the comprehensive evaluation objective function takes the average value across operating conditions, the volatility across operating conditions, and the maximum risk probability as independent variables, and performs a weighted summation with preset weight coefficients; Using the comprehensive evaluation objective function as the optimization objective and the feasible region of the digitized parameters as the constraint, candidate digitized parameters are searched in the feasible region through iterative optimization. Based on the candidate digitization parameters, candidate digitization schemes are constructed, deterministic benchmark prediction sequences and residual prediction sequences are re-acquired, and time-step superposition and statistical analysis are performed until the comprehensive evaluation objective function meets the convergence condition. The corresponding digitization parameters, digitization schemes and simulation results are then output.

[0014] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention provides a simulation method for enterprise digital development paths that integrates temporal adversarial generative networks. First, by constructing a digital test case set containing various typical operating conditions and incorporating a source-based scenario screening strategy, the coverage and representativeness of the simulation verification are significantly improved, enabling the evaluation results to generalize and transfer to different production scenarios. Second, the deterministic baseline prediction output by the mechanistic simulation model is superimposed time-step by time with the residual prediction sequence output by the conditional temporal generator of the pre-trained temporal generative adversarial network. This preserves the physical consistency constraints of the mechanistic model while learning the higher-order temporal dependencies of the residuals and the heterogeneous distribution characteristics of operating conditions from historical operational data through adversarial training. This allows the generated simulation trajectory to realistically reproduce the random fluctuation characteristics of real production logs at the statistical distribution level. Third, based on the statistical distribution of multiple simulation trajectories, multi-level evaluation indicators, including expected value, variance, confidence interval, and risk probability, are extracted, providing a quantitative basis for cross-operating-condition comprehensive evaluation and effectively supporting the iterative search and optimization decision-making of digital parameters. Finally, the simulation verification of the digital development path is achieved, improving the accuracy of uncertainty while maintaining physical credibility. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the enterprise digital development path simulation method using a fusion temporal adversarial generative network provided by the present invention. Figure 2 This is a logical schematic diagram of the enterprise digital development path simulation method that integrates temporal adversarial generative networks provided by the present invention. Detailed Implementation

[0016] This invention provides a simulation method for enterprise digital development path that integrates temporal adversarial generative networks, which solves the technical problem of insufficient representation of uncertainty and randomness caused by relying solely on mechanism simulation models for deterministic prediction in existing technologies.

[0017] The present invention will now be described in detail with reference to the accompanying drawings.

[0018] In one embodiment, such as Figure 1 , Figure 2 As shown, this invention provides a simulation method for enterprise digital development path that integrates temporal adversarial generative networks, the method comprising: S100: Based on the target production scenario, construct a digital test condition set containing multiple typical working conditions; In this embodiment of the invention, a typical operating condition is a representative combination of operating states that the production line may encounter in actual operation; the digital test condition set is a collection of multiple typical operating conditions.

[0019] Specifically, historical operational data of the target production scenario is acquired. By extracting and statistically analyzing the operating condition parameters from this historical data, the frequently occurring operating patterns in actual operation of the production line are identified, forming a source operating condition set. Simultaneously, historical operational data from multiple source production scenarios with the same equipment type and similar processes as the target production scenario are acquired. The reproducibility rate of each operating condition identified from the source production scenarios in these source scenarios is statistically analyzed, and the operating conditions with the highest reproducibility rates are selected to form a source operating condition set. Finally, the source operating condition set and the source operating condition set are merged to form a complete digital test operating condition set.

[0020] In step S100 of the method of this embodiment of the invention, a digital test condition set containing multiple typical working conditions is constructed based on the target production scenario, including: Acquire historical operating data of the target production scenario, and extract and statistically analyze the operating condition parameters of the historical operating data to determine the original operating condition set; Acquire historical operational data from multiple production scenarios that share the same equipment type and similar processes as the target production scenario; Statistical analysis was performed on the reproducibility rate of each working condition in the historical operating data of the same production scenario. Multiple working conditions with the highest reproducibility were selected to determine the set of same working conditions. The original working condition set and the same working condition set are combined into a digital test working condition set, wherein the number of working conditions in the same working condition set is a preset ratio of the number of working conditions in the original working condition set.

[0021] In this embodiment of the invention, firstly, historical operating data of the target production scenario can be acquired through a data acquisition and monitoring control system. Then, corresponding numerical values ​​or category codes are extracted from the structured fields and time-series database of the historical operating data to form a time series of operating condition parameters indexed by timestamps. This completes the extraction of operating condition parameters from the historical operating data. The operating condition parameters extracted from the historical operating data are multi-dimensional and include, but are not limited to: capacity utilization rate, product type identifier, batch size, ambient temperature, equipment service life, and raw material characteristic indicators.

[0022] Statistical analysis is then performed to identify representative operating conditions that frequently occur during the actual operation of the production line. Preferably, a density-based clustering algorithm is used to cluster the operating condition parameter vectors, grouping time periods with similar operating condition parameter characteristics into the same operating condition cluster. For each operating condition cluster, its center vector and total duration are calculated. Subsequently, the clusters are sorted according to their total duration, and the clusters with the highest total duration are selected as representative operating conditions. The number of clusters selected is determined based on the actual operating mode of the target production scenario; for example, clusters covering the top 80% of the total runtime are selected. The source operating condition set consists of these representative operating conditions mined from the target production scenario's own data, representing the most frequently experienced production operation states of the production scenario itself. Finally, the typical operating conditions identified through statistical analysis are summarized to form the source operating condition set.

[0023] Secondly, by using the data interface of the Manufacturing Execution System (MES) or Industrial Internet platform, multiple source production scenarios with the same equipment type and similar processes as the target production scenario are retrieved and acquired within a preset historical time period, such as the past two years. The data structure of the historical operational data is standardized to ensure that its field definitions, sampling frequency, and data format are consistent with the historical operational data of the target production scenario.

[0024] Next, statistical analysis was conducted on the reproducibility rate of each working condition in historical operating data of the same production scenario. Several working conditions with the highest reproducibility were selected to determine the set of the same working conditions.

[0025] Specifically, a method for extracting and clustering operating condition parameters that are identical in the historical operating data of the target production scenarios is used. The historical operating data of each source production scenario is independently divided into operating condition clusters, resulting in a set of operating condition clusters for each source scenario. Subsequently, the operating condition clusters from all source production scenarios are categorized and merged across scenarios. Specifically, using the similarity between operating condition parameter vectors, such as cosine similarity or Euclidean distance, as a metric, operating condition clusters from different source scenarios but with similar operating condition parameter characteristics are grouped into the same candidate operating condition type. For each candidate operating condition type, the total number of times it appears in all source production scenarios and the number of source scenarios covered by that operating condition type are counted, and the reproducibility rate is calculated. The reproducibility rate measures the frequency and breadth of a specific operating condition type appearing in the historical operating data of source production scenarios.

[0026] Preferably, the reproducibility rate is the ratio of the total frequency of the identified working condition type in all similar scenarios to the total number of working condition clusters in all similar scenarios. All candidate working condition types are sorted from high to low according to their reproducibility rates, and a predetermined number of working conditions with the highest reproducibility rates are selected as the similar working condition set.

[0027] Finally, the original test condition set and the similar test condition set are merged, and the union is taken to form the final digital test condition set. After merging, the digital test condition set includes both the original test conditions frequently experienced by the target production scenario and similar test conditions commonly found in similar production lines but less frequently experienced by the target scenario. To ensure the structural rationality and evaluation efficiency of the test condition set, the number of similar test condition sets is constrained, i.e., the number of test conditions in the similar test condition sets is set as a preset proportion of the number of test conditions in the original test condition set. This preset proportion is set according to the coverage requirements of the actual application scenario. For example, in one embodiment, this preset proportion can be set to 100%, meaning the number of similar test conditions is equal to the number of original test conditions.

[0028] In this embodiment of the invention, historical data from similar production scenarios are used to expand the coverage of operating conditions in the evaluation of digital solutions. The similar operating conditions are screened using a reproducibility rate index to ensure the universality and representativeness of the introduced operating conditions, avoiding the waste of simulation computing resources caused by extreme and ineffective operating conditions. Finally, a preset ratio is used to constrain the size of the similar operating condition set, providing a data foundation for executing subsequent steps.

[0029] S200: Obtain the digital scheme to be evaluated, initialize the pre-built mechanism simulation model according to the digital scheme, and run the mechanism simulation model to perform simulation analysis on the digital test case set to obtain the deterministic benchmark prediction sequence under the corresponding test case. In this embodiment of the invention, the mechanism simulation model is a mathematical model based on physical equations and process rules, used to simulate the operating behavior of equipment and production lines. It is assembled from multiple sub-models such as equipment performance degradation sub-model, energy consumption prediction sub-model, failure rate prediction sub-model, data acquisition and communication sub-model, production scheduling sub-model, and maintenance strategy sub-model according to process logic. The deterministic benchmark prediction sequence is the deterministic prediction result of the system state changing over time output by the mechanism simulation model under given digital scheme parameters and test conditions.

[0030] Specifically, the process involves acquiring the digitization scheme to be evaluated and extracting digitization parameters from it, including sensor configuration parameters, edge computing configuration parameters, data acquisition configuration parameters, production management configuration parameters, and maintenance strategy configuration parameters. Then, based on the extracted digitization parameters, the operating parameters of each sub-model in the mechanistic simulation model are initialized using predefined parameter mapping relationships—for example, mapping the sensor sampling frequency to the sampling period parameter of the data acquisition sub-model, and mapping the maintenance early warning threshold to the trigger condition parameter of the maintenance strategy sub-model. Next, each test case in the digitization test case set is traversed, and the corresponding condition parameters for each test case are input into the initialized mechanistic simulation model to perform simulation analysis. Finally, system state prediction sequences are extracted from the simulation analysis results of each test case, constructed as corresponding deterministic benchmark prediction sequences, and association markers are established between each deterministic benchmark prediction sequence and its corresponding test case.

[0031] Step S200 of the method in this embodiment of the invention, the pre-construction of the mechanism simulation model, includes: Construct sub-models for equipment performance degradation, energy consumption prediction, failure rate prediction, data acquisition and communication, production scheduling, and maintenance strategy. Define the mapping relationship between the digitization parameters and the running parameters of each sub-model, where: Sensor configuration parameters are mapped to the data acquisition and communication sub-model and the energy consumption prediction sub-model; Edge computing configuration parameters and data acquisition configuration parameters are mapped to the data acquisition and communication sub-model; Production management configuration parameters are mapped to the production scheduling sub-model; The maintenance strategy configuration parameters are mapped to the maintenance strategy sub-model and the failure rate prediction sub-model; Multiple sub-models are assembled according to the process logic to obtain the mechanism simulation model, wherein the output of each sub-model serves as the input condition for the downstream sub-model.

[0032] In this embodiment of the invention, firstly, a sub-model for equipment performance degradation, a sub-model for energy consumption prediction, a sub-model for failure rate prediction, a sub-model for data acquisition and communication, a sub-model for production scheduling, and a sub-model for maintenance strategy are constructed.

[0033] The equipment performance degradation sub-model describes the performance degradation of production equipment over time due to factors such as wear, fatigue, and corrosion. Its inputs are equipment operating parameters, such as operating time, load rate, and temperature, and its output is the amount of equipment performance degradation or health indicators. The energy consumption prediction sub-model predicts energy consumption during production based on equipment operating status and production plans. Its inputs are equipment operating parameters and production scheduling information, and its output is the predicted energy consumption value for each time period. The failure rate prediction sub-model predicts the probability of future failures based on the current health status of the equipment and historical failure data. Its inputs are the amount of equipment performance degradation and historical failure statistics, and its output is a predicted curve of the failure rate over time. The data acquisition and communication sub-model... The simulation model is used to model the process of sensor data acquisition, edge computing processing, and data communication transmission in a digital system. Its inputs are sensor configuration parameters and data acquisition frequency, and its outputs are simulated data acquisition records and communication latency indicators. The production scheduling sub-model is used to generate the optimal production scheduling plan based on production orders, equipment status, and resource constraints. Its inputs are production order information, equipment availability status, and production management configuration parameters, and its outputs are the production task allocation schemes for each piece of equipment in each time period. The maintenance strategy sub-model is used to generate maintenance plans and maintenance operation instructions based on equipment health status and preset maintenance rules. Its inputs are equipment performance degradation, failure rate prediction results, and maintenance strategy configuration parameters, and its outputs are the time arrangement of maintenance tasks and specific maintenance measures.

[0034] When constructing each sub-model, it is necessary to select an appropriate modeling method based on the specific equipment type, process flow, and operating characteristics of the target production scenario. For example, for the equipment performance degradation sub-model, performance monitoring indicators from historical equipment operating data are used to estimate model parameters; for the energy consumption prediction sub-model, a quantitative relationship between equipment energy consumption and factors such as load rate, operating speed, and ambient temperature is established using a physical mechanism-based energy conservation equation combined with statistical regression methods; for the failure rate prediction sub-model, a proportional hazards model or a state-space based reliability model is used, incorporating equipment performance degradation as a covariate into the failure rate function; for the data acquisition and communication sub-model, simulation models of indicators such as data acquisition latency, packet loss rate, and transmission bandwidth are established based on the actual deployed sensor types, acquisition frequencies, communication protocols, and network topology; for the production scheduling sub-model, mixed integer programming or heuristic optimization algorithms are used to generate production scheduling schemes that meet process and resource constraints with the goal of minimizing the production cycle or maximizing equipment utilization; for the maintenance strategy sub-model, a logical rule model for maintenance decisions is established based on preset preventive maintenance cycles, condition monitoring triggering conditions, and maintenance resource constraints.

[0035] Secondly, the mapping relationship between the digitization parameters and the operating parameters of each sub-model is defined, where: First, sensor configuration parameters are mapped to the data acquisition and communication sub-model and the energy consumption prediction sub-model. In the data acquisition and communication sub-model, sensor configuration parameters are used to set the sensor type, number, sampling frequency, and accuracy level for simulated data acquisition. For example, when the sensor configuration parameters include temperature sensors, vibration sensors, and current sensors, the data acquisition and communication sub-model will generate simulated time series data for temperature, vibration, and current according to the set sampling frequencies. In the energy consumption prediction sub-model, sensor configuration parameters are used to determine the available energy consumption-related monitoring variables. For example, when power sensors are deployed, the energy consumption prediction sub-model can directly use power monitoring data to calculate energy consumption. When no power sensors are deployed, energy consumption values ​​need to be estimated through indirect estimation methods.

[0036] Secondly, edge computing configuration parameters and data acquisition configuration parameters are mapped to the data acquisition and communication sub-model. Edge computing configuration parameters are used to set the processing power of edge computing nodes, data cache size, and the type of data processing algorithm at the edge end in the data acquisition and communication sub-model. For example, if the edge computing configuration parameters specify that a sliding window mean filtering algorithm is used for data preprocessing, the data acquisition and communication sub-model will filter the raw acquired data at the edge end before uploading it to the cloud. Data acquisition configuration parameters are used to set the data acquisition time interval, data upload cycle, and data compression method. For example, if the data acquisition configuration parameters specify an acquisition frequency of once per second and an upload cycle of once per minute, the data acquisition and communication sub-model will generate acquisition records according to this frequency and package and upload the data periodically.

[0037] Secondly, production management configuration parameters are mapped to the production scheduling sub-model. These parameters are used to set order scheduling rules, priority allocation strategies, and resource scheduling constraints within the production scheduling sub-model. For example, if the production management configuration parameters specify a first-come, first-served scheduling rule and an urgent order priority insertion strategy, the production scheduling sub-model will generate production task sequences for each device according to these rules. Similarly, if the production management configuration parameters specify the maximum available working hours and changeover time constraints for the devices, the production scheduling sub-model will incorporate these constraints into the optimization model for scheduling calculations.

[0038] Furthermore, maintenance strategy configuration parameters are mapped to the maintenance strategy sub-model and the failure rate prediction sub-model. In the maintenance strategy sub-model, these parameters are used to set the cycle length of preventative maintenance, the trigger threshold for condition monitoring maintenance, and the available number of maintenance resources. For example, if the maintenance strategy configuration parameters specify preventative maintenance every 200 hours of operation and trigger condition maintenance when the equipment health level is below 0.7, the maintenance strategy sub-model will determine whether to generate a maintenance task based on the equipment's cumulative runtime and real-time health level. In the failure rate prediction sub-model, the maintenance strategy configuration parameters are used to set the maintenance effectiveness coefficient in the failure rate model, i.e., the degree of equipment failure rate recovery after each maintenance. For example, if the maintenance strategy configuration parameters specify a full maintenance or minimum maintenance strategy, the failure rate prediction sub-model will adjust the failure rate prediction curve accordingly after each maintenance event.

[0039] Finally, multiple sub-models are assembled according to the process logic to obtain a mechanism simulation model. The output of each sub-model serves as the input condition for the downstream sub-model. The process logic is the sequence of procedures and material flow relationships in the entire production process from raw material input to product output in the target production scenario.

[0040] Specifically, during assembly, the data flow and temporal dependencies between each sub-model are determined according to the production process. Then, each sub-model is connected in a forward propagation manner, so that the output of the upstream sub-model is used as the input parameter of the downstream sub-model, forming a simulation model of the entire production system's operating status.

[0041] Preferably, the equipment health index output by the equipment performance degradation sub-model is simultaneously input into the failure rate prediction sub-model and the maintenance strategy sub-model to predict equipment failure probability and trigger maintenance decisions; the failure rate prediction curve output by the failure rate prediction sub-model is input into the production scheduling sub-model to consider equipment availability constraints during scheduling; the production task allocation scheme output by the production scheduling sub-model is input into the energy consumption prediction sub-model to calculate production energy consumption for each time period; the maintenance task arrangement output by the maintenance strategy sub-model is input into the production scheduling sub-model to update the equipment availability time window; and the simulated acquisition data output by the data acquisition and communication sub-model serves as the source of status monitoring data required for the operation of each sub-model. Through assembly, the six sub-models are integrated into a unified mechanism simulation model, which receives digital parameters and operating condition parameters as input and outputs a deterministic state prediction sequence of the entire production system under a given digital configuration.

[0042] In this embodiment of the invention, by constructing sub-models for equipment performance degradation, energy consumption prediction, failure rate prediction, data acquisition and communication, production scheduling, and maintenance strategy, and defining the mapping relationship between digital parameters and the operating parameters of each sub-model, the mechanism simulation model can comprehensively represent multi-dimensional production factors, enabling the simulation process to reflect the impact of digital solutions on each production link, and improving the systematicness and credibility of the simulation results.

[0043] S300: Using each test condition and the digital scheme as input conditions, the conditional time series generator of the pre-trained temporal generative adversarial network is combined with a random noise sampling strategy to generate the residual prediction sequence under the corresponding test condition. In this embodiment of the invention, the temporal generative adversarial network (GAN) is a deep learning architecture that extends GAN to sequential data scenarios. It uses a recurrent neural network as its backbone network and can generate sequential data with temporal structure characteristics. The conditional temporal generator is the generator component in the temporal GAN. It receives three inputs: a random noise vector, a working condition vector, and digital scheme parameters. It generates a residual prediction sequence step by step in an autoregressive manner. The residual prediction sequence is a prediction sequence of the deviation between the deterministic benchmark prediction of the mechanism simulation model and the actual operating state. The random noise sampling strategy is a sampling scheme that samples random noise vectors multiple times from a standard normal distribution to generate multiple different residual prediction sequences.

[0044] Specifically, based on historical operational status data of the target production scenario, a stability metric is calculated for each test condition. This metric characterizes the degree of fluctuation of the test condition within the historical operational status data. Then, the sampling count for each test condition is determined based on its stability metric. For each test condition, the conditional parameters and digitized parameters extracted from the digitization scheme are used as conditional inputs, and a randomly sampled noise vector is used as random input. This input is then fed into a pre-trained conditional time series generator of a temporal generative adversarial network to obtain a single residual prediction sequence. Based on the determined sampling count for each test condition, the above generation process is repeated to obtain multiple residual prediction sequences for each test condition.

[0045] In step S300 of the method of this embodiment of the invention, the method involves obtaining a digital scheme to be evaluated, initializing a pre-constructed mechanism simulation model based on the digital scheme, and running the mechanism simulation model to perform simulation analysis on the digital test case set to obtain a deterministic benchmark prediction sequence under the corresponding test case, including: Obtain the digitalization solution to be evaluated, and extract digitalization parameters from the digitalization solution, wherein the digitalization parameters include sensor configuration parameters, edge computing configuration parameters, data acquisition configuration parameters, production management configuration parameters, and maintenance strategy configuration parameters; Based on the digitized parameters, the operating parameters of each sub-model in the mechanism simulation model are initialized and set through a predefined parameter mapping relationship; The test conditions in the digital test condition set are traversed, and the condition parameters corresponding to each test condition are input into the initialized mechanism simulation model to perform simulation analysis. The system state prediction sequence is extracted from the simulation analysis results of each test condition, and the corresponding deterministic benchmark prediction sequence is constructed. The association marker between each deterministic benchmark prediction sequence and the corresponding test condition is established.

[0046] In this embodiment of the invention, firstly, a digital scheme to be evaluated is obtained, and digital parameters are extracted from the digital scheme. The digital parameters include sensor configuration parameters, edge computing configuration parameters, data acquisition configuration parameters, production management configuration parameters, and maintenance strategy configuration parameters.

[0047] Specifically, a digital solution is a comprehensive plan describing the technological means, equipment configuration, and management strategies adopted by an enterprise during its digital transformation process. It is manifested as a set of configurable and adjustable parameters. Digital parameters are quantifiable configuration items extracted from the digital solution that can be input into a mechanism simulation model for quantitative simulation.

[0048] Preferably, sensor configuration parameters describe information such as the type, installation location, measurement range, sampling accuracy, and deployment density of various sensors deployed on the production site, for example, the number, placement, and sampling accuracy of temperature sensors; edge computing configuration parameters describe information such as the computing power, storage capacity, data processing algorithm type, and model update frequency of edge computing nodes deployed near the data source; data acquisition configuration parameters describe information such as the data acquisition time interval, data format specifications, transmission protocol type, data compression method, and storage strategy; production management configuration parameters describe information such as production order scheduling rules, equipment scheduling priority strategies, work-in-process inventory management rules, and order delivery cycle constraints; maintenance strategy configuration parameters describe information such as equipment maintenance cycle setting rules, maintenance triggering conditions, maintenance resource allocation strategies, and maintenance effect evaluation standards. The digital solution can be directly extracted from the enterprise's digital planning document.

[0049] Secondly, based on the digital parameters, the operating parameters of each sub-model in the mechanism simulation model are initialized and set through predefined parameter mapping relationships. These predefined parameter mapping relationships are a set of conversion rules established in advance during the mechanism simulation model construction phase to convert configuration parameters at the digital solution level into operating parameters within each sub-model.

[0050] Since the parameters in the digital solution are high-level strategy configuration descriptions, while each sub-model requires specific model operation parameters, it is necessary to convert the predefined parameter mapping relationship into the operation parameters of each sub-model through mapping relationships.

[0051] Specifically, sensor configuration parameters are mapped to a data acquisition and communication sub-model and an energy consumption prediction sub-model. In the data acquisition and communication sub-model, the sensor type, number, and sampling accuracy in the sensor configuration parameters determine the dimension and quality of the generated simulation data. For example, when the sensor configuration parameters include temperature and vibration sensors, the data acquisition and communication sub-model will generate simulated acquisition data containing temperature and vibration time series, while the sampling accuracy parameter determines the level of noise added to the generated data.

[0052] In the energy consumption prediction sub-model, sensor configuration parameters determine the types of energy consumption-related monitoring variables that can be directly acquired. For example, when deploying power sensors, the energy consumption prediction sub-model can directly calculate real-time energy consumption values ​​using power monitoring data. However, when only temperature and vibration sensors are deployed, energy consumption values ​​need to be estimated through an indirect regression model. Edge computing configuration parameters and data acquisition configuration parameters are mapped to the data acquisition and communication sub-model. The data processing algorithm type in the edge computing configuration parameters determines how the data acquisition and communication sub-model preprocesses the raw data at the edge, such as the filter window length, outlier removal threshold, and data compression ratio. The acquisition frequency and upload cycle in the data acquisition configuration parameters determine the temporal resolution of data sampling and the time interval for data transmission.

[0053] The production management configuration parameters are mapped to the production scheduling sub-model. The scheduling rules and priority strategies in the production management configuration parameters determine the composition of the optimization objective function and the setting of constraints in the production scheduling sub-model. For example, the first-come-first-served rule corresponds to the optimization objective of minimizing the average waiting time, and the emergency order priority rule corresponds to the scheduling strategy of prioritizing the weighted shortest processing time.

[0054] Subsequently, the maintenance strategy configuration parameters are mapped to the maintenance strategy sub-model and the failure rate prediction sub-model. In the maintenance strategy sub-model, the cycle length and trigger threshold in the maintenance strategy configuration parameters determine the triggering conditions and timing of maintenance tasks. In the failure rate prediction sub-model, the maintenance effect coefficient in the maintenance strategy configuration parameters determines the degree of recovery of the equipment failure rate after each maintenance.

[0055] For example, assuming the temperature sensor accuracy is ±0.5℃, the sensor noise parameter of the data acquisition and communication sub-model is set to a normally distributed random noise with a mean of 0 and a standard deviation of 0.5 / 3, based on this accuracy. Assuming a sliding window mean filtering algorithm is used with a window length of 10, the data acquisition and communication sub-model will calculate the mean of every 10 consecutively acquired data points at the edge and output the filtering result. Assuming the acquisition frequency is once per second and the upload cycle is once every 60 seconds, the data acquisition and communication sub-model will generate simulated data with a time step of 1 second, and upload data in batches of 60 data points. Assuming the maximum available working hours of the equipment are 16 hours per day, the production scheduling sub-model will use 16 hours per day as the time boundary constraint for scheduling optimization, and no production tasks will be arranged beyond this time range. Assuming preventive maintenance is performed every 200 hours of operation, the maintenance strategy sub-model will trigger maintenance task generation events when the cumulative operating time of the equipment reaches 200 hours, 400 hours, 600 hours, etc.

[0056] Next, the system iterates through all test conditions in the digital test condition set, inputting the corresponding condition parameters into the initialized mechanism simulation model to perform simulation analysis. This refers to the mechanism simulation model calculating the state evolution process of the entire production system step-by-step, according to the mathematical logic of each sub-model and the data transfer relationships between sub-models, under given input conditions.

[0057] Specifically, for each test condition, the condition parameters of that condition are input into the production scheduling sub-model.

[0058] The production scheduling sub-model generates a production task schedule based on production management configuration parameters and order information in the operating conditions; the initial health status of the equipment is input into the equipment performance degradation sub-model, which calculates the performance degradation step by step based on the equipment's operating time; the health index output by the equipment performance degradation sub-model is input into the failure rate prediction sub-model, which calculates the equipment failure rate over time by combining the maintenance effectiveness coefficient in the maintenance strategy configuration parameters; the failure rate information output by the failure rate prediction sub-model is fed back to the production scheduling sub-model; the production task allocation scheme output by the production scheduling sub-model is input into the energy consumption prediction sub-model, which calculates the corresponding energy consumption value based on the load rate of each piece of equipment in each time period; the maintenance strategy sub-model determines whether to trigger a maintenance task based on the equipment health and cumulative operating time, and feeds back the maintenance task arrangement to the production scheduling sub-model to update the equipment's available time window; the data acquisition and communication sub-model generates simulated sensor acquisition data according to the sensor configuration parameters and data acquisition configuration parameters, serving as the status monitoring data source required for the operation of each sub-model.

[0059] Through collaborative simulation of multiple sub-models, the mechanism simulation model simulates the entire production process from the initial moment to the simulation termination moment on a complete time axis.

[0060] Finally, system state prediction sequences are extracted from the simulation analysis results of each test condition, corresponding deterministic baseline prediction sequences are constructed, and association markers are established between each deterministic baseline prediction sequence and its corresponding test condition. The system state prediction sequence refers to the time-series data output by the mechanistic simulation model, describing the state evolution of the production system along the simulation timeline. The deterministic baseline prediction sequence is the system state prediction sequence calculated by the mechanistic simulation model based on physical mechanisms and deterministic mathematical logic under given digital schemes and operating conditions.

[0061] Specifically, the simulation analysis results output by the mechanistic simulation model include a dataset of time series data for multiple system state variables. According to a pre-defined list of system state indicators, the values ​​of each indicator at each time step are read from this dataset and arranged chronologically to form a complete time series for each system state indicator. The time series of all these indicators are combined to form a multi-dimensional time series data matrix, yielding the deterministic benchmark prediction sequence corresponding to the test condition. Subsequently, this deterministic benchmark prediction sequence is associated with the test condition identifier and stored accordingly, for example, establishing a key-value pair mapping relationship using the test condition number as the key and the deterministic benchmark prediction sequence as the value.

[0062] In step S300 of the method of this embodiment of the invention, the pre-training of the temporal generative adversarial network includes: Acquire historical operating status data, operating condition parameters, and digital parameters of multiple digitally configured production scenarios from the same source, and construct a training sample set; The historical operating state data in the training sample set is subtracted step by step from the deterministic prediction sequence output by the mechanism simulation model under the corresponding operating condition parameters and digital parameters to obtain the residual sequence. A conditional time series generator is constructed, wherein the conditional time series generator takes a random sampled noise vector as random input and operating condition parameters and digital parameters as conditional input, and generates residual prediction sequences step by step in an autoregressive manner based on a recurrent neural network; Construct a time series discriminant, which is used to distinguish between the real residual sequence and the predicted residual sequence; Using the residual sequence as the training target, the network parameters of the conditional time series generator and the time series discriminator are optimized through adversarial training. The loss function of the conditional time series generator consists of adversarial loss and physical constraint regularization loss.

[0063] In this embodiment of the invention, firstly, the historical operating status data, operating condition parameters, and digital parameters of multiple identical production scenarios that have been digitally configured are obtained during actual operation.

[0064] Historical operational status data, operating condition parameters, and digital parameters of various shared production scenarios at the same time are aligned according to timestamps to form structured training samples. For each shared production scenario, its historical operational status data sequence, corresponding operating condition parameter sequence, and digital solution parameters are associated and combined to construct a training sample. All training samples from all shared production scenarios are aggregated to form a training sample set.

[0065] Secondly, the historical operating status data in the training sample set is subtracted from the deterministic prediction sequence output by the mechanism simulation model under the corresponding operating condition parameters and digital parameters step by step to obtain the residual sequence.

[0066] Specifically, for each sample in the training sample set, its corresponding operating condition parameters and digital parameters are extracted. These parameters are then input into a pre-constructed mechanistic simulation model to run the model. The mechanistic simulation model is a digital twin model built based on physical equations and process rules. It can output a deterministic system state prediction sequence under the input operating condition parameters and digital scheme parameters through the collaborative computation of internal sub-models: equipment performance degradation, energy consumption prediction, failure rate prediction, data acquisition and communication, production scheduling, and maintenance strategy.

[0067] The deterministic prediction sequence refers to the unique and deterministic sequence of state values ​​output by the mechanistic simulation model under a given input, which does not contain any random factors. The historical operating state data is subtracted from the deterministic prediction sequence at each identical time step. The difference between the actual observed values ​​and the predicted values ​​of the mechanistic model is calculated at each time step to obtain the residual sequence. The residual sequence represents the deviation between the deterministic prediction of the mechanistic simulation model and the actual operating state. This deviation originates from random factors that cannot be fully described by the mechanistic equations, such as equipment aging drift, random fluctuations in operating conditions, and multivariate coupling noise.

[0068] Next, a conditional time series generator is constructed, which takes a randomly sampled noise vector as random input and operating condition parameters and digitized parameters as conditional input, and generates residual prediction sequences step by step in an autoregressive manner based on a recurrent neural network.

[0069] Specifically, a conditional time series generator is constructed, which is a trainable deep neural network model whose function is to generate a synthetic residual sequence with similar distribution characteristics to the real residual sequence given a conditional input.

[0070] The conditional timing generator receives three types of inputs. The first type is a random sampled noise vector, which is a vector randomly sampled from a standard normal distribution. Its dimension is a preset value, for example, 128 as recommended. This noise vector provides a source of randomness for the generation process, enabling the generator to produce different residual sequences under the same conditions, simulating the diversity of random disturbances. The second type is operating condition parameters, which are feature vectors describing the operating conditions of the production scenario. The third type is digitization parameters, which are configuration parameter vectors of the digitization scheme to be evaluated.

[0071] The conditional time series generator is built upon a recurrent neural network, preferably using a long short-term memory network as its backbone. A recurrent neural network is a neural network structure with recurrent connections, capable of handling variable-length sequence data and capturing the temporal dependencies within the sequence. The autoregressive approach means that when generating the residual value at each time step, the generator uses the residual value of the previous time step as one of the inputs for the current step, thus achieving the stepwise generation of the sequence.

[0072] For example, using LSTM as the backbone of a recurrent neural network, the generation process of the conditional time series generator is as follows: The randomly sampled noise vector is encoded into a noise encoding vector through a first fully connected mapping layer; the operating condition parameters are encoded into an operating condition encoding vector through a second fully connected mapping layer; and the digitized parameters are encoded into a parameter encoding vector through a third fully connected mapping layer. Next, the noise encoding vector and the operating condition encoding vector are added and fused to form the initial hidden state of the LSTM network, with the initial cell state of the LSTM set to zero. Then, at each time step, the residual vector generated in the previous time step is concatenated with the parameter encoding vector to form the input vector for the current step, which is input into the LSTM network to update the hidden state. Finally, the output of the LSTM is converted into the residual vector for the current time step through an output fully connected mapping layer.

[0073] Furthermore, a time series discriminator is constructed to distinguish between the actual residual sequence and the predicted residual sequence.

[0074] Specifically, a time series discriminator is constructed. The time series discriminator is a trainable deep neural network model whose function is to distinguish whether the input residual sequence is a residual sequence derived from real data or a residual prediction sequence generated by a conditional time series generator.

[0075] The temporal discriminator employs a bidirectional long short-term memory (LSTM) network as its feature extraction backbone. The LSM network captures temporal dependencies and extracts sequence features from both the forward and reverse directions of the sequence. The output of the BiLSTM is processed through multi-scale one-dimensional convolutional layers for feature extraction. These layers use kernels of different sizes to capture short-term and medium-term temporal patterns respectively, and then concatenate the features from different scales. The concatenated features are compressed into a fixed-dimensional feature vector by a global average pooling layer, and finally, a scalar discriminant value is output through a fully connected discriminant head layer.

[0076] The discriminant value represents the degree of confidence the discriminator has in the input residual sequence being a true residual sequence. During adversarial training, the discriminator aims to classify the true residual sequence as "true" and the generator-generated residual prediction sequence as "false"; while the generator aims to generate residual prediction sequences that can "deceive" the discriminator into classifying them as "true." Through this adversarial game, the generator gradually learns the data distribution of the true residual sequence.

[0077] Finally, using the residual sequence as the training target, adversarial training is conducted on the conditional time series generator and the time series discriminator to optimize their network parameters. The adversarial training is a process where the generator and discriminator are alternately optimized. In each iteration, the discriminator's parameters are updated first, followed by the generator's parameters, allowing them to co-evolve in the adversarial game.

[0078] Specifically, a Wasserstein generative adversarial network combined with a gradient penalty framework is used for adversarial training. The generator's total loss function consists of adversarial loss and physical constraint regularization loss. During training, a progressive training strategy is employed: the generator is trained for 200 epochs using only adversarial loss to initially learn the basic distribution characteristics of the residual sequence; subsequently, physical constraint regularization loss is gradually introduced over 100 epochs, with each constraint weight increasing linearly from zero to a preset target value; finally, all loss terms are used for fine-tuning over 50 epochs, with the learning rate gradually decreasing. This progressive training avoids constraints interfering with the normal convergence of adversarial learning in the early stages of training, while ensuring that the final generated residual sequence possesses both data authenticity and meets physical consistency requirements.

[0079] The loss function of the conditional time series generator consists of adversarial loss and physical constraint regularization loss.

[0080] Specifically, the physical constraint regularization loss includes the output bounded constraint loss, the physical relationship constraint loss between variables, and the state transition smoothing constraint loss. That is, the loss L=L 对抗 +λ1×L 有界约束 +λ2×L 变量间物理关系约束 +L 状态转移平滑约束×λ3. The adversarial loss aims to enable the residual sequence generated by the generator to deceive the discriminator. Here, λ1, λ2, and λ3 are the penalty coefficients for the physical constraint regularization loss. These penalty coefficients balance the relative importance of the adversarial loss and the various physical constraint regularization losses in the total loss function; their values ​​directly affect the training convergence performance of the conditional time series generator and the physical consistency of the final generated residual sequence.

[0081] In this embodiment, the penalty coefficients λ1, λ2, and λ3 are determined using a hyperparameter grid search method based on a validation set. Grid search is a systematic hyperparameter optimization method. Its basic principle is to train the model for each possible hyperparameter combination within a predefined hyperparameter value space, evaluate the model performance of each combination on an independent validation set, and finally select the hyperparameter combination that optimizes the predefined evaluation index on the validation set as the final penalty coefficient value. For example, the combination λ1=0.2, λ2=0.5, λ3=0.3 has the optimal generation quality index when the physical constraint violation rate is only 0.5%, and is therefore selected as the optimal penalty coefficient combination.

[0082] In step S300 of the method of this embodiment of the invention, each test condition and the digitization scheme are used as input conditions. A conditional time series generator of a pre-trained time series generative adversarial network is used, combined with a random noise sampling strategy, to generate a residual prediction sequence for the corresponding test condition. This includes: Based on the historical operating status data of the target production scenario, calculate the operating condition stability metric corresponding to each test condition, wherein the operating condition stability metric is used to characterize the degree of fluctuation of the test condition in the historical operating status data. Based on the stability metric of each test condition, the number of samplings corresponding to each test condition is determined, wherein the lower the stability metric of the test condition, the greater the number of samplings. For each test condition, the condition parameters of the test condition and the digital parameters extracted from the digital scheme are used as conditional inputs, and the random sampling noise vector is used as random input. The inputs are then fed into the conditional time series generator of the time series generative adversarial network to obtain the corresponding single residual prediction sequence. Based on the number of samplings determined for each test condition, the generation process is repeated to obtain multiple residual prediction sequences for each test condition.

[0083] In this embodiment of the invention, firstly, based on the historical operating status data of the target production scenario, a stability metric is calculated for each test condition. This stability metric is a quantitative indicator used to characterize the degree of fluctuation of the test condition within the historical operating status data. A lower stability metric indicates greater system state fluctuation and stronger randomness under that condition; a higher stability metric indicates a more stable system state and stronger regularity under that condition.

[0084] Specifically, for each test condition in the digital test condition set, the time period data belonging to that condition is selected from the historical operating status data of the target production scenario, and the statistical fluctuation characteristics of each system status index within that time period are calculated. Then, the coefficient of variation of the system status index under that condition is calculated, that is, the ratio of the standard deviation to the mean; or the root mean square value of the system status change at adjacent time steps under that condition is calculated as a measure of the stability of the condition.

[0085] Secondly, based on the stability metric of each test condition, the sampling count for each test condition is determined, with lower stability metrics corresponding to higher sampling counts. The sampling count refers to the number of times the residual prediction sequence is repeatedly generated using the conditional time series generator for each test condition; that is, the number of residual prediction sequences ultimately obtained under that condition. A higher sampling count results in more accurate statistical distributions and more reliable confidence interval estimations when performing statistical analysis based on multiple simulation trajectories, but also increases computational overhead.

[0086] The lower the stability metric of a test condition, the greater the number of samples required. This is because under conditions with low stability metrics, the system state fluctuates wildly and is highly random, requiring more samples to fully capture the uncertainty distribution characteristics under that condition in order to ensure the statistical reliability of the evaluation results.

[0087] Specifically, the number of samples for each operating condition can be determined according to a preset sampling number mapping rule. For example, the stability metric value of the operating condition can be divided into multiple levels in descending order, with each level corresponding to a sampling number value, and the lowest stability metric level corresponding to the largest number of samples.

[0088] Furthermore, for each test condition, the conditional parameters of the test condition and the digital parameters extracted from the digital scheme are used as conditional inputs, and the random sampled noise vector is used as random input. These are input to the conditional time series generator of the time series generative adversarial network to obtain the corresponding single residual prediction sequence.

[0089] Specifically, the operating condition parameters of the test condition and the digital parameters extracted from the digital scheme are feature-encoded to form a conditional feature vector. A noise vector is randomly sampled from a preset noise distribution, and this noise vector, along with the conditional feature vector, is input into a conditional time series generator. The conditional time series generator uses a recurrent neural network to generate residual prediction sequences step-by-step in an autoregressive manner. The generated value at each time step depends on the generated value at the previous time step, the conditional feature at the current time step, and the noise input at the current time step. The hidden state of the recurrent neural network is updated progressively over time to ensure that the generated residual prediction sequence has temporal dependence and dynamic consistency. The length of the output residual prediction sequence is the same as the length of the deterministic benchmark prediction sequence.

[0090] For example, suppose the operating conditions of a certain test scenario are: order quantity 800 units, delivery period 5 days, initial equipment health 0.90, and digital solution parameters are: temperature sensor accuracy ±0.5℃, acquisition frequency once per second, and preventive maintenance cycle 200 hours. The conditional parameters are encoded into a 64-dimensional conditional feature vector. A 100-dimensional noise vector is randomly sampled from a standard normal distribution. This conditional feature vector and the noise vector are input into a conditional time series generator constructed based on a Long Short-Term Memory (LSTM) network. This generator contains two layers of LSM networks, each containing 128 hidden units. At the first time step, the generator initializes the decoding state based on the conditional feature vector and the noise vector, generating the residual value for the first time step. At the second time step, the generator generates the residual value for the current time step based on the conditional feature vector, the current noise vector component, and the residual value from the previous time step. This process continues, generating a residual prediction sequence of 1000 time steps, including multiple channels such as equipment health residual sequence, energy consumption residual sequence, and production residual sequence.

[0091] Finally, based on the number of samplings determined for each test condition, the generation process is repeated to obtain multiple residual prediction sequences for each test condition.

[0092] Specifically, for the same test condition, keeping the condition parameters and digitization parameters unchanged, the following operations are performed cyclically according to the sampling number N for that condition: a new noise vector is randomly sampled from the noise distribution, and the noise vector and the condition feature vector of that condition are input into the conditional timing generator to generate a residual prediction sequence; the generated residual prediction sequence is associated with the condition identifier of that condition and stored; the above process is repeated until N residual prediction sequences are generated.

[0093] Finally, multiple residual prediction sequences are obtained for each test condition. The condition with the lower stability metric has more residual prediction sequences. When multiple simulation trajectories are generated by superimposing these sequences, the quantification of the volatile condition is more thorough.

[0094] In this embodiment of the invention, the stability metric of each test condition is calculated based on historical operational status data of the target production scenario, and the inherent volatility of each condition is quantitatively assessed. The sampling frequency for each condition is determined based on the stability metric, ensuring that conditions with lower stability metrics receive more sampling frequencies. With a fixed total sampling budget, more computational resources are allocated to conditions with greater uncertainty, improving sampling efficiency and the accuracy of uncertainty quantification for key conditions. A residual prediction sequence is generated by using the test condition parameters and the digital parameters of the digital scheme as input conditions, combined with a random sampling noise vector input conditional time series generator. By repeatedly executing the generation process for each test condition, multiple residual prediction sequences are obtained, providing a sufficient data foundation for subsequent time-step superposition to generate multiple simulation trajectories.

[0095] S400: Using the test conditions as an index, the deterministic benchmark prediction sequence and each of the residual prediction sequences are superimposed step-by-step to generate multiple simulation trajectories. In this embodiment of the invention, time-step superposition is the operation of adding the deterministic baseline prediction value output by the mechanism simulation model to the residual prediction value generated by the temporal generative adversarial network at each time step; the simulation trajectory is a complete sequence that reflects the evolution of the operating state of the digital scheme under the influence of random factors over time, obtained by superimposing the deterministic baseline prediction sequence and the residual prediction sequence.

[0096] Specifically, using the test conditions as indices, for each test condition, the deterministic baseline prediction sequence corresponding to that condition is superimposed on each residual prediction sequence under that condition, step by step. Specifically, for a test condition at a given time step, the deterministic state prediction value output by the mechanism simulator is added to the residual value of a certain residual prediction sequence at that time step to obtain the simulation state value for that time step. After performing this operation on all time steps, a complete simulation trajectory is obtained. The superposition operation is then performed on multiple residual prediction sequences under the same test condition to generate multiple simulation trajectories for that condition. After traversing all test conditions, a set of simulation trajectories for all test conditions is obtained.

[0097] In this embodiment of the invention, multiple simulation trajectories with fusion mechanism and data-driven bias compensation are generated by superimposing the deterministic benchmark prediction sequence and the residual prediction sequence step by step. This allows the simulation results to simultaneously contain physical consistency and random uncertainty, providing a diverse sample basis for the subsequent extraction of evaluation indicators based on trajectory statistical distribution.

[0098] S500: Perform statistical analysis on the simulation trajectory based on all test conditions, obtain the statistical distribution and extract the corresponding evaluation indicators, and output the evaluation indicators as simulation results by associating them with the digital scheme.

[0099] In this embodiment of the invention, statistical analysis is the process of performing mathematical statistical processing on the set of simulation trajectories; evaluation indicators are quantitative performance metrics used to measure the operational effectiveness of digital solutions, including system status indicators such as overall equipment efficiency, unit product energy consumption, and failure downtime rate, as well as their statistical characteristics; statistical distribution refers to the distribution of values ​​of the simulation trajectory on each system status indicator, reflecting the distribution of the effectiveness of the digital solution under the influence of random factors.

[0100] Specifically, for each test condition, the expected value and variance of each system state index are calculated based on multiple simulation trajectories corresponding to that condition. For each test condition, based on the corresponding multiple simulation trajectories, the confidence interval of each system state index at a preset confidence level and the risk probability that the observed value of each system state index is lower than the corresponding preset target threshold are calculated, and these are output as individual evaluation indicators. Subsequently, across all test conditions, the cross-condition average value, cross-condition volatility, and maximum risk probability of each system state index across all test conditions are obtained as cross-condition comprehensive evaluation indicators. Finally, the individual evaluation indicators and cross-condition comprehensive evaluation indicators for each test condition are associated and labeled with the digital scheme, and the simulation results are output.

[0101] Step S500 of the method in this embodiment of the invention involves performing statistical analysis on the simulation trajectory based on all test conditions, obtaining the statistical distribution, extracting corresponding evaluation indicators, and associating the evaluation indicators with the digital scheme to output the simulation results, including: For each test condition, the expected value and variance of the system state index are calculated based on the corresponding multiple simulation trajectories. For each test condition, based on the corresponding multiple simulation trajectories, the confidence interval of each system state index under the preset confidence level and the risk probability that the observed value of each system state index is lower than the corresponding preset target threshold are calculated, and the output is a single evaluation index. Across all test conditions, the average value of each system status indicator across all test conditions, the fluctuation of each system status indicator across all test conditions, and the maximum risk probability of each system status indicator across all test conditions are analyzed and obtained as comprehensive evaluation indicators across all test conditions. The individual evaluation indicators and the comprehensive evaluation indicators across test conditions are associated and labeled with the digital scheme, and the results are output as simulation results.

[0102] In this embodiment of the invention, firstly, each test condition is traversed, and based on the corresponding multiple simulation trajectories, the expected value and variance of the system status indicators are calculated. Taking a production line as an example, the system status indicators include one or more of the following: overall equipment efficiency, unit product energy consumption, downtime rate, and product quality pass rate.

[0103] Specifically, under a certain test condition, N simulated trajectories are generated through multiple random noise samplings, and the time length of each trajectory is T. For the system state index m, N observations can be obtained at the t-th time step. Then, the expected value at that time step is: The variance is: Where N represents the sample size, X it The observed values ​​are represented by these values. The above calculations yield the expected value curve and variance curve for each system state index over the entire simulation time axis, providing fundamental statistics for subsequent interval estimation and risk measurement.

[0104] Secondly, for each test condition, based on the corresponding multiple simulation trajectories, the confidence interval of each system state index under the preset confidence level and the risk probability that the observed value of each system state index is lower than the corresponding preset target threshold are calculated, and the output is a single evaluation index.

[0105] Specifically, based on the expected value and variance, and using the assumptions of a t-distribution or normal distribution, the upper and lower limits of the confidence interval for each system state index at each time step are calculated at a preset confidence level, such as 95%. For example, at a confidence level of 95%, the confidence interval for a certain time step can preferably be [μ]. t -1.96σ t ,μ t +1.96σ t The narrower this range, the less uncertain the simulation results are, and the higher the stability of the digital solution.

[0106] For indicators with a "positive" direction, the risk probability refers to the probability that the observed value is lower than a preset lower threshold; for indicators with a "negative" direction, the risk probability refers to the probability that the observed value is higher than a preset upper threshold. For each time step, the number *l* of N simulated trajectories that meets the "worst than the target threshold" condition is counted, and the ratio of this number to the total number of trajectories *N* is used as the empirical risk probability estimate for that time step. By comprehensively analyzing the expected values, confidence intervals, and risk probabilities for all time steps, an evaluation index for this test condition is formed.

[0107] Furthermore, across all test conditions, the average value of each system status indicator across all test conditions, the fluctuation of each system status indicator across all test conditions, and the maximum risk probability of each system status indicator across all test conditions are analyzed and obtained as comprehensive evaluation indicators across all test conditions.

[0108] Specifically, the average value of the expected value for each test condition is calculated by averaging it over the time axis across all test conditions. The formula is as follows: Where K is the total number of test conditions, μ j Let be the mean expected value of a certain indicator during the stable phase under the j-th test condition. This indicator reflects the average effectiveness of the digital solution under various production scenarios. It is measured using standard deviation or range, and can be calculated as follows: The smaller this value, the stronger the adaptability of the solution to different production scenarios and the more stable its performance. The maximum risk probability refers to the highest risk probability value that can be observed across all time steps under each test condition, i.e., P. max =max j , t (P), this indicator is used to identify the maximum risk exposure of a digital solution under the most unfavorable scenario, and is a key basis for assessing the security and robustness of the solution.

[0109] Finally, after calculating all individual and comprehensive evaluation indicators, the individual evaluation indicators and the cross-condition comprehensive evaluation indicators for each test condition are associated and labeled with the digital scheme currently being evaluated, forming a complete simulation result dataset.

[0110] In the method of this embodiment of the invention, the pre-training of the temporal generative adversarial network further includes: Acquire historical operational status data, operating condition parameters, and digital parameters from multiple production scenarios with the same origin; Using the operating condition parameters of each production scenario as input, adversarial training is performed on all historical operating status data to complete source domain pre-training and obtain the source domain pre-trained model. Historical operational status data of the target production scenario is obtained, and the working condition parameters of the target production scenario are used as input conditions. The conditional temporal generator in the source domain pre-trained model is fine-tuned with a learning rate lower than that of the source domain pre-training model to obtain the target domain fine-tuned model as the pre-trained temporal generative adversarial network.

[0111] In this embodiment of the invention, firstly, through a data acquisition interface, historical operating status data, operating condition parameters, and digital parameters of multiple production scenarios that have undergone digital transformation are extracted from their information systems in a unified data format and sampling period.

[0112] Secondly, using the operating condition parameters of each production scenario as input, adversarial training is performed on all historical operating status data to complete source domain pre-training and obtain the source domain pre-trained model.

[0113] Specifically, historical operational status data from all acquired production scenarios of the same origin are used as training targets, while corresponding operating condition parameters and digitization parameters are used as input conditions. These are jointly input into a conditional time series generator and a time series discriminator for adversarial training. During training, the conditional time series generator attempts to fit the distribution pattern of the true residual sequences in the source domain, while the time series discriminator attempts to distinguish between the generated residual prediction sequences and the true residual sequences. The two work together through a minimax game for iterative optimization. After sufficient training, the network parameters of the conditional time series generator encode the common dynamic evolution patterns and residual patterns inherent in the source domain, forming a pre-trained model of the source domain.

[0114] Finally, historical operational status data of the target production scenario is acquired, and the working condition parameters of the target production scenario are used as input conditions. The conditional temporal generator in the source domain pre-trained model is fine-tuned with a learning rate lower than that of the source domain pre-training model, resulting in a fine-tuned target domain model as the pre-trained temporal generative adversarial network. Fine-tuning involves iteratively updating the model parameters using data from the target domain, based on the pre-trained model, to adapt the model to the specific data distribution of the target domain.

[0115] Specifically, it involves acquiring historical operational status data and corresponding operating condition parameters recorded during the actual production process of the target production scenario. During fine-tuning training, the network parameters of the time-series discriminator are kept frozen or only the output layer parameters of the time-series discriminator are updated.

[0116] Historical operational data and corresponding working condition parameters of the target domain are used as training samples. The parameters of the pre-trained model in the source domain are used as initial values, and training is performed using a learning rate lower than that used in the source domain pre-training stage. This lower learning rate could be, for example, one-tenth of the source domain pre-training learning rate. The parameters of the temporal discriminator can be frozen, and only the parameters of the conditional temporal generator can be updated to reduce the risk of overfitting and improve training efficiency. After fine-tuning, a fine-tuned model of the target domain is obtained, which is then used as a pre-trained temporal generative adversarial network for subsequent residual prediction of digital solutions for the target production scenario.

[0117] The method of this invention, which integrates temporal adversarial generative networks for simulating enterprise digital development paths, further includes: A comprehensive evaluation objective function is constructed, wherein the comprehensive evaluation objective function takes the average value across operating conditions, the volatility across operating conditions, and the maximum risk probability as independent variables, and performs a weighted summation with preset weight coefficients; Using the comprehensive evaluation objective function as the optimization objective and the feasible region of the digitized parameters as the constraint, candidate digitized parameters are searched in the feasible region through iterative optimization. Based on the candidate digitization parameters, candidate digitization schemes are constructed, deterministic benchmark prediction sequences and residual prediction sequences are re-acquired, and time-step superposition and statistical analysis are performed until the comprehensive evaluation objective function meets the convergence condition. The corresponding digitization parameters, digitization schemes and simulation results are then output.

[0118] In this embodiment of the invention, firstly, a comprehensive evaluation objective function is constructed, wherein the comprehensive evaluation objective function uses the average value across operating conditions, the volatility across operating conditions, and the maximum risk probability as independent variables, and is weighted and summed using preset weight coefficients.

[0119] Specifically, the cross-condition average is the arithmetic mean of the expected values ​​of the system state indicators under each test condition, used to characterize the average performance level of the digital solution under various conditions; the cross-condition volatility is the variance of the system state indicators under each test condition, used to characterize the stability of the digital solution's performance under different conditions; the maximum risk probability is the maximum probability that the observed value of each system state indicator is lower than the corresponding preset target threshold across all test conditions, used to characterize the risk tolerance of the digital solution under the most unfavorable conditions. The preset weighting coefficients are three weight values ​​pre-set according to the actual needs of the target production scenario, with the sum of the three equal to 1, used to adjust the relative importance of average performance, stability, and worst-case risk in the comprehensive evaluation.

[0120] Preferably, the preset weight coefficients can be pre-set according to the actual needs of the target production scenario, specifically determined through the analytic hierarchy process (AHP) or expert scoring. The three evaluation dimensions included in the comprehensive evaluation objective function are determined: average efficiency, stability, and worst-case risk. Experts in production management are invited to use a nine-level scaling method to compare the relative importance of each dimension pairwise, constructing a judgment matrix. Finally, the eigenvectors of the judgment matrix are calculated and normalized to obtain the weight coefficients for each dimension.

[0121] For example, in the high-end electronics manufacturing scenario where production capacity is the priority, experts judged that average efficiency is more important than stability and worst-case risk. The constructed judgment matrix yielded weight coefficients of 0.6, 0.25, and 0.15 for average efficiency, stability, and worst-case risk, respectively. In contrast, in the chemical production scenario where safe and stable operation is the primary goal, experts judged that stability and worst-case risk are more important. The corresponding weight coefficients were calculated to be 0.2, 0.4, and 0.4, respectively.

[0122] By weighted summation, the three different evaluation indicators mentioned above are combined into a standardized objective function value. The comprehensive evaluation objective function = (average value across operating conditions × preset weighting coefficient of average value across operating conditions) + (variability across operating conditions × preset weighting coefficient of volatility across operating conditions) + (1 - maximum risk probability × preset weighting coefficient of volatility across operating conditions). The higher this value, the better the overall effectiveness of the corresponding digital solution.

[0123] Secondly, with the comprehensive evaluation objective function as the optimization objective and the feasible region of the digitized parameters as the constraint, candidate digitized parameters are searched in the feasible region through iterative optimization.

[0124] Specifically, the digital parameters include sensor configuration parameters, edge computing configuration parameters, data acquisition configuration parameters, production management configuration parameters, and maintenance strategy configuration parameters. Each type of digital parameter has its preset value range, and the Cartesian product of all parameter value ranges constitutes the feasible region. With maximizing the comprehensive evaluation objective function value as the optimization direction, an iterative optimization algorithm is used to search for the optimal combination of digital parameters within the feasible region. The iterative optimization algorithm can be any one of Bayesian optimization, genetic algorithm, or particle swarm optimization. In each iteration, the algorithm constructs a surrogate model or updates the population based on the currently evaluated digital parameter combinations and their corresponding comprehensive evaluation objective function values, thereby determining the candidate digital parameter combinations to be evaluated in the next round. Through multiple iterations, the algorithm gradually approaches the global optimal solution within the feasible region, avoiding the computational explosion problem caused by exhaustive search, while ensuring optimization efficiency and solution quality.

[0125] Finally, based on the candidate digitization parameters, candidate digitization schemes are constructed, deterministic baseline prediction sequences and residual prediction sequences are re-acquired, and time-step stacking and statistical analysis are performed until the comprehensive evaluation objective function satisfies the convergence condition, and the corresponding digitization parameters, digitization schemes and simulation results are output.

[0126] Specifically, in the iterative optimization process, each round of selected candidate digitization parameters undergoes a re-execution of simulation analysis. This involves initializing a mechanistic simulation model based on the candidate digitization parameters, running the model to obtain a deterministic baseline prediction sequence, obtaining a residual prediction sequence through a conditional time series generator of a temporal generative adversarial network, superimposing these sequences to generate multiple simulation trajectories, performing statistical analysis, extracting evaluation indicators, and finally calculating the comprehensive evaluation objective function value. This process is iteratively executed until a preset convergence condition is met. The convergence condition is defined as the change in the comprehensive evaluation objective function value being lower than a preset change threshold for a consecutive preset number of rounds, or reaching a preset maximum number of iteration rounds. When the convergence condition is met, the candidate digitization parameters corresponding to the current round are taken as the optimal digitization parameters, and they are correlated with the corresponding digitization scheme and all simulation results, outputting this as the final output of the optimization method.

[0127] In this embodiment of the invention, the expected value and variance of the system state index are calculated, confidence intervals and individual evaluation indices are calculated, and a comprehensive evaluation index across operating conditions is extracted to achieve a comprehensive evaluation of single-operating-condition uncertainty quantification and multi-operating-condition robustness. After pre-training in the source domain, the conditional time series generator is fine-tuned in the target domain with a learning rate lower than that of the source domain pre-training. Finally, by constructing a comprehensive evaluation objective function with the average value across operating conditions, the volatility across operating conditions, and the maximum risk probability as independent variables and weighted by preset weight coefficients, candidate digital parameters are iteratively searched under the constraints of the feasible domain of digital parameters until the convergence condition is met and the optimal digital scheme is output, thus achieving automatic optimization of multiple objectives.

[0128] The embodiments of this application, through the above specific implementation methods, achieve the following technical effects: In this embodiment, firstly, the scope of operating conditions covered in the digital solution evaluation is expanded by using historical data from similar production scenarios. The similar operating conditions are then screened using a reproducibility rate metric to ensure the universality and representativeness of the introduced operating conditions, avoiding the waste of simulation computing resources caused by extreme and ineffective operating conditions. Finally, the size of the similar operating condition set is constrained by a preset ratio to provide a data foundation for executing subsequent steps.

[0129] Secondly, by constructing sub-models for equipment performance degradation, energy consumption prediction, failure rate prediction, data acquisition and communication, production scheduling, and maintenance strategy, and defining the mapping relationship between digital parameters and the operating parameters of each sub-model, the mechanism simulation model can comprehensively represent multi-dimensional production factors, enabling the simulation process to reflect the impact of digital solutions on each production link and improving the systematicness and credibility of the simulation results.

[0130] Secondly, based on historical operational data of the target production scenario, a stability metric for each test condition is calculated to quantitatively assess the inherent volatility of each condition. The sampling frequency for each condition is determined based on the stability metric, ensuring that conditions with lower stability metrics receive more sampling times. With a fixed total sampling budget, more computational resources are allocated to conditions with higher uncertainty, improving sampling efficiency and the accuracy of uncertainty quantification for key conditions. A residual prediction sequence is generated by using the test condition parameters and the digital parameters of the digital scheme as input conditions, combined with a random sampling noise vector input conditional time series generator. By repeatedly executing the generation process for each test condition, multiple residual prediction sequences are obtained, providing a sufficient data foundation for subsequent time-step overlay generation of multiple simulation trajectories.

[0131] Furthermore, by superimposing the deterministic benchmark prediction sequence and the residual prediction sequence step by step, multiple simulation trajectories that integrate the fusion mechanism and data-driven bias compensation are generated, so that the simulation results simultaneously contain physical consistency and random uncertainty, providing a diverse sample basis for the subsequent extraction of evaluation indicators based on trajectory statistical distribution.

[0132] Finally, the expected value and variance of the system state indicators are calculated, confidence intervals and individual evaluation indicators are calculated, and a comprehensive evaluation indicator across operating conditions is extracted to achieve a comprehensive evaluation of single-condition uncertainty quantification and multi-condition robustness. After pre-training in the source domain, the conditional time series generator is fine-tuned in the target domain with a learning rate lower than that of the source domain pre-training. Finally, by constructing a comprehensive evaluation objective function with cross-condition average value, cross-condition volatility and maximum risk probability as independent variables and weighted by preset weight coefficients, candidate digital parameters are iteratively searched under the constraints of the feasible domain of digital parameters until the convergence condition is met and the optimal digital solution is output, thus achieving automatic optimization of multiple objectives.

Claims

1. A simulation method for enterprise digital development path integrating temporal adversarial generative networks, characterized in that, include: Based on the target production scenario, construct a digital test case set containing multiple typical working conditions; Obtain the digital scheme to be evaluated, initialize the pre-built mechanism simulation model according to the digital scheme, and run the mechanism simulation model to perform simulation analysis on the digital test case set to obtain the deterministic benchmark prediction sequence under the corresponding test case. Using each test condition and the aforementioned digitization scheme as input conditions, a conditional time series generator of a pre-trained temporal generative adversarial network is used to generate the residual prediction sequence for the corresponding test condition, combined with a random noise sampling strategy. Using the test conditions as an index, the deterministic benchmark prediction sequence and each residual prediction sequence are superimposed step-by-step to generate multiple simulation trajectories. Statistical analysis is performed on the simulation trajectory based on all test conditions to obtain the statistical distribution and extract the corresponding evaluation indicators. The evaluation indicators are then associated with the digital scheme and output as simulation results.

2. The enterprise digital development path simulation method using a fusion temporal adversarial generative network as described in claim 1, characterized in that, Based on the target production scenario, a digital test case set containing various typical working conditions is constructed, including: Acquire historical operating data of the target production scenario, and extract and statistically analyze the operating condition parameters of the historical operating data to determine the original operating condition set; Acquire historical operational data from multiple production scenarios that share the same equipment type and similar processes as the target production scenario; Statistical analysis was performed on the reproducibility rate of each working condition in the historical operating data of the same production scenario. Multiple working conditions with the highest reproducibility were selected to determine the set of same working conditions. The original working condition set and the same working condition set are combined into a digital test working condition set, wherein the number of working conditions in the same working condition set is a preset ratio of the number of working conditions in the original working condition set.

3. The enterprise digital development path simulation method using a fusion temporal adversarial generative network as described in claim 1, characterized in that, The pre-construction of the mechanism simulation model includes: Construct sub-models for equipment performance degradation, energy consumption prediction, failure rate prediction, data acquisition and communication, production scheduling, and maintenance strategy. Define the mapping relationship between the digitization parameters and the running parameters of each sub-model, where: Sensor configuration parameters are mapped to the data acquisition and communication sub-model and the energy consumption prediction sub-model; Edge computing configuration parameters and data acquisition configuration parameters are mapped to the data acquisition and communication sub-model; Production management configuration parameters are mapped to the production scheduling sub-model; The maintenance strategy configuration parameters are mapped to the maintenance strategy sub-model and the failure rate prediction sub-model; Multiple sub-models are assembled according to the process logic to obtain the mechanism simulation model, wherein the output of each sub-model serves as the input condition for the downstream sub-model.

4. The enterprise digital development path simulation method using a fusion temporal adversarial generative network as described in claim 1, characterized in that, Obtain the digital scheme to be evaluated, initialize a pre-built mechanism simulation model based on the digital scheme, and run the mechanism simulation model to perform simulation analysis on the digital test case set to obtain the deterministic benchmark prediction sequence under the corresponding test case, including: Obtain the digitalization solution to be evaluated, and extract digitalization parameters from the digitalization solution, wherein the digitalization parameters include sensor configuration parameters, edge computing configuration parameters, data acquisition configuration parameters, production management configuration parameters, and maintenance strategy configuration parameters; Based on the digitized parameters, the operating parameters of each sub-model in the mechanism simulation model are initialized and set through a predefined parameter mapping relationship; The test conditions in the digital test condition set are traversed, and the condition parameters corresponding to each test condition are input into the initialized mechanism simulation model to perform simulation analysis. The system state prediction sequence is extracted from the simulation analysis results of each test condition, and the corresponding deterministic benchmark prediction sequence is constructed. The association marker between each deterministic benchmark prediction sequence and the corresponding test condition is established.

5. The enterprise digital development path simulation method using a fusion temporal adversarial generative network as described in claim 1, characterized in that, The pre-training of the temporal generative adversarial network includes: Acquire historical operating status data, operating condition parameters, and digital parameters of multiple digitally configured production scenarios from the same source, and construct a training sample set; The historical operating state data in the training sample set is subtracted step by step from the deterministic prediction sequence output by the mechanism simulation model under the corresponding operating condition parameters and digital parameters to obtain the residual sequence. A conditional time series generator is constructed, wherein the conditional time series generator takes a random sampled noise vector as random input and operating condition parameters and digital parameters as conditional input, and generates residual prediction sequences step by step in an autoregressive manner based on a recurrent neural network; Construct a time series discriminant, which is used to distinguish between the real residual sequence and the predicted residual sequence; Using the residual sequence as the training target, the network parameters of the conditional time series generator and the time series discriminator are optimized through adversarial training. The loss function of the conditional time series generator consists of adversarial loss and physical constraint regularization loss.

6. The enterprise digital development path simulation method using a fusion temporal adversarial generative network as described in claim 1, characterized in that, Using each test condition and the aforementioned digitization scheme as input conditions, a pre-trained temporal generative adversarial network (GAN) conditional temporal generator, combined with a random noise sampling strategy, generates the residual prediction sequence for the corresponding test condition, including: Based on the historical operating status data of the target production scenario, calculate the operating condition stability metric corresponding to each test condition, wherein the operating condition stability metric is used to characterize the degree of fluctuation of the test condition in the historical operating status data. Based on the stability metric of each test condition, the number of samplings corresponding to each test condition is determined, wherein the lower the stability metric of the test condition, the greater the number of samplings. For each test condition, the condition parameters of the test condition and the digital parameters extracted from the digital scheme are used as conditional inputs, and the random sampling noise vector is used as random input. The inputs are then fed into the conditional time series generator of the time series generative adversarial network to obtain the corresponding single residual prediction sequence. Based on the number of samplings determined for each test condition, the generation process is repeated to obtain multiple residual prediction sequences for each test condition.

7. The enterprise digital development path simulation method using a fusion temporal adversarial generative network as described in claim 1, characterized in that, Statistical analysis is performed on the simulation trajectories based on all test conditions to obtain statistical distributions and extract corresponding evaluation indicators. These evaluation indicators are then correlated with the digital scheme and output as simulation results, including: For each test condition, the expected value and variance of the system state index are calculated based on the corresponding multiple simulation trajectories. For each test condition, based on the corresponding multiple simulation trajectories, the confidence interval of each system state index under the preset confidence level and the risk probability that the observed value of each system state index is lower than the corresponding preset target threshold are calculated, and the output is a single evaluation index. Across all test conditions, the average value of each system status indicator across all test conditions, the fluctuation of each system status indicator across all test conditions, and the maximum risk probability of each system status indicator across all test conditions are analyzed and obtained as comprehensive evaluation indicators across all test conditions. The individual evaluation indicators and the comprehensive evaluation indicators across test conditions are associated and labeled with the digital scheme, and the results are output as simulation results.

8. The enterprise digital development path simulation method using a fusion temporal adversarial generative network as described in claim 1, characterized in that, The pre-training of the temporal generative adversarial network also includes: Acquire historical operational status data, operating condition parameters, and digital parameters from multiple production scenarios with the same origin; Using the operating condition parameters of each production scenario as input, adversarial training is performed on all historical operating status data to complete source domain pre-training and obtain the source domain pre-trained model. Historical operational status data of the target production scenario is obtained, and the working condition parameters of the target production scenario are used as input conditions. The conditional temporal generator in the source domain pre-trained model is fine-tuned with a learning rate lower than that of the source domain pre-training model to obtain the target domain fine-tuned model as the pre-trained temporal generative adversarial network.

9. The enterprise digital development path simulation method using a fusion temporal adversarial generative network as described in claim 7, characterized in that, Also includes: A comprehensive evaluation objective function is constructed, wherein the comprehensive evaluation objective function takes the average value across operating conditions, the volatility across operating conditions, and the maximum risk probability as independent variables, and performs a weighted summation with preset weight coefficients; Using the comprehensive evaluation objective function as the optimization objective and the feasible region of the digitized parameters as the constraint, candidate digitized parameters are searched in the feasible region through iterative optimization. Based on the candidate digitization parameters, candidate digitization schemes are constructed, deterministic benchmark prediction sequences and residual prediction sequences are re-acquired, and time-step superposition and statistical analysis are performed until the comprehensive evaluation objective function meets the convergence condition. The corresponding digitization parameters, digitization schemes and simulation results are then output.