Digital twin driven smart factory ai intelligent decision system

Abstract

The application relates to the technical field of intelligent factory decision-making, and discloses a digital-twin-driven intelligent factory AI intelligent decision-making system, which comprises a data acquisition and preprocessing module, a twin modeling and identification module, an alignment evaluation and credibility module, a strategy simulation and optimization module, a gray online release module and a running evaluation and audit module.The data acquisition and preprocessing module is used for acquiring real-time and historical data of a production line and performing preprocessing; the twin modeling and identification module is used for constructing a twin body and performing parameter identification and uncertainty quantification; the alignment evaluation and credibility module is used for comparing twin body prediction data with real data, generating availability labels and trust scores; the strategy simulation and optimization module is used for generating candidate strategies and risk evidence based on constraints and key performance indicators; the gray online release module is used for screening and releasing strategies in gray according to the trust scores and performance results; the running evaluation and audit module is used for collecting running data and generating an audit package; and the drift detection and update module is used for detecting distribution drift and recalibrating and updating the twin body and the strategy library.The application realizes the reliability, robustness and continuous optimization of intelligent factory production decision-making.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent factory decision-making, and particularly relates to an intelligent factory AI intelligent decision-making system driven by digital twinning. BACKGROUND

[0002] With the development of digitalization and intelligentization of manufacturing industry, intelligent factory gradually becomes an important way to improve production efficiency, flexibility and resource utilization. In traditional production management, although production line operation data can be collected through sensors and information systems, there are often problems of data missing, noise interference and difficulty in timely removing abnormal points, resulting in lack of unified and reliable input for subsequent modeling and analysis. At the same time, the data formats and statistical indicators between different processes and links are inconsistent, which also limits the comprehensive optimization of the overall production line.

[0003] The existing decision-making methods mostly rely on static models or single prediction results, and it is usually difficult to reflect the advantages of AI intelligent decision-making, and the dynamic evolution and uncertainty of the production process cannot be described. For example, when the order demand fluctuates, the device state is abnormal or external environmental disturbance occurs, the existing method often cannot adjust the parameters in time, resulting in deviation of the strategy from the actual demand. Although digital twinning technology has been applied to device health monitoring, energy consumption prediction and other scenes in recent years, most of them are limited to local links and cannot fully combine the autonomous learning and autonomous decision-making ability of AI, lacking end-to-end modeling, optimization and feedback loop. In addition, the existing production line optimization means also has shortcomings in the execution level, and often directly applies the offline optimization results to the real production line, lacking the gray release and compliance verification mechanism based on AI evaluation. Once there is deviation between the online strategy and the actual production line environment, it may cause production interruption or efficiency decline, lacking effective rollback means and risk control ability. SUMMARY

[0004] The application provides an intelligent factory AI intelligent decision-making system driven by digital twinning, which solves the technical problems of inconsistent production line operation data quality, lack of uncertainty description in model prediction, and lack of gray verification in strategy online in related technologies.

[0005] The application provides an intelligent factory AI intelligent decision-making system driven by digital twinning, which includes:

[0006] A data acquisition and preprocessing module is configured to acquire real-time data and historical statistical data of the production line, standardize, suppress noise and detect abnormalities of the real-time data of the production line, and obtain a production line state vector, real observation data, noise parameters and abnormal markers;

[0007] A twinning modeling and identification module is configured to construct a twin body based on the production line state vector and the real observation data, and perform parameter identification and uncertainty quantification, and output twin body prediction data;

[0008] An alignment evaluation trusted module is configured to align and evaluate the twin prediction data and the real observation data, and obtain an availability label and a trust score according to a preset alignment threshold and a preset trust threshold;

[0009] A strategy simulation optimization module is configured to, when the availability label is available, solve a candidate strategy set in the counterfactual scenario set according to the key performance indicators and constraint conditions, and give a key performance indicator prediction result and risk evidence; wherein the constraint conditions include: physical constraints, capacity constraints and risk constraints.

[0010] A gray release module is configured to apply a preset online threshold to the candidate strategy set according to the trust score and the key performance indicator prediction result, and issue the candidate strategy set to generate an online strategy version in a gray release range.

[0011] A running evaluation audit module is configured to collect online strategy running data corresponding to the online strategy version, form a contrast evaluation, and generate an audit package.

[0012] A drift detection and update module is configured to detect drift based on a feature distribution stability indicator and an online error threshold, and recalibrate and update the twin and the strategy library.

[0013] Further, the process of standardizing, noise suppressing and anomaly detecting the production line real-time data includes:

[0014] Step 11, time alignment and integrity check are performed on the production line real-time data, wherein the time alignment generates a unified timestamp using a preset resampling period, and the integrity check removes continuous missing data segments according to the upper limit of the missing rate in the historical statistical data to obtain real observation data.

[0015] Step 12, the real observation data is standardized based on the mean and standard deviation in the historical statistical data, and a standby standard deviation is used to replace the standard deviation when the standard deviation is zero, to obtain a standardized sequence.

[0016] Step 13, Kalman filtering is applied to the standardized sequence channel by channel to generate a noise reduction sequence, and the residual error is used to estimate the observation noise variance and the process noise variance to form noise parameters.

[0017] Step 14, according to the noise parameters, determine the double-sided threshold, and perform preliminary threshold judgment on the noise reduction sequence, use the graph neural network model to analyze the multi-channel correlation of the noise reduction sequence, identify the abnormal points across channels and generate an abnormal label, and after compensation processing of the time stamp covered by the abnormal label using the historical mean, aggregate it into a production line state vector.

[0018] Further, the construction of the twin and the generation of the twin prediction data include:

[0019] Step 21, construct a twin taking the in-process state vector as input and taking the predicted value of each channel as output, the twin including mechanism terms and data terms, and set the initial value of the model parameter according to historical statistical data;

[0020] Step 22, remove the corresponding sample using the abnormality label, determine the sample weight based on the observation noise variance and the process noise variance in the noise parameter, and combine the remaining samples to form a weighted training set;

[0021] Step 23, establish a target function of the prediction error sum of squares based on the weighted training set, update the twin parameters using an iterative optimization algorithm, stop iteration when the target function decreases by less than a first preset threshold, and obtain an identified twin;

[0022] Step 24, use the identified twin to infer the in-process state vector, output the twin prediction data, and synthesize the prediction variance based on the model parameter covariance and the noise parameter to obtain the uncertainty quantification result.

[0023] Further, the generation process of the availability label and the trust score includes:

[0024] Step 31, remove abnormal points based on the abnormality label, calculate the error between the twin prediction data and the true observation data for each channel for the remaining data, and obtain an alignment error sequence;

[0025] Step 32, generate a sample weight based on the sum of the observation noise variance and the process noise variance in the noise parameter, and square and normalize the alignment error sequence according to the sample weight to obtain an overall error indicator as a weighted root mean square error;

[0026] Step 33, compare the overall error indicator with a second preset threshold set according to historical statistical data, and compare the uncertainty quantification result with a third preset threshold, and generate an availability label as available when both are satisfied, otherwise generate an unavailable label;

[0027] Step 34, set a normalization constant according to historical statistical data, map the overall error indicator to a score through an exponential monotone decreasing function, and multiply the score by the availability label to obtain a trust score.

[0028] Further, the key performance indicators include: throughput, work-in-process, energy consumption, and overdue rate.

[0029] Further, the strategy simulation optimization module specifically includes:

[0030] Step 41, construct a counterfactual scenario set according to the in-process state vector and a preset disturbance, and weight and normalize the scenario probability using the trust score to form a weighted scenario set;

[0031] Step 42, taking the weighted sum of key performance indicators as the optimization target, and integrating physical constraints, capacity constraints, and risk constraints defined by a preset risk threshold into the constraint conditions;

[0032] Step 43, calling the twin to simulate the control parameter sequence under the weighted scenario set, eliminating solutions that do not meet the constraint conditions or are inferior to other solutions in all key performance indicators, and selecting a plurality of candidate strategies with high scores to form a candidate strategy set;

[0033] Step 44, outputting key performance indicator prediction results and risk evidence for the candidate strategy, and packaging the candidate strategy identifier, control parameter sequence, evaluation time window, key performance indicator prediction result, risk evidence, and twin version number into a set, wherein the risk evidence includes the most adverse scenario result and risk threshold determination.

[0034] Further, the step 43 further comprises: obtaining a prediction variance by synthesizing model parameter covariance and noise parameters, tightening physical constraints and capacity constraints at a preset confidence level, so that they are converted into deterministic constraints at the preset confidence level; and adaptively adjusting the length and precision of the control parameter sequence using the prediction variance; obtaining candidate strategies according to the tightened constraint conditions and the adaptively set control parameter sequence and performing screening.

[0035] Further, the gray release module specifically comprises:

[0036] Step 51, setting a trust score threshold, a key performance indicator boundary, and a preset risk threshold based on historical statistical data, comparing the candidate strategy set one by one, and if the threshold set is met, including it in the online candidate; if the screening result is empty, relaxing a single threshold according to a preset de-escalation strategy, and only when the key performance indicator improvement of the candidate strategy relative to the baseline strategy reaches a preset improvement amplitude, determining it as an effective candidate;

[0037] Step 52, setting a gray ratio based on the trust score and key performance indicator margin of the effective candidate strategy through a monotonically increasing function, setting a gray time domain window based on the risk level through a monotonically decreasing function, and preferentially selecting units that have less impact on bottleneck processes within the coverage range;

[0038] Step 53, generating an online strategy version for the gray release candidate strategy, and recording the rollback trigger condition and the baseline strategy binding relationship, and switching to the baseline strategy when triggered.

[0039] Further, collect online strategy running data corresponding to the online strategy version, form a comparative evaluation, and generate an audit package, including:

[0040] Step 61, collect running data within the coverage and validity period window of the online policy version, and resample and align according to the data snapshot identifier, fill in the missing data and remove the abnormal points to form a consistent running data set;

[0041] Step 62, construct a control set based on the running data, preferentially use the true control of the unit without intervention, supplement the insufficient part with the corresponding twin version, and form the counterfactual key performance indicator by weighted fusion;

[0042] Step 63, calculate the observed key performance indicator using the unified caliber, generate the improvement amount by comparing with the counterfactual key performance indicator, multiply the improvement amount of each key performance indicator by the preset business weight one by one, sum to get the comprehensive effect score, and output the confidence interval at the significance level;

[0043] Step 64, within the validity period window, calculate the proportion of time when the actual execution value falls within the tolerance range of the preset control parameter sequence to obtain the compliance rate, and when the compliance rate is lower than the preset compliance rate threshold, mark it as low compliance, and detect the distribution distance of the key exogenous variable to generate the drift flag;

[0044] Step 65, encapsulate the policy identifier, control parameter sequence, twin version number, data snapshot identifier, validity period window, running data result, control result, effect score, compliance rate and drift flag field to generate an audit package.

[0045] Further, based on the feature distribution stability index and the online error threshold to detect drift, recalibrate and update the twin and policy library, including:

[0046] Step 71, extract the feature distribution of the running data within the sliding time window, and align with the baseline snapshot, and calculate the stability index using the binning method;

[0047] Step 72, double compare the stability index with the preset upper threshold and the online error threshold of the key performance indicator, trigger when multiple windows exceed the threshold continuously, and generate a drift determination flag;

[0048] Step 73, when the drift determination triggers, select samples that meet the compliance rate requirement and have no abnormal marks, generate weights according to the inverse of noise variance, and update the twin mechanism item and data item parameters using weighted regularized least squares iteration, stop when the parameter change is less than the preset change threshold, and get the updated twin parameters;

[0049] Step 74, re-simulate the policy library based on the updated twin in the weighted scenario, retain the candidate policies that meet the preset risk threshold and improve the baseline policy by the fourth preset threshold, and remove the schemes that are dominated on all key performance indicators to generate an updated policy library.

[0050] The beneficial effects of the present application are: the present application realizes accurate simulation and dynamic updating of the operation state of the production line by introducing digital twin modeling, parameter identification and uncertainty quantification, ensuring the reliability of the prediction data; through the availability mark and trust score generation method based on error alignment and threshold determination, the credibility of the twin can be quantified and fed back to the strategy optimization link; by combining the counter-factual scenario set to carry out strategy simulation and screening under multiple constraints, both physical and production capacity constraints are considered, and risk constraints are also included, making the candidate strategy more robust; through the trust score and key performance indicator driven online threshold control, as well as the dynamic setting of the gray scale ratio and gray time domain window, the gradual landing of the strategy in the real production line environment is ensured, and the online risk is reduced; through the drift detection based on stability indicators and online errors, the twin and the strategy library are recalibrated and updated, ensuring the long-term adaptability and continuous optimization capability of the system. Overall, the reliability, stability and continuous optimization capability of the intelligent factory production decision are ensured. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 is a module schematic diagram of the digital twin driven intelligent factory AI intelligent decision system of the present application. DETAILED DESCRIPTION

[0052] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can be changed in function and arrangement without departing from the scope of the present specification. Various processes or components can be omitted, substituted, or added according to desired implementations. Additionally, features described with respect to some examples can be combined in other examples.

[0053] As shown in Figure 1 The digital twin driven intelligent factory AI intelligent decision system includes:

[0054] The data acquisition and preprocessing module 1 is used to obtain real-time data and historical statistical data of the production line, standardize, suppress noise and detect abnormalities of the real-time data of the production line, and obtain a production line state vector, real observation data, noise parameters and abnormal markers;

[0055] The twin modeling and identification module 2 is used to construct a twin based on the production line state vector and real observation data, and perform parameter identification and uncertainty quantification, and output twin prediction data;

[0056] The alignment evaluation and credibility module 3 is used to align and evaluate the twin prediction data and the real observation data, and obtain an availability mark and a trust score according to a preset alignment threshold and a preset trust threshold;

[0057] The strategy simulation optimization module 4 is configured to solve a candidate strategy set in the counterfactual scenario set according to the key performance indicators and constraint conditions when the availability is marked as available, and give a key performance indicator prediction result and risk evidence; wherein the constraint conditions include: physical constraints, capacity constraints and risk constraints.

[0058] The gray release module 5 is configured to apply a preset online threshold to the candidate strategy set according to the trust score and the key performance indicator prediction result, and issue the candidate strategy set in the gray release range to generate an online strategy version.

[0059] The running evaluation audit module 6 is configured to collect online strategy running data corresponding to the online strategy version, form a contrast evaluation and generate an audit package.

[0060] The drift detection update module 7 is configured to detect drift based on a feature distribution stability indicator and an online error threshold, and recalibrate and update the twin and the strategy library.

[0061] In an embodiment of the present application, the production line real-time data refers to the running data such as process parameters, equipment state quantities and material information continuously collected by sensors or information systems in the production line during the production process, which reflects the instant running situation of the production line in the form of time series. The historical statistical data refers to the results obtained by archiving and statistically analyzing the production line running data in a long time period, including the mean, standard deviation, missing rate and distribution characteristics of each channel, which are used as a reference for data processing and model initialization.

[0062] The process of standardizing, noise suppressing and anomaly detecting the production line real-time data includes:

[0063] Step 11, time alignment and integrity check are performed on the production line real-time data, wherein the time alignment generates a unified timestamp by using a preset resampling period, so as to ensure the correspondence between the data of different channels in the time dimension, and the integrity check removes the continuously missing data segment according to the upper limit of the missing rate in the historical statistical data, so as to avoid the data quality not meeting the requirements and obtain real observation data;

[0064] Step 12, the real observation data is standardized based on the mean and standard deviation in the historical statistical data, and a standby standard deviation is used to replace the standard deviation when the standard deviation is zero, to obtain a standardized sequence; specifically, the standardization refers to subtracting the historical mean from the original data and dividing by the historical standard deviation, so that the data sequence has consistent dimension and scale between different channels;

[0065] Step 13, Kalman filtering is applied to the standardized sequence channel by channel to generate a noise reduction sequence, and the residual error is used to estimate the observation noise variance and process noise variance to form noise parameters;

[0066] Step 14, determine the double-sided threshold according to the noise parameter, and perform preliminary threshold judgment on the noise reduction sequence to quickly identify data points that exceed the reasonable range; the double-sided threshold refers to the limits set in the upper and lower directions of the data distribution; on this basis, further multi-channel correlation analysis of the noise reduction sequence is performed by using a graph neural network model, so as to identify abnormal points across channels and generate abnormal markers.

[0067] Specifically, the graph neural network model is constructed based on graph structure data, the nodes in the graph structure data represent the monitoring data of different channels in the production line, and the edges in the graph structure data represent the logical or physical connection between the channels, such as the sequence of processes, the material transfer relationship between devices, etc. By iteratively transmitting information on the graph structure, the graph neural network model can automatically capture high-order dependency relationships between different channels, thereby effectively identifying coupled abnormalities that are difficult to find by single-channel threshold methods. If the features of a node show obvious inconsistency with the neighbor nodes, the timestamp data corresponding to the node is marked as an abnormal point; the system generates an abnormal marker for the corresponding timestamp, and compensates the abnormal data using the historical mean value to ensure data continuity and stability; this process embodies the intelligent perception and abnormal cognition ability of AI, enabling the system to not only handle local abnormalities, but also maintain data quality under complex process coupling conditions.

[0068] Through the above process, data objects that meet the modeling requirements can be extracted from the original real-time production line data, i.e., production line state vectors, real observation data, noise parameters, and abnormal markers. Among them, the production line state vector is used to show the running state of the production line under a unified time scale, the real observation data provides the original input corresponding to the actual working condition, the noise parameter provides the basis for subsequent uncertainty modeling and weight allocation, and the abnormal marker is used to ensure that abnormal points do not interfere with the modeling accuracy.

[0069] In an embodiment of the present application, the construction of the twin and the generation of the twin prediction data include:

[0070] Step 21, construct a twin with the production line state vector as input and the prediction value of each channel as output, the twin includes a mechanism term and a data term, and the initial value of the model parameter is set according to historical statistical data; specifically, the twin refers to a hybrid model based on the real production line running mechanism, which combines mechanism modeling and data-driven modeling, wherein the mechanism term is a function relationship established based on the known physical laws and process constraints of the production line, including but not limited to material balance model, energy consumption model, dynamic transfer function model, used to describe the causal relationship of physical processes, and the data term is used to fit nonlinear or disturbance effects that are not fully covered by the mechanism;

[0071] Step 22, when training the twin, the corresponding samples are removed by using the anomaly label to avoid the interference of noise and anomaly on the training accuracy, and the sample weight is determined based on the observation noise variance and the process noise variance in the noise parameter, and the reserved samples are combined to form a weighted training set;

[0072] Step 23, a target function of prediction error square sum is established based on the weighted training set, an iterative optimization algorithm is used to update the twin parameters, and the iteration is stopped when the descending amplitude of the target function is lower than a first preset threshold, and an identified twin is obtained; specifically, the calculation formula of the target function is: , , represents the value of the target function, T represents the time index number, t represents the time index, C represents the total number of channels, c represents the channel index, , represents the sample weight of channel c at time t, and the weight can be determined by the reciprocal of the observation noise variance and the process noise variance, , represents the real observation data, , represents the line state vector at time t, , represents the prediction value of channel c under the twin parameters , the target function is used to measure the difference between the twin prediction value and the real observation data;

[0073] Step 24, the line state vector is inferred by using the identified twin, the twin prediction data is output, the prediction variance is synthesized based on the model parameter covariance and the noise parameter, and the uncertainty quantization result is obtained. Wherein, the model parameter covariance refers to the covariance matrix of parameter estimation in the iterative optimization process, which is used to describe the uncertainty of the model parameters; the uncertainty quantization result refers to the variance or interval estimation given on the basis of the prediction output, which is used to reflect the confidence range of the twin prediction result.

[0074] Through the above steps, the twin reflecting the actual running state of the production line can be constructed under the premise of ensuring the data quality, and the stability and reliability of the twin are improved through weighted identification and uncertainty quantization; the twin prediction data obtained in this way not only can provide accurate numerical prediction for subsequent decision-making, but also can provide a measure of prediction reliability, avoiding the complete dependence of the decision-making process on single-point prediction values; embodies the learning and modeling ability of AI, can identify the law from historical and real-time data, dynamically adapt to unknown disturbances, and thus provide more intelligent and adaptive input for subsequent decision-making.

[0075] In an embodiment of the present application, the generation process of the availability label and the trust score includes:

[0076] Step 31, based on the abnormal marker, the abnormal points are removed, the errors between the twin prediction data and the real observation data are calculated for the remaining data channel by channel, and an alignment error sequence is obtained, which reflects the prediction deviation of the twin in the current period;

[0077] Step 32, based on the sum of the observation noise variance and the process noise variance in the noise parameter, a sample weight is generated, and the alignment error sequence is squared and normalized according to the sample weight to obtain an overall error indicator as a weighted root mean square error, which considers the error size and noise level, thereby avoiding the dominance of the channel with high noise in the overall result;

[0078] Step 33, compare the overall error indicator with the second preset threshold set according to historical statistical data, and compare the uncertainty quantification result with the third preset threshold, when both meet, generate an availability marker as available, otherwise generate unavailable, and represent them with 1 and 0 respectively;

[0079] Step 34, set a normalization constant according to historical statistical data, map the overall error indicator to a score through an exponential monotone decreasing function, and multiply it by the availability marker to obtain a trust score; specifically, the calculation formula of the trust score is: S represents the trust score, represents the availability marker, exp represents the exponential function, represents the overall error indicator, represents the normalization constant.

[0080] Through the above steps, the embodiment can not only generate the availability marker of the twin based on the dual standards of error and uncertainty, but also further generate the trust score as a measure; on the one hand, it can eliminate unreliable twin prediction to avoid its entering the subsequent decision; on the other hand, it can establish ranking and priority among multiple candidate schemes through the trust score, enhancing the decision stability and interpretability of the system in the intelligent factory scenario; the above process is equivalent to AI self-cognition, which does not blindly rely on model output, but makes a judgment on whether to adopt it through the availability marker and the trust score.

[0081] In an embodiment of the present application, the key performance indicators include: throughput, work-in-process quantity, energy consumption and overdue rate. Among them, the throughput refers to the number of qualified products completed by the production line within a preset time window; the work-in-process quantity refers to the number of semi-finished products or materials being processed in each process of the production line; the energy consumption refers to the total energy consumption of the production line within a preset time window; and the overdue rate refers to the proportion of the number of production orders that cannot be completed on time within a preset time window to the total number of orders.

[0082] In an embodiment of the present application, the strategy simulation optimization module specifically includes:

[0083] Step 41, according to the production line state vector and the preset disturbance, a counterfactual scenario set is constructed, and the scenario probability is weighted and normalized by using the trust score to form a weighted scenario set; specifically, the preset disturbance refers to order demand fluctuation, equipment failure simulation, or external supply delay; the weighted scenario probability is used to measure the importance of different scenarios;

[0084] Step 42, taking the weighted sum of key performance indicators as the optimization target, wherein the weights are set according to historical statistical data, and physical constraints, capacity constraints, and risk constraints defined by a preset risk threshold are uniformly included in the constraint conditions; specifically, the physical constraints define the safety limits of the equipment, and the capacity constraints define the maximum load of the process;

[0085] Step 43, under the weighted scenario set, the twin body is called to simulate the control parameter sequence, the schemes that do not meet the constraint conditions or are inferior to other solutions in all key performance indicators are removed, and a plurality of candidate strategies with high scores are selected to form a candidate strategy set; the control parameter sequence refers to the timing setting of adjustable variables in the production line operation process within a preset time window, including production rate, equipment switching, resource allocation, and scheduling window, etc., and the control parameter sequence is the core input of the candidate strategy set for twin body simulation and evaluation;

[0086] Step 44, the key performance indicator prediction results and risk evidence are output for the candidate strategies, and the candidate strategy identifier, control parameter sequence, evaluation time window, key performance indicator prediction results, risk evidence, and twin body version number are packaged into a set; the risk evidence refers to the results extracted in the evaluation process of the candidate strategy for proving the robustness of the strategy, including the most adverse scenario result and risk threshold determination.

[0087] Through the above steps, the embodiment not only can generate candidate strategies that meet physical and capacity constraints under multiple counterfactual scenarios, but also can ensure the robustness and explainability of the candidate strategies through trust score weighting and risk evidence output; further improving the reliability of the strategy optimization process driven by digital twin, so that the generated strategy not only has high expected return, but also has strong anti-risk ability in uncertain environment.

[0088] In an embodiment of the present application, the step 43 further comprises: obtaining a prediction variance by synthesizing the model parameter covariance and the noise parameter, tightening the physical constraints and the capacity constraints at a preset confidence level to convert them into deterministic constraints at the preset confidence level, and adaptively adjusting the length and precision of the control parameter sequence according to the prediction variance; and obtaining a candidate strategy according to the tightened constraints and the adaptively set control parameter sequence and performing screening. Specifically, the prediction variance is obtained by synthesis, which can quantify the uncertainty level of the prediction result; the constraint boundary is tightened at the preset confidence level, so that the constraint can still be met under high probability conditions; and when the prediction variance is large, the length of the control parameter sequence is shortened and the step interval is increased to reduce the uncertainty risk of long-term prediction; and when the prediction variance is small, the control length is extended and the step interval is reduced to fully exploit the optimization space.

[0089] In an embodiment of the present application, the gray release module specifically comprises:

[0090] Step 51, according to historical statistical data, set a trust score threshold, a key performance indicator boundary and a preset risk threshold, compare the candidate strategy set one by one, if the threshold set is met, it is included in the online candidate; if the screening result is empty, relax the single threshold according to the preset de-escalation strategy, and only when the key performance indicator improvement amplitude of the candidate strategy relative to the baseline strategy reaches the preset improvement amplitude, it is determined as an effective candidate; this way can avoid the strategy from being unable to go online due to too strict screening under the premise of ensuring safety and stability; wherein the preset de-escalation strategy includes but is not limited to: appropriately relaxing the throughput improvement amplitude or the energy consumption boundary, but keeping other indicators and risk thresholds unchanged;

[0091] Step 52, set the gray ratio based on the trust score and the key performance indicator margin of the effective candidate strategy through a monotonically increasing function, that is, initially only cover a part of the production units during the system online process, the higher the trust score and the indicator margin, the larger the gray ratio; set the gray time domain window based on the risk level through a monotonically decreasing function, that is, gradually release the time length within the online period, the higher the risk level, the shorter the time domain window; and preferentially select units that have less impact on the bottleneck process within the coverage range to reduce the impact of gray release on the overall stability of the production line; the gray ratio refers to the proportion of the candidate strategy covering the production line units in the early stage of online; the gray time domain window refers to the execution length allowed by the gray strategy during online;

[0092] Step 53, generate an online strategy version for the gray release candidate strategy, the minimum fields encapsulated include strategy identification, control parameter sequence, key performance indicator prediction result, risk evidence, trust score, gray ratio, gray time domain window, coverage range, twin version number, data snapshot identification and valid period window, and record the rollback trigger condition and the baseline strategy binding relationship, and switch to the baseline strategy when triggered.

[0093] Through the above steps, the embodiment can convert the candidate strategy into an online strategy version with an online execution condition; through multi-dimensional threshold screening and degrading strategy mechanism, the usability and improvement of the candidate strategy are ensured; through dynamic adjustment of the gray scale ratio and the gray time domain window, the online risk is reduced and the strategy is gradually expanded; through the rollback mechanism and the baseline binding, it is ensured that the system can quickly recover to a safe running state in abnormal conditions; it embodies the risk perception and adaptive online decision-making ability of AI, and ensures the stable landing of the strategy in the smart factory environment; and provides reliable execution guarantee for digital twin AI intelligent decision-making in the smart factory environment.

[0094] In an embodiment of the present application, the online strategy version corresponding online strategy running data is collected, a contrast evaluation is formed, and an audit package is generated, including:

[0095] Step 61, collecting running data within the coverage range and the effective period window of the online strategy version, and resampling and aligning according to the data snapshot identifier, using missing data completion and outlier rejection to form a consistent running data set; the coverage range refers to the actual application production unit set when the strategy is released in gray scale, and the effective period window refers to the effective time interval set for the strategy version;

[0096] Step 62, constructing a contrast set based on the running data, preferentially using the real contrast of the non-strategy unit, supplementing the insufficient part by the corresponding twin version, and weighted fusion to form counterfactual key performance indicators; wherein the non-strategy unit refers to the equipment or time slice in the same production environment that has not applied the candidate strategy, which is used as a real contrast; the counterfactual key performance indicator refers to the expected performance of the production line if the candidate strategy is not used in the current production environment, which is used for differential comparison with the actual observation result;

[0097] Step 63, calculating the observed key performance indicators using a unified scale, comparing them with the counterfactual key performance indicators to generate improvement amounts, multiplying each key performance indicator improvement amount by the preset business weight one by one, summing to obtain a comprehensive effect score, and outputting a confidence interval at a significance level; preferably, the significance level is 0.05; the comprehensive effect score is used to show the overall effect of the candidate strategy;

[0098] Step 64, within the effective period window, calculate the proportion of the period when the actual execution value falls within the tolerance range of the preset control parameter sequence to obtain the compliance rate, and when the compliance rate is lower than the preset compliance rate threshold, mark it as low compliance, and detect the distribution distance of the key exogenous variables to generate a drift flag; the key exogenous variable values include order demand, supply delay, and environment temperature; specifically, the difference between the distribution of the key exogenous variable and the historical distribution is detected, the distribution distance is calculated, and if the distance exceeds the corresponding threshold, a drift flag is generated to prompt that the external conditions significantly deviate from the training or evaluation environment;

[0099] Step 65, encapsulate the policy identification, control parameter sequence, twin version number, data snapshot identification, validity period window, running data result, control result, effect score, compliance rate and drift flag field to generate an audit package, and write fingerprint information to support reproduction and audit.

[0100] Through the above steps, the embodiment realizes the full-link closed loop from running data collection, control construction, effect evaluation, compliance and drift detection to audit encapsulation; the process not only provides quantitative evidence for the effect of the candidate strategy, but also provides guarantee for subsequent reproduction and external audit through the generation and fingerprint storage of the audit package, thereby enhancing the transparency and controllability of the system in the intelligent factory environment.

[0101] In an embodiment of the present application, based on the feature distribution stability index and the online error threshold value to detect drift, the twin and the strategy library are recalibrated and updated, including:

[0102] Step 71, extract the feature distribution of the running data in the sliding time window, and align it with the baseline snapshot, and calculate the stability index by using the binning method; specifically, the baseline snapshot refers to the running data distribution saved at the time of the last update of the twin, which is used as a comparison benchmark; to calculate the feature distribution stability index, the feature variable is divided into multiple intervals, and the difference between the current distribution and the baseline distribution is compared to obtain the stability index value; the index is used to quantify the statistical consistency between the running data and the historical benchmark;

[0103] Step 72, compare the stability index with the preset upper threshold value and the online error threshold value of the key performance indicator, and trigger when the threshold is exceeded for a plurality of windows in succession, and generate a drift determination flag; the drift determination flag is a binary signal, which is used to indicate whether the prediction performance of the twin has decreased significantly or the input feature distribution has deviated significantly, and recalibration is needed;

[0104] Step 73, when the drift determination is triggered, select samples that meet the compliance rate requirement and have no abnormal marks, generate weights according to the inverse of the noise variance, that is, the smaller the noise variance, the higher the weight of the sample in training, update the mechanism term and data term parameters of the twin by using weighted regularized least squares iteration, stop when the parameter change is less than the preset change threshold, obtain the updated twin parameters and record the version number;

[0105] Step 74, re-simulate the strategy library based on the updated twin in the weighted scenario, retain the candidate strategies that meet the preset risk threshold and improve the baseline strategy by the fourth preset threshold, eliminate the schemes that are dominated in all key performance indicators, and generate an updated strategy library. The domination refers to if a certain candidate strategy is not superior to another strategy in all key performance indicators, it is determined that the former is dominated and is eliminated.

[0106] Through the above steps, when the running data distribution or model error is detected to drift, the embodiment can automatically trigger the recalibration of the twin parameters, and update the strategy library on this basis, so that the system keeps synchronized with the actual state of the production line. This mechanism not only improves the long-term effectiveness of the twin prediction and optimization results, but also ensures that the candidate strategies retained in the strategy library have high stability and improvement value through the dominant screening and improvement amplitude constraint, which means that the AI has the ability of continuous learning and self-correction, and can automatically evolve when the environment changes or the data distribution deviates, thereby enhancing the adaptability and reliability of AI intelligent decision-making in the intelligent factory environment.

[0107] It should be noted that the setting of the interval and the threshold size is for the convenience of comparison, wherein the size of the threshold depends on the number of sample data and the base number set by the person skilled in the art for each group of sample data, as long as it does not affect the proportional relationship of the parameters and the quantized values. And the above formula is a calculation of the dimensionless value, and the formula is obtained by software simulation of a large amount of data to obtain a formula of the nearest true situation, and the preset parameters in the formula are set by the person skilled in the art according to the actual situation.

[0108] The above describes the embodiments of the present application, but the present application is not limited to the above specific embodiments, and the above specific embodiments are only illustrative and not limiting. Those skilled in the art can make many forms under the inspiration of the present embodiment, which are all within the protection of the present embodiment.

Claims

1. A digital twin-driven AI-powered intelligent decision-making system for smart factories, characterized in that: include: The data acquisition and preprocessing module is used to acquire real-time and historical statistical data of the production line, standardize, suppress noise and detect anomalies in the real-time production line data, and obtain the production line state vector, actual observation data, noise parameters and anomaly markers. The twin modeling and identification module is used to construct twins based on production line state vectors and real observation data, perform parameter identification and uncertainty quantification, and output twin prediction data. The alignment assessment confidence module is used to align twin prediction data with real observation data, and obtain availability labels and confidence scores based on preset alignment thresholds and preset confidence thresholds. The strategy simulation and optimization module is used to solve for a set of candidate strategies in a counterfactual scenario set when availability is marked as available, based on key performance indicators and constraints. It then provides key performance indicator predictions and risk evidence, including: Step 41: Construct a counterfactual scenario set based on the production line state vector and preset disturbances, and use trust scores to weight and normalize the scenario probabilities to form a weighted scenario set; Step 42: Use the weighted sum of key performance indicators as the optimization objective, and unify physical constraints, capacity constraints, and risk constraints limited by preset risk thresholds into the constraint conditions; the constraint conditions include: physical constraints, capacity constraints, and risk constraints. Step 43: In the weighted scenario set, call the twin to simulate the control parameter sequence, eliminate the schemes that do not meet the constraints or are inferior to other solutions in all key performance indicators, and select the top-scoring candidate strategies to form a candidate strategy set. The step also includes: synthesizing the predicted variance through model parameter covariance and noise parameters, and tightening the physical constraints and capacity constraints according to the preset confidence level, so that they are transformed into deterministic constraints under the preset confidence level; and adaptively adjusting the duration and accuracy of the control parameter sequence using the predicted variance; and obtaining candidate strategies and screening them based on the tightened constraints and the adaptively set control parameter sequence. Step 44: Output the key performance indicator prediction results and risk evidence for the candidate strategy, and encapsulate the candidate strategy identifier, control parameter sequence, evaluation time window, key performance indicator prediction results, risk evidence and twin version number into a set. The risk evidence includes the worst-case scenario result and risk threshold determination. The gray-scale release module is used to apply a preset release threshold to the candidate strategy set according to the trust score and key performance indicator prediction results, and to distribute the candidate strategy set to generate an online strategy version within the gray-scale release scope. The operation evaluation and audit module is used to collect online policy operation data corresponding to the online policy version, form a comparison evaluation and generate an audit package; The drift detection and update module is used to detect drift based on the feature distribution stability index and online error threshold, and to recalibrate and update the twin and the policy library.

2. The AI-powered intelligent decision-making system for smart factories driven by digital twins according to claim 1, characterized in that, The process of standardizing, noise suppressing, and anomaly detection of real-time production line data includes: Step 11: Perform time alignment and integrity verification on the real-time production line data. Time alignment uses a preset resampling cycle to generate a unified timestamp. Integrity verification removes consecutive missing data segments according to the upper limit of the missing rate in historical statistical data to obtain the true observation data. Step 12: Standardize the actual observation data based on the mean and standard deviation in historical statistical data. When the standard deviation is zero, use the spare standard deviation to replace it, and obtain the standardized series. Step 13: Apply Kalman filtering to the standardized sequence channel by channel to generate a denoised sequence, and use the residuals to estimate the variance of the observed noise and the variance of the process noise to form noise parameters; Step 14: Determine the two-sided threshold based on the noise parameters, perform preliminary threshold determination on the denoised sequence, use a graph neural network model to perform multi-channel correlation analysis on the denoised sequence, identify cross-channel anomalies and generate anomaly markers, and aggregate the timestamps covered by the anomaly markers into a production line state vector after compensation processing with historical averages.

3. The AI-powered intelligent decision-making system for smart factories driven by digital twins according to claim 1, characterized in that, The construction of twins and the generation of twin prediction data include: Step 21: Construct a twin with the production line state vector as input and the predicted value of each channel as output. The twin includes a mechanism item and a data item, and the initial values ​​of the model parameters are set according to historical statistical data. Step 22: Remove corresponding samples using anomaly markers, and determine sample weights based on the observed noise variance and process noise variance in the noise parameters. Combine the retained samples to form a weighted training set. Step 23: Establish an objective function for the sum of squared prediction errors based on the weighted training set, update the twin parameters using an iterative optimization algorithm, and stop iterating when the decrease in the objective function is lower than the first preset threshold to obtain the identified twin. Step 24: Use the identified twin to infer the production line state vector, output twin prediction data, and synthesize prediction variance based on model parameter covariance and noise parameters to obtain uncertainty quantification results.

4. The AI-powered intelligent decision-making system for smart factories driven by digital twins according to claim 1, characterized in that, The process of generating availability tags and trust scores includes: Step 31: Based on the anomaly markers, outliers are removed, and the error between the twin prediction data and the actual observation data is calculated for each channel of the remaining data to obtain the alignment error sequence; Step 32: Generate sample weights based on the sum of the observation noise variance and the process noise variance in the noise parameters, and perform squared normalization on the alignment error sequence according to the sample weights to obtain the overall error index as the weighted root mean square error. Step 33: Compare the overall error index with the second preset threshold set according to historical statistical data, and compare the uncertainty quantification result with the third preset threshold. If both are satisfied, generate an availability mark as available; otherwise, generate an unavailable mark. Step 34: Set a normalization constant based on historical statistical data, map the overall error index to a score through an exponentially decreasing monotonically decreasing function, and multiply it with the availability label to obtain the trust score.

5. The AI-powered intelligent decision-making system for smart factories driven by digital twins according to claim 1, characterized in that, The key performance indicators include: throughput, work-in-process inventory, energy consumption, and overdue rate.

6. The digital twin-driven AI intelligent decision-making system for smart factories according to claim 1, characterized in that, The gray-scale deployment module specifically includes: Step 51: Based on historical statistical data, set the trust score threshold, key performance indicator boundary and preset risk threshold, compare the candidate strategy set one by one, and include the candidate strategy in the online candidate if the threshold set is met; if the screening result is empty, relax the single threshold according to the preset step-down strategy, and only when the improvement of the key performance indicator of the candidate strategy relative to the baseline strategy reaches the preset improvement range, it is determined to be a valid candidate. Step 52: Based on the trust score and key performance indicator margin of the effective candidate strategy, the gray scale ratio is set by a monotonically increasing function, and the gray scale time domain window is set by a monotonically decreasing function based on the risk level. Within the coverage area, units with less impact on the bottleneck process are selected first. Step 53: Generate an online strategy version for the canary release candidate strategy, and record the binding relationship between the rollback trigger conditions and the baseline strategy. When triggered, switch to the baseline strategy.

7. The AI-powered intelligent decision-making system for smart factories driven by digital twins according to claim 1, characterized in that, Collect online policy execution data corresponding to the online policy version, form a comparative evaluation, and generate an audit package, including: Step 61: Collect operational data within the coverage and validity window of the online policy version, and resample and align it according to the data snapshot identifier. Use missing data completion and outlier removal to form a consistent set of operational data. Step 62: Construct a set of controls based on the operational data, prioritizing the use of real controls from units that have not implemented policies, supplementing any deficiencies with corresponding twin versions, and then weighted and fused to form counterfactual key performance indicators; Step 63: Calculate the observed key performance indicators using a unified standard, compare them with counterfactual key performance indicators to generate improvement amounts, multiply the improvement amounts of each key performance indicator by the preset business weights one by one, sum them to obtain the comprehensive effect score, and output the confidence interval at the significance level. Step 64: Within the validity window, calculate the proportion of time periods in which the actual execution value falls within the tolerance range of the preset control parameter sequence to obtain the compliance rate. When the compliance rate is lower than the preset compliance rate threshold, it is marked as low compliance, and the distribution distance of key exogenous variables is detected to generate a drift flag. Step 65: Encapsulate the strategy identifier, control parameter sequence, twin version number, data snapshot identifier, validity window, running data results, control results, effect score, compliance rate, and drift flag fields to generate an audit package.

8. The AI-powered intelligent decision-making system for smart factories driven by digital twins according to claim 1, characterized in that, Based on the feature distribution stability index and online error threshold to detect drift, the twin and policy library are recalibrated and updated, including: Step 71: Extract the feature distribution of the running data within the sliding time window, align it with the baseline snapshot, and calculate the stability index using the binning method; Step 72: The stability index is compared with the preset upper limit threshold and the online error threshold of the key performance index. When the threshold is exceeded for multiple consecutive windows, a drift judgment flag is generated. Step 73: When the drift determination is triggered, select samples that meet the compliance rate requirements and have no abnormal markings, generate weights according to the inverse of the noise variance, and use weighted regularized least squares to iteratively update the twin mechanism and data parameters. Stop when the parameter change is lower than the preset change threshold, and obtain the updated twin parameters. Step 74: Based on the updated twin, re-simulate the strategy library in the weighted scenario, retain candidate strategies that meet the preset risk threshold and whose improvement relative to the baseline strategy reaches the fourth preset threshold, remove schemes that are dominated by all key performance indicators, and generate an updated strategy library.

Images