Flexible production line AI fool-proof simulation verification method based on digital twinning

Abstract

The invention provides a flexible production line AI fool-proof simulation verification method based on digital twinning. The method belongs to the technical field of simulation verification and intelligent manufacturing. The method comprises the following steps: carrying out multi-physics field coupling modeling on a flexible production line, constructing a high-fidelity digital twinborn body, and generating production line total factor digital model data; and deploying a virtual sensor network based on the digital model data, synchronously mapping signal acquisition logic of a real production line, and forming a virtual-real synchronous simulation verification basic environment. Various physical characteristics of a flexible production line are accurately mapped to a virtual space by constructing a high-fidelity multi-physics field digital twinborn body, just like a digital branch is created for the production line, various production scenes can be simulated in advance, the tedious process of carrying out a large number of trial and error verification on a real production line is reduced, and the production efficiency is improved. The problems of high cost and long period caused by trial and error are avoided, and the deployment risk is remarkably reduced.

Description

TECHNICAL FIELD

[0001] The application provides a flexible production line AI fool-proof simulation verification method based on digital twinning, and belongs to the technical field of simulation verification and intelligent manufacturing. BACKGROUND

[0002] In the field of high-end manufacturing such as new energy vehicles, the introduction of flexible production lines greatly improves the flexibility and efficiency of production, but the fool-proof problem that follows is increasingly critical. The traditional flexible production line fool-proof system adopts an engineering mode of "first deployment, then optimization", which has many drawbacks.

[0003] The verification link is seriously lagging behind, and fool-proof logic, such as visual rules and force control thresholds, needs to be verified through a large number of trial and error on the real production line. This process not only has a long cycle and high cost, but also is difficult to cover low-probability high-risk scenarios, such as the insertion of correct parts in the opposite direction, which may cause serious losses.

[0004] The traditional simulation method has a single scene and can only reproduce normal working conditions or a small number of preset faults. It lacks the ability to explore unknown abnormal combinations and cannot fully evaluate the performance of the fool-proof system. At the same time, existing digital twinning focuses on geometric and kinematic simulation, and lacks simulation of key physical details such as sensor noise, material deformation, and environmental interference, resulting in a large deviation between virtual verification results and actual measurements.

[0005] In addition, the evaluation of fool-proof effect relies on the experience of engineers, and lacks quantitative robustness indicators such as maximum tolerance offset, false positive rate-false negative rate trade-off curve, etc., making it difficult to scientifically and accurately judge the pros and cons of the fool-proof system. Therefore, a new simulation verification method is needed to realize high-fidelity modeling, intelligent scene generation, automatic verification closed loop and quantitative evaluation output, and to provide a reliable virtual verification infrastructure for intelligent manufacturing. SUMMARY

[0006] The application provides a flexible production line AI fool-proof simulation verification method based on digital twinning, which solves the problems mentioned in the background art:

[0007] The flexible production line AI fool-proof simulation verification method based on digital twinning provided by the application comprises the following steps:

[0008] S1, multi-physics field coupling modeling of the flexible production line is performed to construct a high-fidelity digital twin and generate production line full-element digital model data; based on the digital model data, a virtual sensor network is deployed to synchronize the signal acquisition logic of the real production line and form a virtual-real synchronous simulation verification environment;

[0009] S2, using a generative adversarial network to reverse deconstruct normal working condition data and recombine abnormal features, autonomously evolving to generate a massive virtual abnormal case library covering low-probability high-risk scenarios; through reinforcement learning algorithm to dynamically expand the case library, generate high-coverage abnormal scenario data set;

[0010] S3, embed AI foolproof strategy into digital twin, conduct full-automatic stress test in virtual abnormal scenario data set, record the foolproof response results under each scenario, generate foolproof strategy verification original data; perform noise filtering and time series alignment on the original data to generate standardized verification data;

[0011] S4, based on the standardized verification data, a quantitative robustness evaluation system is constructed; through cluster analysis to identify the weak scene cluster of the foolproof strategy, combined with Monte Carlo simulation to predict the failure probability under extreme working conditions, to generate a foolproof strategy robustness evaluation report;

[0012] S5, according to the robustness evaluation report, generate foolproof strategy optimization suggestions, verify the optimization effect through digital twin closed loop; iteratively execute steps S2-S5 until all key indicators meet the preset threshold, finally output the AI foolproof strategy configuration package and robustness level certificate passed by quantitative authentication.

[0013] The flexible production line AI foolproof simulation verification system based on digital twinning proposed by the application comprises:

[0014] One or more processors;

[0015] Memory for storing one or more programs;

[0016] When the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of the above.

[0017] The present application has the following advantages: by constructing a high-fidelity multi-physical field digital twin, the various physical characteristics of the flexible production line are accurately mapped to the virtual space, like creating a digital twin for the production line, which can simulate various production scenarios in advance, reduce the tedious process of a large number of trial and error verification on the real production line, avoid high cost and long cycle problems caused by trial and error, significantly reduce the deployment risk, change the post-failure verification to pre-prevention, and improve the efficiency and reliability of the entire verification process; using the generated adversarial network to autonomously evolve a large number of abnormal scenarios, like preparing a super training ground for the foolproof strategy, which can cover low-probability high-risk scenarios and unknown abnormal combinations, and fully explore the boundaries of the foolproof strategy. This effectively avoids the shortcomings of single simulation scenario in traditional simulation, which cannot cope with complex abnormal situations, reduces the probability of foolproof failure caused by not considering special scenarios, and enhances the adaptability of the foolproof strategy to various complex working conditions; based on the standardized verification data, a quantitative robustness evaluation system is constructed to evaluate the foolproof strategy with scientific and accurate indicators (such as maximum tolerance offset, false alarm rate-miss rate trade-off curve, etc.), which breaks away from the subjective judgment of relying on engineers' experience. This can clearly show the pros and cons of the foolproof strategy, provide a clear direction for optimization, avoid resource waste caused by blind optimization, and improve the quality and stability of the foolproof strategy. Through closed-loop verification and iterative optimization, the foolproof strategy is continuously polished, and finally a configuration package and a robustness level certificate that pass quantitative authentication are output, providing a reliable and fully verified foolproof solution for intelligent manufacturing, making the operation of the flexible production line more secure and stable, and enhancing the competitiveness of the entire intelligent manufacturing system. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The method steps described in the present application. DETAILED DESCRIPTION

[0019] The preferred embodiments of the present application are described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application.

[0020] As shown in one embodiment of the present application, Figure 1 The AI foolproof simulation verification method for the flexible production line based on digital twinning includes:

[0021] S1, multi-physical field coupling modeling is performed on the flexible production line to construct a high-fidelity digital twin containing geometric structure, kinematic characteristics, sensor noise, material deformation and environmental interference factors, and generate a production line full-element digital model data; based on the digital model data, a virtual sensor network is deployed to synchronize the signal acquisition logic of the real production line and form a virtual-real synchronous simulation verification basic environment;

[0022] S2, using a generative adversarial network (GAN) to reverse deconstruct normal working condition data and reorganize abnormal features, autonomously evolving to generate a massive virtual abnormal case library covering low-probability high-risk scenarios (such as reverse insertion of parts, non-standard deformation, and multiple fault superposition); through reinforcement learning algorithm, the case library is dynamically expanded to ensure exponential growth of scene complexity and diversity index, generating a high-coverage abnormal scene dataset;

[0023] S3, embedding AI foolproof strategy (including visual recognition rules, force control threshold, and logic judgment tree) into digital twin, performing full-automatic stress testing in virtual abnormal scene dataset, recording foolproof response results (triggered / not triggered, response time, false alarm / missed alarm type) under each scene, generating foolproof strategy verification original data; performing noise filtering and time series alignment on the original data to generate standardized verification data;

[0024] S4, based on the standardized verification data, a quantitative robustness evaluation system is constructed to calculate key indicators including maximum tolerance deviation (MTD), false alarm rate-missed alarm rate trade-off curve (ROC-AUC), and scene coverage index (SCI); through cluster analysis, weak scene clusters of the foolproof strategy are identified, and the failure probability under extreme working conditions is predicted by Monte Carlo simulation, generating a foolproof strategy robustness evaluation report;

[0025] S5, according to the robustness evaluation report, generating foolproof strategy optimization suggestions (such as adjusting visual recognition threshold, increasing force control redundancy, and optimizing logic judgment priority), and verifying the optimization effect through digital twin closed loop; iteratively executing steps S2-S5 until all key indicators meet the preset threshold, finally outputting the AI foolproof strategy configuration package and robustness level certificate that pass quantitative authentication.

[0026] The working principle and effect of the above technical solution are: high-fidelity digital twin and virtual-real synchronous simulation environment make the test of AI foolproof strategy more close to real production scenarios, greatly improving the credibility of the verification. Through GAN and reinforcement learning, a large number of abnormal cases containing low-probability high-risk scenarios are generated, significantly enhancing the comprehensiveness of scene coverage and avoiding production accidents caused by missing key fault scenarios; full-automatic stress testing replaces manual testing, reducing human error in the testing process, and greatly reducing testing cost and improving testing efficiency. Combined with cluster analysis to locate weak scenes and Monte Carlo simulation to predict failure probability under extreme working conditions, the foolproof strategy can be accurately optimized to effectively reduce false alarm and missed alarm rates in actual production, enhancing the robustness of the strategy; through closed-loop iterative optimization, the foolproof strategy is continuously improved to avoid production line downtime or defective product output caused by strategy defects; this method can not only guarantee high-precision assembly requirements, but also improve production line

[0027] In an embodiment of the present application, the S1 comprises:

[0028] S11, for the core elements of the flexible production line, collect multi-dimensional data, the multi-dimensional data includes geometric structure data (such as mechanical arm joint parameters, conveying track size, tooling fixture model), kinematics characteristic data (such as the motion speed of each actuator, acceleration limit, joint rotation angle range), material attribute data (such as part elastic modulus, surface friction coefficient), environmental disturbance data (such as workshop temperature fluctuation range, vibration frequency, electromagnetic interference intensity) and real sensor data (such as visual camera resolution, force sensor range, position sensor sampling frequency), generate a flexible production line multi-dimensional original data set;

[0029] S12, based on the generated multi-dimensional original data set, a multi-physical field coupling model covering structural mechanics (simulating material deformation), dynamics (simulating motion interference), sensor noise (simulating signal fluctuation) and environmental disturbance (simulating external influence) is constructed by using finite element analysis, and a preliminary digital twin model of the flexible production line is generated;

[0030] S13, compare the generated preliminary digital twin model with the key operation parameters (such as mechanical arm repeated positioning accuracy, sensor data error range) of the real production line, calculate the deviation value; for the modules with excessive deviation (such as material deformation simulation module, environmental disturbance simulation module), the parameters are corrected until the consistency of the model and the real production line meets the preset threshold (such as consistency ≥ 98%), and the production line full-element digital model data is generated;

[0031] S14, based on the generated production line full-element digital model data, according to the sensor layout (such as visual camera installation position, force sensor deployment node) of the real production line, signal acquisition logic (such as sampling frequency, data transmission protocol) is deployed virtual sensor; configure the interaction interface of virtual sensor and digital twin, ensure that virtual sensor can collect real-time operation data of digital twin, generate preliminary virtual sensor network;

[0032] S15, virtual-real synchronization test is performed on the generated preliminary virtual sensor network: simulate typical working conditions (such as normal assembly of parts, uniform speed running of mechanical arm) in the real production line, compare the time delay and data deviation of the data stream collected by the virtual sensor and the data stream collected by the real sensor; for the items that do not meet the synchronization (such as delay > 10ms, deviation > 5%), optimize the interface protocol or sampling frequency, and form a virtual-real synchronous simulation verification environment.

[0033] The working principle and effect of the above technical solution are:

[0034] An embodiment of the present application, the S13, comprising:

[0035] S131, based on the core functional requirements of the flexible production line (such as assembly accuracy, operation stability), key operating parameters for comparison are selected from the parameter system of the real production line. In addition to the previously mentioned mechanical arm repeatability, sensor data error range, tooling fixture positioning error, actuator response delay, and environmental disturbance parameter fluctuation range are added. The definition, unit, and measurement standard of each parameter are defined, and a key operating parameter comparison list is generated.

[0036] S132, according to the generated comparison list, the measurement equipment (such as laser interferometer for positioning accuracy, oscilloscope for response delay) is used to collect parameters continuously for 24 hours, and the actual values of each parameter under different working conditions (such as empty load, full load, normal temperature, high temperature interference) are recorded. After removing abnormal values (such as extreme data caused by sudden equipment failure), the real production line key operating parameter dataset is generated.

[0037] S133, the generated preliminary digital twin model of the flexible production line is run under the same working conditions as the real production line, and the key operating parameters simulated by the model are output. The model parameters are compared with the generated real production line key operating parameter dataset, and the relative deviation value is calculated using (model parameter-real parameter) / real parameter x 100%. According to the parameter type, a model and real parameter deviation analysis table is arranged.

[0038] S134, set the deviation threshold (such as positioning accuracy class parameter deviation ≤2%, response delay class parameter deviation ≤5%), compare the generated deviation analysis table, and select the model modules corresponding to the parameters with deviation values exceeding the threshold, such as material deformation simulation module (corresponding to part deformation parameter deviation exceeding 3%) and environmental disturbance simulation module (corresponding to temperature interference parameter fluctuation deviation exceeding 6%). The deviation exceeding model module list is generated.

[0039] S135, for the generated exceeding module list, analyze the deviation reasons (such as the material deformation module does not consider the part fatigue coefficient, the environmental disturbance module does not consider the workshop air flow influence), adjust the core parameters of the module (such as supplementing the part fatigue coefficient to the material attribute library, adding the air flow disturbance coefficient to the environmental model), and integrate the corrected module into the preliminary digital twin model again. Under the same working conditions, the simulation is run again, and the preliminary digital twin model with corrected parameters is generated.

[0040] S136, the generated preliminary digital twin model with corrected parameters is compared with the real production line key parameters again, and the corrected deviation value is calculated. If the deviation values of all key parameters meet the preset threshold (such as consistency ≥98%), the production line full-element digital model data is directly output. If there are still exceeding parameters, return to S134 to reselect the exceeding modules and iterate the correction until the model consistency meets the standard.

[0041] The working principle and effects of the above technical solution are as follows: comprehensive collection of multi-dimensional data covers core data such as geometry, kinematics and environment, provides solid data source support for subsequent modeling, and greatly improves the accuracy of model construction. A multi-physics field coupling model is constructed by using finite element analysis, and key factors such as material deformation and motion interference are included, thereby enhancing the model's ability to reproduce the real running state of the production line. By comparing and correcting with the real production line parameters, the model consistency is improved to more than 98%, effectively avoiding the deviation problem of subsequent verification results caused by model distortion. Virtual sensors are deployed according to the real layout and logic, reducing the logical difference between virtual and real signal collection, and making the virtual data more consistent with the real scene. After virtual-real synchronous testing optimization, the delay and deviation are controlled in a very low range, ensuring the synchronization of the simulation environment and the real production line, and avoiding the verification failure caused by different data synchronization. The whole process not only builds a reliable simulation foundation for subsequent AI foolproof strategy testing, but also reduces the data correction cost in subsequent testing, significantly improving the overall verification efficiency.

[0042] In an embodiment of the present application, the S2 comprises:

[0043] S21, collect the working condition data of the flexible production line running continuously for 72 hours, the working condition data of the normal running including visual image data (part assembly posture, tool positioning state), force control data (mechanical arm assembly pressure, clamping force time curve), sensor monitoring data (position sensor coordinates, temperature sensor readings); the data is preprocessed, the preprocessing including deduplication, noise reduction (for example, Gaussian filtering is used to process the fluctuation of force control data), standardization (for example, the image size is unified to 1024*1024 pixels), and a preprocessed normal working condition data set is generated;

[0044] S22, input the generated preprocessed normal working condition data set into a generative adversarial network: learn the feature distribution of the normal working condition data (for example, the geometric features of the normal assembly of parts, the smooth interval of the force control curve) through the discriminator, and inversely deconstruct the feature law and recombine the abnormal features (for example, the geometric deviation of the reverse insertion of parts, the contour anomaly of the non-standard deformation, and the sudden change of the sensor signal of a single fault) through the generator; iteratively train the generative adversarial network until the generated abnormal cases meet the physical rationality (for example, the part deformation does not exceed the material limit), and generate an initial virtual abnormal case library;

[0045] S23, purify the generated initial virtual abnormal case library by using rule checking (for example, eliminate cases where the size of the part exceeds the adaptation range of the production line); delete invalid cases (for example, cases where the physical law is not established) and repeated cases (for example, cases with a feature similarity of more than 90%), and generate a clean initial abnormal case library;

[0046] S24, based on the generated clean initial abnormal case library, build a reinforcement learning agent: set the reward mechanism (for example, generate an uncovered abnormal type reward +20 points, generate a multi-fault superposition scene reward +50 points), the punishment mechanism (for example, generate a repeated scene penalty-10 points); let the agent autonomously explore low-probability high-risk scenarios (for example, part reverse + sensor noise superposition, non-standard deformation + mechanical arm load limit), dynamically expand the complexity and diversity of the case library, and generate an expanded abnormal case library;

[0047] S25, define scene coverage (number of covered abnormal types / total number of potential abnormal types), complexity index (number of fault types contained in a single scene x fault severity) as evaluation indicators, and quantitatively evaluate the generated expanded abnormal case library; if the scene coverage is greater than or equal to 95% and the complexity index is greater than or equal to a preset threshold (for example, greater than or equal to 3), a high-coverage abnormal scene data set is directly generated; if the requirements are not met, return to S24 for continuous expansion until the evaluation requirements are met.

[0048] The working principle and effect of the above technical solution are: through 72 hours of continuous collection of normal working condition data and preprocessing, the data quality after deduplication and noise reduction is more reliable, which lays a solid foundation for subsequent generation of abnormal cases, and greatly improves the rationality of the initial cases. Through the generation of adversarial network to learn the normal data features and reorganize the abnormality, the boundary between abnormality and normality can be accurately grasped, and it is ensured that the generated abnormal cases conform to the physical law, avoiding invalid cases such as part deformation exceeding the material limit. The rule verification purification link eliminates invalid and repeated cases, reduces the redundancy of subsequent testing, and improves the cleanliness of the case library. The reward and punishment mechanism of reinforcement learning is very important, which can drive the agent to actively explore low-probability high-risk scenarios such as part reverse superposition sensor noise, significantly enhancing the complexity and diversity of the case library. Finally, through quantitative evaluation, it is ensured that the scene coverage is not less than 95% and the complexity meets the requirements before output, avoiding the problem of missing key fault scenarios. This way can not only provide comprehensive and practical materials for AI foolproof strategy pressure testing, but also reduce the downtime loss caused by real production line trial and error, making the foolproof strategy testing more targeted.

[0049] In an embodiment of the present application, the S3 comprises:

[0050] S31, disassemble the constituent modules of the AI foolproof strategy, including a visual recognition module (for example, edge detection algorithm, part posture judgment logic), a force control threshold module (for example, maximum clamping force, minimum assembly pressure), a logic judgment tree module (for example, position detection-force control verification-assembly confirmation judgment node and priority); extract the key parameters of each module (for example, confidence threshold of visual recognition, alarm threshold of force control), and generate an AI foolproof strategy analysis document;

[0051] S32, based on the generated AI fool-proof strategy analysis document, the visual recognition module is connected to the virtual camera data stream of the digital twin (for example, real-time acquisition of part assembly image), the force control threshold module is connected to the force feedback unit of the virtual mechanical arm (for example, real-time collection of clamping force data), and the logic judgment tree module is connected to the working condition decision interface of the digital twin; through interface debugging, it is ensured that each module can receive data in real time and output judgment results, and the integration of AI fool-proof strategy and digital twin is completed;

[0052] S33, according to the generated high-coverage abnormal scene data set, the scene is grouped according to the abnormal type (for example, part posture abnormality, force control overrun abnormality, and multi-fault superposition abnormality); set the test times of each group of scenes (for example, test 50 times for each group to ensure the significance of result statistics), test working condition parameters (for example, mechanical arm running speed, tool fixture positioning accuracy), data record items (for example, fool-proof triggering state, response time, false alarm / missed alarm type), and generate a stress test execution plan;

[0053] S34, according to the generated stress test execution plan, run the high-coverage abnormal scene in the digital twin by group; record the response results of the AI fool-proof strategy under each scene in real time: including the state of triggering fool-proof / triggering fool-proof, the time from scene occurrence to fool-proof response (accurate to milliseconds), the specific type of false alarm (normal scene misjudged as abnormal) / missed alarm (abnormal scene not identified) (for example, visual false alarm, force control missed alarm), and generate a fool-proof strategy verification original data set;

[0054] S35, pretreatment of the generated fool-proof strategy verification original data set: adopt Kalman filter to eliminate noise interference of time series data (for example, response time sequence), align response data of different modules through time stamp (for example, time synchronization of visual recognition result and force control judgment result), adopt box plot method to eliminate abnormal value (for example, invalid response time caused by test interruption); the processed data is regularized according to the format of scene number-fool-proof response result-key parameter, and the standardized verification data is generated.

[0055] The working principle and effect of the above technical solution are that: the anti-fooling strategy module is disassembled and key parameters are extracted, the analysis document formed makes the docking target of each module more clear, effectively avoids the problems of interface confusion and parameter mismatch during integration, and greatly improves the integration efficiency of the strategy and the digital twin. After the precise docking of the module and the twin, the visual and force control modules can obtain virtual data in real time and feed back the results, making the test scene closer to the real running state and enhancing the credibility of the stress test. Group testing according to abnormal types and setting 50 repeated times can ensure the statistical significance of the results and avoid the influence of accidental errors of single testing on judgment. Real-time recording of details such as trigger state, response time and false and missing report types makes the original data more complete, provides a clear basis for subsequent analysis of strategy problems, and reduces the difficulty of problem positioning caused by missing data. The preprocessing operations such as Kalman filter denoising and timestamp alignment effectively eliminate invalid data and noise interference, greatly improve the standardization degree of the verification data, and avoid the interference of dirty data on subsequent robustness evaluation. The whole process can not only reduce the labor cost through full automation testing, but also ensure the accuracy of the verification results through standardized data processing, and provides a solid support for strategy optimization.

[0056] In one embodiment of the present application, the S4 comprises:

[0057] S41, in combination with the production needs of the flexible production line (such as the high-precision requirement of automobile part assembly and the high-efficiency requirement of 3C product assembly), the core evaluation dimensions of AI anti-fool strategy robustness are determined: including fault tolerance ability (ability to cope with part deviation and noise interference), judgment accuracy (balance ability of false alarm and missed alarm), and scene adaptability (ability to cover various abnormal scenes), and a robustness evaluation dimension definition document is generated;

[0058] S42, based on the generated evaluation dimension definition document, the key indicators are designed to be quantifiable and the calculation method is clear:

[0059] Maximum Tolerable Deviation (MTD): the maximum position deviation value of the part that the anti-fool strategy can still respond correctly is calculated by linear interpolation;

[0060] False Alarm Rate-Missed Alarm Rate Trade-off Curve (ROC-AUC): the curve is drawn with false alarm rate as horizontal axis and missed alarm rate as vertical axis, and the area under the curve (AUC) is calculated;

[0061] Scene Coverage Index (SCI): the number of abnormal scenes correctly identified / total number of abnormal scenes x 100%;

[0062] The index definition and calculation method are sorted out, and a quantitative robustness evaluation index manual is generated.

[0063] S43, put the generated standardized verification data into the prepared index calculation method: extract the verification data of the part offset scene to calculate MTD, count the number of false positives / misreporting of all scenes to draw an ROC curve and calculate AUC, count the number of correct recognitions of each abnormal scene to calculate SCI; organize the calculation results according to the index type to generate a robustness index quantification result table;

[0064] S44, adopt K-means clustering algorithm to cluster the generated standardized verification data: take abnormal type, fault severity and foolproof response result as clustering features, and cluster scenes where the foolproof strategy fails (misses) or response delay exceeds the standard into one class; analyze the common features of each cluster (for example, small size parts + visual recognition miss cluster, high load + force control non-triggering cluster), and generate a foolproof strategy weak scene cluster analysis report;

[0065] S45, based on the identified weak scene cluster parameter range (for example, part size <5mm, load >120% of the rated value), 10000 sets of extreme working condition samples (for example, scenes where small size parts + high load + sensor noise exist at the same time) are generated by Monte Carlo simulation; input the samples into the digital twin, count the number of failures of the foolproof strategy, and calculate the failure probability; integrate the robustness index quantification result, the weak scene cluster analysis, and the extreme working condition failure probability to generate a foolproof strategy robustness evaluation report.

[0066] The working principle and effect of the above technical solution are: combined with the actual demand of the production line to determine the core evaluation dimensions such as fault tolerance, so that the robustness evaluation does not deviate from the actual production, effectively avoiding the evaluation error of generalization, and greatly improving the pertinence of the evaluation. Design MTD, ROC-AUC, SCI and other quantitative indexes and clearly define the calculation method, use data to replace qualitative judgment, significantly improve the accuracy of the evaluation, and reduce the error caused by subjective speculation. Put the standardized data into the calculation to obtain a quantification result table, so that the performance of the foolproof strategy is clear at a glance, providing solid data support for subsequent analysis and avoiding the lack of data leading to insufficient evaluation persuasiveness. Through K-means clustering, the weak scene clusters such as small size part visual miss and high load force control non-triggering are accurately located, so that the strategy problem has no place to hide, and the invalid investment of blind optimization is reduced. Monte Carlo simulation generates 10000 sets of extreme working condition samples, calculates the failure probability in advance, and effectively avoids the problem of stop line or defective product caused by extreme failure in actual production. The whole process integrates the multi-dimensional evaluation results to form a report, which can not only reflect the performance of the strategy, but also point out the accurate direction for optimization, and lay a solid foundation for subsequent strategy improvement.

[0067] In one embodiment of the present application, the S44 comprises:

[0068] Extract the cluster core features in the standardized verification data, including abnormal type, fault severity, foolproof response result, and perform quantitative conversion, including assigning values to abnormal types (e.g., part posture abnormality, force control overrun) according to coding rules (e.g., 1-part abnormality, 2-force control abnormality), fault severity (mild / medium / severe) corresponding to 1 / 2 / 3 values, and foolproof response result (missed report / response delay / normal) corresponding to 1 / 2 / 0 labels, to generate a cluster feature quantization dataset;

[0069] Use the elbow rule combined with the silhouette coefficient method to evaluate the generated cluster feature quantization dataset: test different cases from K=2 to K=8, calculate the within-cluster sum of squares (WCSS) and silhouette coefficient for each K value, select the K value with the maximum silhouette coefficient at the WCSS curve inflection point (e.g., K=4), and generate the optimal cluster number determination result for K-means;

[0070] Input the cluster feature quantization dataset and the determined optimal K value into the K-means algorithm, initialize the cluster centers, and iteratively calculate the Euclidean distance of each sample to the cluster center to continuously update the cluster center until convergence (e.g., the center variation is less than 0.001 for 3 consecutive iterations), and generate the preliminary scene clustering result (including the cluster number and corresponding cluster center of each scene);

[0071] From the preliminary scene clustering result, filter out clusters where the majority of samples are mainly due to foolproof failure (missed report) or response delay exceeding the standard (e.g., a cluster where more than 80% of samples are missed report scenarios), and remove invalid clusters dominated by normal response samples, to generate an effective weak scene cluster list (including the number of samples in each cluster and the proportion of failure types);

[0072] For the effective weak scene cluster list, extract the common features of each cluster one by one: for example, 75% of samples in cluster 1 are part size <5mm + visual recognition missed report, and 90% of samples in cluster 2 are load >120% of rated value + force control not triggered, and organize the abnormal types, fault parameter ranges, and failure modes of each cluster to generate a weak scene cluster common feature analysis table;

[0073] Integrate the weak scene cluster common feature analysis table, supplement the risk impact of each cluster on the production line (e.g., missed report of small size parts easily leads to assembly rework), sample proportion, etc., and organize the content according to the structure of cluster number-common feature-risk level, to finally generate a foolproof strategy weak scene cluster analysis report.

[0074] The working principle and effect of the above technical solution are that: the non-numeric characteristic values such as abnormal types and fault degrees are quantitatively assigned, which provides a calculable basis for cluster analysis, effectively avoids the problem that cluster cannot be carried out or the result is chaotic due to the chaotic form of characteristics, and greatly improves the accuracy of analysis. The elbow rule is combined with the contour coefficient to select the K value, which is not set empirically, and the number of clusters most suitable for the characteristics of the data can be accurately found, the rationality of the clustering result is enhanced, and the problem of too rough cluster number missing key problems or too fine cluster number causing analysis redundancy is avoided. Iterative calculation is performed until the cluster center converges stably, ensuring that the division of the scene cluster is solid and reliable, and reducing the clustering deviation caused by accidental fluctuations. The effective cluster mainly composed of failures and delays is specially screened out, and the invalid cluster with normal response is eliminated, so that the analysis can focus on the real weak points, without wasting energy on useless data, and the analysis efficiency is improved. When extracting the common characteristics of each cluster, the specific problems such as visual miss report of small size parts and non-triggering of high load force control can be accurately located, and the blindness of general problem finding is avoided. The risk influence in the report is supplemented with a standard structure, which can not only point out the accurate direction for subsequent optimization and make the adjustment targeted, but also enable decision makers to clearly understand the risk level of each problem, thereby providing a solid basis for decision making and significantly reducing the trial and error cost of optimization.

[0075] In one embodiment of the present application, the S5 comprises:

[0076] S51, analyze the generated robustness evaluation report, including for the problem of insufficient MTD, suggesting adjusting the edge detection threshold of visual recognition (for example, reducing the threshold from 0.8 to 0.7 to expand the recognition range); for the problem of low ROC-AUC, suggesting increasing the redundancy of force control (for example, setting up "main threshold + backup threshold" double judgment); for the weak scene cluster with low SCI, suggesting optimizing the priority of logical judgment (for example, advancing "small size part detection" to the first node of the judgment tree); and sorting the suggestions to form an AI foolproof strategy optimization suggestion list;

[0077] S52, according to the generated optimization suggestion list, updating the parameter configuration (such as visual threshold, force control redundancy parameter, judgment tree node order) of the AI foolproof strategy, and re-embedding the digital twin; selecting typical scenes (such as small size part reverse assembly, force control anomaly under high load) in the identified weak scene cluster, and performing closed loop verification: comparing the response time, false positive rate and false negative rate changes of the foolproof strategy before and after optimization, and generating a foolproof strategy optimization effect verification report.

[0078] S53, checking whether the key indicators (MTD, ROC-AUC, SCI) in the generated optimization effect verification report meet the preset threshold (for example, MTD≥1.2 times of the design requirement, ROC-AUC≥0.92, SCI≥95%):

[0079] If not up to standard: return to step S2, generate new abnormal cases based on the current weak scene (for example, expand the case library for small-size part abnormal scenes not covered), and re-execute the iterative process of S2-S5;

[0080] If up to standard: proceed to the next step;

[0081] S54, collate the complete parameters of the verified AI foolproof strategy: including the algorithm type and threshold of the visual recognition module, the main / backup threshold and redundancy logic of the force control module, and the complete node sequence and trigger condition of the logical judgment tree; encapsulate the parameters in a compatible format (such as XML, JSON) of the production line control system, and generate an AI foolproof strategy configuration package that can be directly deployed to the real production line;

[0082] S55, determine the robustness level according to the key indicator standard (for example, A level: all indicators are greater than or equal to the optimal threshold; B level: core indicators meet the standard, and secondary indicators are close to the threshold), and make a robustness level certificate containing the basis for level evaluation, key indicator values, and evaluation date; finally output the complete product of AI foolproof strategy configuration package + robustness level certificate, and complete the flexible production line AI foolproof simulation verification task.

[0083] The working principle and effect of the above technical solution are: after analyzing the evaluation report, the optimization suggestions are targeted, such as adjusting the visual threshold when MTD is insufficient, and adding force control redundancy when ROC-AUC is low, which avoids blind parameter adjustment and greatly improves the accuracy of optimization. Embed the updated strategy into the twin body, and specifically select the weak scene for closed-loop verification. By comparing the response time and false alarm rate before and after optimization, the effect can be clearly seen, ensuring that the optimization measures really take effect. The mechanism of iterative expansion of cases for optimization when not up to standard can force the strategy to continuously improve until all key indicators meet the standard, effectively avoiding production problems such as small-size part false alarm and high-load force control failure after the strategy with defects is put into operation, reducing assembly rework and downtime loss. The complete parameters are packaged in a compatible format to generate a configuration package that can be directly deployed to the real production line, without the need for further time-consuming adaptation and debugging, significantly reducing deployment costs. According to the indicator level, a certificate is issued, which provides a quantitative authentication basis for the performance of the strategy, making it easy for the production line to quickly judge the applicability and providing a clear reference for subsequent maintenance and upgrade, making the entire verification result more practical.

[0084] One embodiment of the present application is a flexible production line AI foolproof simulation verification system based on digital twinning, which includes:

[0085] One or more processors;

[0086] Memory for storing one or more programs;

[0087] wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to carry out the method of any of the above.

[0088] Obviously, many modifications and changes can be made to the present application without departing from the spirit and scope of the application. It is not therefore desired to limit the present application to the exact details shown and described herein. Accordingly, wherever a discretion is possible, it is intended to embrace all such modifications and changes as fall within the scope of the present application including its general concept and features.

Claims

1. A flexible production line AI fool-proof simulation verification method based on digital twinning, characterized in that, The method comprises: S1, multi-physical field coupling modeling is performed on the flexible production line to construct a high-fidelity digital twin, generate a production line full-element digital model data, and deploy a virtual sensor network based on the digital model data to synchronize the signal acquisition logic of the real production line and form a virtual-real synchronous simulation verification basic environment; S2, the normal working condition data is inversely deconstructed and abnormal feature is reorganized by using a generative adversarial network to autonomously evolve to generate a large number of virtual abnormal case libraries covering low-probability high-risk scenarios; the case library is dynamically expanded by using a reinforcement learning algorithm to generate a high-coverage abnormal scenario data set; S3, the AI foolproof strategy is embedded into the digital twin, and full-automatic stress testing is performed in the virtual abnormal scenario data set, the foolproof response results under each scenario are recorded, and foolproof strategy verification original data are generated; the original data are subjected to noise filtering and time sequence alignment processing to generate standardized verification data; S4, a quantitative robustness evaluation system is constructed based on the standardized verification data; the weak scenario clusters of the foolproof strategy are identified by cluster analysis, the failure probability under extreme working conditions is predicted by Monte Carlo simulation, and a foolproof strategy robustness evaluation report is generated; S5, the foolproof strategy optimization suggestions are generated according to the robustness evaluation report, the optimization effect is verified through the digital twin closed loop, and steps S2-S5 are iteratively executed until all key indicators meet the preset threshold, and finally an AI foolproof strategy configuration package and a robustness level certificate that pass quantitative authentication are output.

2. The flexible production line AI prevention of fooling simulation verification method based on digital twinning according to claim 1, characterized in that, The S1 comprises: S11, multi-dimensional data is collected for the core constituent elements of the flexible production line to generate a flexible production line multi-dimensional original data set; S12, a multi-physical field coupling model is constructed by using finite element analysis based on the generated multi-dimensional original data set to generate a preliminary digital twin model of the flexible production line; S13, the generated preliminary digital twin model is compared with the key operating parameters of the real production line to calculate the deviation value; the parameters of the modules with excessive deviation are corrected until the consistency of the model and the real production line meets the preset threshold to generate production line full-element digital model data; S14, based on the generated production line full-element digital model data, virtual sensors are deployed according to the sensor layout and signal acquisition logic of the real production line; the interaction interface of the virtual sensors and the digital twin is configured to ensure that the virtual sensors can collect the operating data of the digital twin in real time to generate a preliminary virtual sensor network; S15, the generated preliminary virtual sensor network is subjected to virtual-real synchronous testing; the interface protocol or sampling frequency is optimized for the items that do not meet the synchronization to form a virtual-real synchronous simulation verification basic environment.

3. The flexible production line AI prevention of fooling simulation verification method based on digital twinning according to claim 1, characterized in that, The S2 comprises: S21, working condition data of the flexible production line continuously running for 72 hours is collected; the data is preprocessed to generate a preprocessed normal working condition data set; S22, the generated preprocessed normal working condition data set is input into the generative adversarial network; the generative adversarial network is iteratively trained until the generated abnormal cases meet the physical rationality to generate an initial virtual abnormal case library; S23, the generated initial virtual abnormal case library is purified by rule checking; invalid cases and repeated cases are deleted to generate a clean initial abnormal case library; S24, based on the generated clean initial abnormal case library, an intelligent agent of reinforcement learning is constructed; the intelligent agent is allowed to autonomously explore low-probability high-risk scenarios, dynamically expand the complexity and diversity of the case library, and generate an expanded abnormal case library; S25, scene coverage and complexity index are defined as evaluation indexes, and the generated expanded abnormal case library is quantitatively evaluated; if the scene coverage is greater than or equal to 95% and the complexity index is greater than or equal to a preset threshold, a high-coverage abnormal scenario data set is directly generated; if the evaluation requirements are not met, S24 is returned to continue expansion until the evaluation requirements are met.

4. The flexible production line AI prevention of fooling simulation verification method based on digital twinning according to claim 1, characterized in that, The S3 comprises: S31, the constituent modules of the AI foolproof strategy are disassembled; key parameters of each module are extracted, and an AI foolproof strategy analysis document is generated; S32, based on the generated AI foolproof strategy analysis document, the visual recognition module is connected to the virtual camera data stream of the digital twin, the force control threshold module is connected to the force feedback unit of the virtual mechanical arm, and the logic judgment tree module is connected to the working condition decision interface of the digital twin; S33, according to the generated high-coverage abnormal scenario data set, the scenes are grouped according to the abnormal types; the test number, test working condition parameters and data record items of each group of scenes are set, and a stress test execution plan is generated; S34, according to the generated stress test execution plan, the high-coverage abnormal scenarios are run in the digital twin in groups; the response results of the AI foolproof strategy under each scene are recorded in real time, and a foolproof strategy verification original data set is generated; S35, the generated foolproof strategy verification original data set is preprocessed; the processed data are normalized according to the format of scene number-foolproof response result-key parameter, and a standardized verification data is generated.

5. The flexible production line AI prevention of fooling simulation verification method based on digital twinning according to claim 1, characterized in that, The S4 comprises: S41, in combination with the production requirements of the flexible production line, core evaluation dimensions of the AI foolproof strategy robustness are determined, and a robustness evaluation dimension definition document is generated; S42, based on the generated evaluation dimension definition document, quantifiable key indicators are designed and the calculation method is specified; S43, the generated standardized verification data is substituted into the specified index calculation method; the calculation results are arranged according to the index type, and a robustness index quantification result table is generated; S44, the K-means clustering algorithm is used to cluster the generated standardized verification data, and a foolproof strategy weak scene cluster analysis report is generated; S45, based on the identified weak scene cluster parameter range, 10,000 groups of extreme working condition samples are generated by using the Monte Carlo simulation; the samples are input into the digital twin, the failure number of the foolproof strategy is counted, and the failure probability is calculated; the robustness index quantification result, the weak scene cluster analysis and the extreme working condition failure probability are integrated, and a foolproof strategy robustness evaluation report is generated.

6. The flexible line AI prevention simulation verification method based on digital twinning according to claim 5, wherein, The design of quantifiable key indicators and the specification of calculation methods comprise: Maximum tolerable offset: the maximum position offset value of the part that the foolproof strategy can still correctly respond to is calculated by linear interpolation; False positive rate-missed call rate trade-off curve: a curve is drawn with the false positive rate as the horizontal axis and the missed call rate as the vertical axis, and the area under the curve is calculated; Scene coverage index: the number of correctly identified abnormal scenes / total abnormal scenes x 100%; The index definition and calculation method are arranged, and a quantified robustness evaluation index manual is generated.

7. The flexible line AI prevention simulation verification method based on digital twinning according to claim 5, wherein, The S44 comprises: Extracting the clustering core features in the standardized verification data, performing quantitative conversion, and generating a clustering feature quantitative data set; Using the elbow rule combined with the contour coefficient method to evaluate the K value of the generated clustering feature quantitative data set, and generating a K-means optimal cluster number determination result; Inputting the clustering feature quantitative data set and the determined optimal K value into the K-means algorithm, initializing the cluster center, and through iterative calculation of the Euclidean distance of each sample to the cluster center, constantly updating the cluster center until convergence, and generating a preliminary scene clustering result; From the preliminary scene clustering result, filtering out clusters in which samples are mainly failed or have excessive response delay, and removing invalid clusters in which samples are mainly normal, and generating an effective weak scene cluster list; For the effective weak scene cluster list, extracting the feature commonality of each cluster one by one, and generating a weak scene cluster commonality feature analysis table; Integrating the weak scene cluster commonality feature analysis table, and finally generating a weak scene cluster analysis report of the foolproof strategy.

8. The flexible line AI prevention simulation verification method based on digital twinning according to claim 5, wherein, The K value evaluation of the generated clustering feature quantitative data set is specifically: testing different cases of K=2 to K=8, calculating the within-cluster sum of squares and the contour coefficient under each K value, and selecting the K value with the maximum contour coefficient at the inflection point of the WCSS curve.

9. The flexible line AI prevention simulation verification method based on digital twinning according to claim 1, wherein, The S5 comprises: S51, analyzing the generated robustness evaluation report, organizing suggestions to form an AI foolproof strategy optimization suggestion list; S52, updating the parameter configuration of the AI foolproof strategy according to the generated optimization suggestion list, and re-embedding the digital twin; selecting a typical scene in the identified weak scene cluster, performing closed-loop verification, and generating a foolproof strategy optimization effect verification report; S53, checking whether the key indicators in the generated optimization effect verification report meet the preset threshold: If not up to standard: return to step S2, supplement new abnormal cases based on the current weak scene, and re-execute the iteration process of S2-S5; If up to standard: go to the next step; S54, organizing the complete parameters of the verified AI foolproof strategy; packaging the parameters in a compatible format of the production line control system to generate an AI foolproof strategy configuration package that can be directly deployed to the real production line; S55, determining the robustness level according to the key indicator up-to-standard condition, and making a robustness level certificate; finally outputting the complete achievement of the AI foolproof strategy configuration package + robustness level certificate.

10. A flexible production line AI foolproof simulation verification system based on digital twinning, comprising: One or more processors; Memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1-9.

10. A flexible production line AI foolproof simulation verification system based on digital twinning, comprising: One or more processors; Memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1-9.

Images