Stamping die health prediction system based on digital twinning and historical data

Abstract

The application provides a stamping die health prediction system based on digital twinning and historical data, relates to the technical field of stamping dies, and comprises a data generation module for generating virtual degradation historical data of a die; a simulation unit based on a digital twinning model of the stamping die, configured to simulate a degradation index sequence of the die in a virtual running process according to input working condition parameters; a construction unit connected to the simulation unit and configured to form the virtual degradation historical data; a data training unit configured to use the simulation unit and the construction unit as a generator, use real historical data sequences as positive samples, perform adversarial training, optimize parameters of the simulation unit, and make the generated virtual degradation historical data conform to statistical characteristics of the real historical data; and a prediction module configured to predict a health state or residual life of the die based on real-time data and obtained by mixing training of virtual degradation historical data generated by the data generation module and real historical data.

Description

Technical Field

[0001] This invention belongs to the field of stamping die technology, specifically relating to a stamping die health prediction system based on digital twins and historical data. Background Technology

[0002] In the field of digital twin-based health prediction for stamping dies, the core challenge lies in how to construct a high-precision, highly reliable prediction model. Current methods mainly face two major technical bottlenecks:

[0003] First, data-driven model training relies on massive amounts of labeled historical data, especially data reflecting the complete degradation process of a mold from a healthy state to failure. In actual industrial scenarios, mold failure is a low-probability event, resulting in extremely scarce effective failure samples that can be collected, while normal operation data constitutes the vast majority. This small sample size and imbalanced data situation makes complex data-driven models, represented by deep learning, prone to overfitting or insensitive to early failures, making it difficult to achieve accurate remaining life prediction.

[0004] Secondly, while simulation models based on physical mechanisms can generate data, they are difficult to directly use for training data-driven models. Existing technologies often utilize simulation tools such as finite element analysis to generate stress, strain, and theoretical wear data for molds. However, this type of simulation data is generated under idealized assumptions and simplified models, and its data distribution differs significantly from historical data collected by real sensors (affected by multiple uncertainties such as noise, assembly errors, material batch fluctuations, and complex working conditions). Directly mixing simulation data with real historical data for training can lead to the model learning incorrect feature mappings, resulting in decreased performance in practical applications.

[0005] Therefore, how to effectively integrate limited, high-value real historical data with large-scale, physically meaningful simulation data, so that the two can complement each other and thus train a high-precision, highly robust health prediction model under data-scarce conditions, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] In view of the above-mentioned defects or deficiencies in the prior art, a stamping die health prediction system based on digital twin and historical data is provided, comprising:

[0007] The data generation module is configured to generate virtual degradation history data for stamping dies; including:

[0008] The simulation unit is built based on a digital twin model of the stamping die and is configured to simulate and output a sequence of degradation indicators of the die during virtual operation based on the input working parameters.

[0009] A construction unit, connected to the simulation unit, is configured to convert the degradation index sequence into a training feature sequence usable by the model to form virtual degradation history data;

[0010] A data training unit, connected to the construction unit, and pre-stores real historical data sequences; the data training unit is configured to: use the simulation unit and the construction unit as generators, and use the real historical data sequences as positive samples to perform adversarial training, so as to optimize the parameters of the simulation unit and make the generated virtual degraded historical data conform to the statistical characteristics of real historical data.

[0011] The prediction module is configured to predict the health status or remaining lifespan of the mold based on real-time data, and is obtained by mixing the virtual degradation history data generated by the data generation module with the real history data.

[0012] According to the technical solution provided in this application, the data training unit includes a discriminator, and the discriminator is used to perform adversarial training with the generator, wherein the real historical data sequence is used as the positive sample of the discriminator;

[0013] The data training unit is specifically configured for:

[0014] During the initialization phase, based on the limited sequence of real historical data, the initial parameters of the simulation unit are meta-optimized to minimize the difference between the generated initial virtual degradation historical data and the real historical data in a preset distribution metric.

[0015] During the adversarial training phase, the simulated unit and the constructed unit that have undergone meta-optimization are used as the initial state of the generator, and adversarial training is conducted with the discriminator.

[0016] According to the technical solution provided in this application, the digital twin model in the simulation unit is a parametric finite element model, and its specific configuration is used for:

[0017] The stress-strain field of the die during the stamping process is simulated, and based on the Achard wear model and the fatigue damage accumulation model, the stress-strain field is mapped to the wear depth increment and fatigue damage increment of the key area.

[0018] The construction unit is specifically configured for:

[0019] The wear depth increment and fatigue damage increment are fused with the corresponding simulated stress-strain field characteristics in a multidimensional way to construct a feature sequence that simultaneously contains macroscopic degradation index characteristics and microscopic mechanical state characteristics, which serves as the virtual degradation history data.

[0020] The optimization objective of the adversarial training includes making the micromechanical state characteristics in the generated virtual degraded historical data consistent in distribution with the mechanical state characteristics inferred from real historical data.

[0021] According to the technical solution provided in this application, the discriminator in the data training unit is a multi-task discriminator, which includes a mechanical feature inference sub-network and is configured to execute simultaneously:

[0022] The first discrimination task is to distinguish whether the input sequence comes from the virtual degenerate historical data of the generator or from the real historical data sequence.

[0023] The second inference task is to infer the corresponding mechanical state features from the input real historical data sequence through its internal feature extraction network.

[0024] The optimization objective of the adversarial training includes minimizing the difference between the micromechanical state features in the virtual degraded historical data generated by the generator and the mechanical state features inferred by the mechanical feature inference subnetwork from the corresponding positive sample real historical data sequence.

[0025] According to the technical solution provided in this application, the training of the mechanical feature inference subnetwork is achieved through the following steps:

[0026] The virtual degradation history data generated by the generator and its corresponding micromechanical state features calculated by the simulation unit are used as real labels to pre-train the mechanical feature inference subnetwork, so that it establishes an initial mapping relationship from sequence features to mechanical state features.

[0027] During the adversarial training process between the discriminator and the generator, the mechanical feature inference subnetwork is simultaneously optimized, and its optimization objectives include:

[0028] The mechanical feature inference subnetwork infers the mechanical state features of the real historical data sequence, and the overall distribution of these features is consistent with the distribution of the microscopic mechanical state features corresponding to the virtual degraded historical data.

[0029] The mechanical feature inference subnetwork infers the mechanical state features from the newly generated virtual degradation history data by the generator, and minimizes the difference between the micromechanical state features corresponding to the newly generated data calculated by the simulation unit.

[0030] According to the technical solution provided in this application, it also includes:

[0031] The on-site data acquisition and processing module is configured to acquire and process raw signals from the stamping production line to generate the real historical data sequence with time sequence and working condition labels.

[0032] The on-site data acquisition and processing module is configured for:

[0033] The synchronization signals acquired by multiple sensors are aligned using the stamping cycle signal as the timing reference.

[0034] Based on the production event tags obtained from the production line, continuous stamping production is divided into multiple working condition segments with stable working conditions.

[0035] Signal features are extracted from each of the aforementioned working condition segments and arranged according to the stamping time sequence to form the real historical data sequence, wherein each data point is associated with a corresponding stamping number and working condition label.

[0036] According to the technical solution provided in this application, it also includes:

[0037] The verification and correction module has a pre-set verification rule library based on the failure mechanism of stamping dies. The verification rule library includes geometric constraint rules, energy constraint rules and trend constraint rules.

[0038] The verification and correction module is connected to the data generation module and is configured to:

[0039] After the simulation unit completes a set number of simulated stamping operations, or after the construction unit generates a piece of virtual degradation history data, the verification rule base is called to verify the intermediate or final results.

[0040] If the verification fails, the input operating parameters of the simulation unit or the uncertainty parameters in the digital twin model are adjusted in reverse according to the triggered rule type, and the simulation is restarted from the starting point corresponding to this verification until the generated virtual degradation history data passes the verification of all relevant rules.

[0041] According to the technical solution provided in this application, the geometric constraint rule is: the depth of the simulated cutting edge wear morphology should not exceed the minimum cutting edge height in the mold design drawings;

[0042] The energy constraint rule is that the energy dissipated by a single press calculated from the simulated stress-strain field should be on the same order of magnitude as the approximate work done calculated by integrating the actual pressure curve of the corresponding working condition.

[0043] The trend constraint rule is as follows: under fixed working conditions, in continuous simulated stamping, the sequence of wear depth increments should satisfy a monotonically non-decreasing trend, and its fluctuations should conform to the stochastic statistical model of material wear.

[0044] According to the technical solution provided in this application, the on-site data acquisition and processing module further includes:

[0045] The identification and cleaning unit is configured for:

[0046] Real-time monitoring of changes in the production event tags; when a tag change is detected during a stamping process, the stamping is marked as a cross-condition stamping.

[0047] For each of the marked cross-condition strokes, perform at least one of the following operations:

[0048] Using the precise moment when the label change occurs as the boundary, the sensor signal across the working condition stroke is divided into two sub-strokes, and working condition labels before and after the change are assigned respectively.

[0049] If effective segmentation is not possible, all signal data across the operating conditions will be removed from the sequence of the current operating condition segment; the cleaned data will be used to regenerate or update the real historical data sequence.

[0050] According to the technical solution provided in this application, the prediction module is configured to: synchronously output a confidence evaluation value when performing prediction;

[0051] The credibility assessment value is calculated based on the degree of agreement between the features of the current real-time input data and the overall feature distribution of the virtual degradation historical data and the real historical data used to train the prediction module.

[0052] When the credibility assessment value is lower than a preset threshold, the system is configured to automatically trigger the data generation module to generate targeted supplementary virtual degradation historical data, focusing on the operating parameters corresponding to the current real-time data.

[0053] The prediction module uses the newly generated supplementary virtual degradation history data for incremental training.

[0054] Compared with the prior art, the beneficial effects of this application are as follows:

[0055] First, it fundamentally solves the model training challenge under small sample conditions: the system embeds a physical simulation model into a generative adversarial network framework, making it an intelligent data generator. By conducting adversarial training using a small amount of real historical data as anchors, the system can automatically and continuously generate virtual degraded historical data that is highly consistent with the statistical distribution of real data. This essentially constructs an inexhaustible and highly faithful training data source, completely breaking the dependence of model training on scarce fault historical data, and laying a solid foundation for building high-performance prediction models.

[0056] Second, it achieves a deep and effective fusion of physical simulation data and real historical data: Unlike simple data mixing, this application uses an adversarial training mechanism to force the parameters of the mechanistic simulation model (generator) to continuously optimize, so that the statistical characteristics of its output data (virtual degraded historical data) actively approximate the distribution of real data. This fusion method, which aligns simulation data with the real world, ensures that the generated virtual data retains both clear physical meaning and the statistical characteristics of real data, thereby enabling the subsequently trained prediction model to have both physical interpretability and real-world generalization ability.

[0057] Third, it significantly improves the accuracy and reliability of the health prediction model: The prediction module trained with the aforementioned high-quality mixed data learns more comprehensive and realistic degradation patterns. It can not only learn empirical patterns from real data, but also understand the underlying physical constraints from mechanistic data. Thus, it exhibits superior prediction accuracy, stronger robustness, and higher early warning reliability when facing unseen operating conditions or early, subtle signs of failure. Attached Figure Description

[0058] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0059] Figure 1 A schematic diagram of the structure of the stamping die health prediction system based on digital twin and historical data provided in this application.

[0060] The text labels in the image represent:

[0061] 1. Data generation module; 11. Simulation unit; 12. Construction unit; 13. Data training unit; 2. Prediction module. Detailed Implementation

[0062] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0063] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0064] As mentioned in the background section, this application proposes a stamping die health prediction system based on digital twins and historical data, such as... Figure 1 As shown, it includes:

[0065] Data generation module 1 is configured to generate virtual degradation history data for stamping dies; including:

[0066] Simulation unit 11 is constructed based on a digital twin model of the stamping die and is configured to simulate and output a sequence of degradation indicators of the die during virtual operation based on the input working parameters.

[0067] Construction unit 12, connected to the simulation unit 11, is configured to convert the degradation index sequence into a training feature sequence usable by the model to form virtual degradation history data;

[0068] The data training unit 13 is connected to the construction unit 12 and has a pre-stored real historical data sequence. The data training unit 13 is configured to: use the simulation unit 11 and the construction unit 12 as generators and the real historical data sequence as positive samples to perform adversarial training to optimize the parameters of the simulation unit 11 so that the generated virtual degraded historical data conforms to the statistical characteristics of real historical data.

[0069] Prediction module 2 is configured to predict the health status or remaining lifespan of the mold based on real-time data, and is obtained by mixing the virtual degradation history data generated by the data generation module 1 with the real history data for training.

[0070] Specifically, virtual degradation history data refers to a time-series data set generated by computer simulation, simulating the gradual performance degradation of a die during long-term use. It does not originate directly from physical sensors but aims to approximate the real degradation process from statistical and physical perspectives. The degradation index sequence refers to a series of key parameter values ​​reflecting the die's condition degradation, calculated and recorded in the virtual simulation at set time or stamping intervals. Examples include simulated cutting edge wear depth and accumulated fatigue damage in key areas. These parameters collectively describe the degradation trajectory. Simulation unit 11 is the physical knowledge engine of the entire system. Its implementation involves establishing a parameterized finite element analysis model based on the three-dimensional geometric model of the stamping die, material properties (such as hardness and yield strength), and stamping process parameters (such as stamping speed, clamping force, and sheet metal properties). This model can perform one or more mechanical simulations of the virtual stamping process based on the input working parameters. After each simulation, it doesn't just output a single-point result. Instead, based on a preset physical degradation model (such as a wear model), it calculates the degradation index increments of key parts of the die (such as the cutting edge of the blanking punch and the fillet radius of the drawing die), and arranges these increments in order of the number of simulated stampings to form a degradation index sequence. Construction unit 12 then plays a role in data conversion and encapsulation. It transforms the relatively raw degradation index sequence (which may contain time series of multiple physical quantities) output by simulation unit 11 into a more structured training feature sequence suitable for machine learning models through feature engineering techniques (e.g., calculating sliding window statistics, performing principal component analysis for dimensionality reduction, or directly splicing relevant features). This set of feature sequences constitutes the virtual degradation history data.

[0071] Specifically, the data training unit 13 achieves a deep integration of mechanism simulation and real data. During implementation, a small amount of real historical data sequences collected and processed from actual production lines need to be pre-stored. The core operation of this unit is to conduct a specific adversarial training: it treats the combination of simulation unit 11 and construction unit 12 as a generator, using the pre-stored real historical data sequences as positive samples. By constructing a discriminator (e.g., a deep learning classification network), the discriminator learns to distinguish whether the input data comes from virtual data generated by the generator or from real samples. During training, the generator aims to produce virtual data that is difficult for the discriminator to distinguish, while the discriminator strives to accurately differentiate. The gradient signal generated by this adversarial game process is backpropagated to simulation unit 11, specifically used to optimize the uncertain parameters or simplified empirical parameters in its digital twin model (e.g., empirical coefficients in the wear model, parameters that are difficult to determine precisely in the material model), rather than changing its core physical laws. Through this optimization, the virtual degradation history data generated by simulation unit 11 is driven to continuously approximate the real historical data in terms of overall statistical distribution characteristics (such as mean, variance, higher-order moments, and temporal patterns), thereby achieving realistic calibration of the physical simulation data. Prediction module 2 is the final application output of the system. It is essentially a trained machine learning model (e.g., a long short-term memory network, a temporal convolutional network, or a gradient boosting tree model). Its training data is a dataset composed of virtual degradation history data calibrated through adversarial training generated by data generation module 1, mixed with directly collected real historical data. In the online application phase, prediction module 2 receives real-time sensor data streams from the production line, performs feature extraction processing in the same manner as during training, and outputs an assessment of the current health status of the mold (e.g., healthy, alert, severe) or a quantitative prediction of its remaining service life (e.g., remaining safe stamping cycles).

[0072] The technical principle of this implementation lies in transforming the task of generating high-quality virtual data into an optimization problem where a physical simulation model (generator) fits the statistical distribution of real data using an adversarial training framework. This cleverly bypasses the difficulty of directly and accurately modeling complex degradation processes, instead pursuing statistical consistency. Its effect is that it can use a small amount of real data to guide the generation of massive amounts of high-quality labeled (known degradation levels) training data, fundamentally overcoming the limitations imposed on model training by small samples, especially the scarcity of fault samples, thus laying a reliable data foundation for subsequently building high-precision, highly robust prediction models.

[0073] In a preferred embodiment, the data training unit 13 includes a discriminator, which is used to perform adversarial training against the generator, wherein the real historical data sequence is used as a positive sample for the discriminator;

[0074] The data training unit 13 is specifically configured for:

[0075] During the initialization phase, based on the limited sequence of real historical data, the initial parameters of the simulation unit 11 are meta-optimized to minimize the difference between the generated initial virtual degradation historical data and the real historical data in a preset distribution metric.

[0076] During the adversarial training phase, the simulation unit 11 and the construction unit 12, which have undergone meta-optimization, are used as the initial state of the generator to conduct adversarial training with the discriminator.

[0077] Specifically, the discriminator here is the comparison component. In implementation, it is typically constructed as a neural network that takes a data sequence (whether virtual or real) as input and outputs a scalar value or a probability distribution representing the confidence level that the sequence is classified as real. During adversarial training, the discriminator aims to maximize its accuracy in distinguishing between real and generated sequences.

[0078] The initialization phase, also known as the meta-optimization phase, is a warm-up step performed before formal adversarial training begins. Its specific implementation is as follows: First, using a pre-stored, finite number of real historical data sequences, their statistical values ​​on certain preset distribution metrics are calculated. These distribution metrics can be the mean vector or covariance matrix of the overall data, or higher-level features such as latent space distribution features extracted by an autoencoder. Then, using the initial virtual degenerate historical data generated by simulation unit 11 (which has not yet engaged in adversarial training with the discriminator) as the object, the same distribution metrics are calculated. Next, using optimization algorithms such as gradient descent, the initial parameters of simulation unit 11 (these parameters also refer to empirical coefficients or uncertain parameters in the digital twin model) are adjusted to minimize the difference between the generated initial virtual data and the real data on the aforementioned distribution metrics (such as Euclidean distance and KL divergence). This process can be performed independently and quickly, without involving the training of the discriminator. The aim is to find a better starting point for simulation unit 11, so that it can generate virtual data that is roughly similar to the distribution of real data from the beginning, avoiding a long process of trial and error starting from a completely random initial distribution.

[0079] The adversarial training phase is the core optimization process. This phase uses the meta-optimized simulation unit 11 and construction unit 12 as the starting point for the generator, while the discriminator is also randomly initialized or starts from simple pre-training. Subsequently, the standard generative adversarial network (GAN) training process is initiated: the generator is fixed, and the discriminator is trained for several steps using real and generated data; then, the discriminator is fixed again, data is generated using the generator, and the realism score provided by the discriminator is used as the loss, with gradient backpropagation optimizing the parameters of simulation unit 11. This process is iterated repeatedly. Because the generator already has a good initial state, this adversarial training process converges more stably and finds simulation parameters that highly match the distribution of virtual data with the distribution of real data more efficiently.

[0080] The technical advantage of this implementation lies in its significant improvement of the dynamic process of adversarial training. The underlying principle is that meta-optimization provides strong prior guidance, reducing the complexity of the generator search space, minimizing potential pattern collapse (i.e., insufficient diversity in generated data) during training, accelerating the overall training convergence speed, and ultimately enabling the optimized simulation unit 11 parameters to generate more statistically high-quality and stable virtual degenerate historical data, thereby indirectly improving the performance ceiling and training efficiency of the final prediction model.

[0081] In a preferred embodiment, the digital twin model in the simulation unit 11 is a parametric finite element model, specifically configured for:

[0082] The stress-strain field of the die during the stamping process is simulated, and based on the Achard wear model and the fatigue damage accumulation model, the stress-strain field is mapped to the wear depth increment and fatigue damage increment of the key area.

[0083] The construction unit 12 is specifically configured for:

[0084] The wear depth increment and fatigue damage increment are fused with the corresponding simulated stress-strain field characteristics in a multidimensional way to construct a feature sequence that simultaneously contains macroscopic degradation index characteristics and microscopic mechanical state characteristics, which serves as the virtual degradation history data.

[0085] The optimization objective of the adversarial training includes making the micromechanical state characteristics in the generated virtual degraded historical data consistent in distribution with the mechanical state characteristics inferred from real historical data.

[0086] Specifically, simulation unit 11 is implemented using a parametric finite element model. This means that the mold mechanics model established in the computer, including its mesh generation, boundary conditions, and material constitutive relations, is designed to be flexibly adjustable through a set of parameters. The implementation steps are as follows: First, specific working parameters (such as sheet thickness, material grade, and stamping speed) are input. Then, the finite element solver calculates the stress and strain fields inside the mold during a single stamping process. Next, instead of simply outputting stress and strain values, it integrates and applies two classic physical failure models: the Achard wear model and the fatigue damage accumulation model. The Achard wear model calculates the volumetric wear of the material in the contact area between the mold and the sheet metal based on parameters such as contact stress and sliding distance, and then converts this into the wear depth increment of critical areas (such as the cutting edge). The fatigue damage accumulation model (such as one based on Miner's linear accumulation rule or a more complex nonlinear model) calculates the fatigue damage increment at critical danger points (such as stress concentration points) based on stress amplitude and average stress. Each virtual stamping operation calculates an increment, and by simulating multiple stamping operations, the wear depth increment sequence and fatigue damage increment sequence are obtained, which together constitute the degradation index sequence.

[0087] The implementation of construction unit 12 is further deepened. It no longer simply converts degradation indices, but performs multi-dimensional fusion operations. Specifically, it correlates and splices the two macroscopic degradation indices—wear depth increment and fatigue damage increment—with the root cause of these increments—the stress-strain field characteristics at the corresponding time points calculated in the simulation (e.g., the maximum contact stress at the cutting edge, the equivalent stress amplitude at the critical point, and the strain energy density during the stamping). The resulting feature sequence, in which the feature vector at each time point simultaneously contains both the result (macroscopic degradation indices) and the cause (microscopic mechanical state characteristics), constitutes higher-quality virtual degradation history data.

[0088] The optimization objective of adversarial training has therefore been upgraded. The goal is no longer merely matching overall statistical characteristics, but rather emphasizing that the distribution of microscopic mechanical state characteristics inherent in the generated virtual degradation history data must be consistent with the distribution of mechanical state characteristics inferred from real historical data. Inference refers to estimating the implicit mechanical state from sensor signals through data analysis methods (described later). This objective means that adversarial training must not only make the virtual data look like real data, but also ensure that the distribution patterns of the internal physical causes driving the virtual data are consistent with the physical laws indirectly reflected in the real world, thereby achieving a deeper, more mechanistic alignment between the virtual and real worlds.

[0089] The physically rich training data in this implementation allows the prediction model to learn not only phenomenological correlations but also potential causal relationships, thereby improving its generalization ability and robustness to abnormal operating conditions. Secondly, it provides a possible data foundation and comparison standard for subsequent, more refined adversarial training and physics-based verification.

[0090] In a preferred embodiment, the discriminator in the data training unit 13 is a multi-task discriminator, which includes a mechanical feature inference subnetwork and is configured to execute simultaneously:

[0091] The first discrimination task is to distinguish whether the input sequence comes from the virtual degenerate historical data of the generator or from the real historical data sequence.

[0092] The second inference task is to infer the corresponding mechanical state features from the input real historical data sequence through its internal feature extraction network.

[0093] The optimization objective of the adversarial training includes minimizing the difference between the micromechanical state features in the virtual degraded historical data generated by the generator and the mechanical state features inferred by the mechanical feature inference subnetwork from the corresponding positive sample real historical data sequence.

[0094] Specifically, the multi-task discriminator is a composite neural network model that shares a low-level feature extraction network but has multiple output task heads. In this embodiment, it specifically includes a mechanical feature inference sub-network. This sub-network can be several fully connected layers following the shared feature extraction layer, and its output dimension is consistent with the dimension of the microscopic mechanical state features of interest (such as equivalent stress and principal stress direction).

[0095] The discriminator is configured to perform two tasks simultaneously. The first discriminative task is the basic task, a binary classification task: given an input sequence (which may come from the generator or from real data), the discriminator needs to output a probability value representing its authenticity. Its implementation is the same as that of a standard GAN discriminator. The second inference task is an innovative task. The mechanical feature inference subnetwork is activated only when the discriminator identifies (or assumes) the input sequence as coming from real historical data, and attempts to infer the corresponding implicit micromechanical state feature sequence from this real sensor signal feature sequence. For example, given a real vibration and pressure signal sequence, the inference subnetwork outputs an estimated sequence of stress changes at key points of the mold within the same time period. The supervision signal for this task is indirect in the early stages of training, coming from the overall constraints of the adversarial process.

[0096] The optimization objective of adversarial training thus becomes specific and actionable. The objective requires that the microscopic mechanical state features directly calculated by simulation unit 11 in the virtual data generated by the generator be as close as possible to the mechanical state features inferred by the discriminator's mechanical feature inference subnetwork from corresponding (i.e., under the same or similar working conditions) real historical data positive samples. In practice, this introduces an additional term into the loss function: calculating the difference between the two (e.g., mean squared error), and using this difference as part of the loss, while simultaneously updating the inference subnetworks in both the generator (simulation unit 11) and the discriminator.

[0097] This approach involves a subtle game: the generator (simulation unit 11), in order to make the microscopic features in its data more closely resemble the discriminator's inferences about real data, adjusts its parameters to make the simulation more consistent with the discriminator's understanding of real-world physical laws. Meanwhile, the inference sub-network in the discriminator, in order to make more accurate inferences (which helps it better perform the task of distinguishing between true and false data), continuously improves the mapping relationship from real signals to mechanical features. The two mutually reinforce each other, ultimately making the distribution of microscopic mechanical features generated by the simulation model tend to be consistent with the feature distribution inferred from real data.

[0098] This implementation significantly enhances the physical reliability of virtual degradation history data, enabling predictive models trained on this data to incorporate stronger physical consistency. Simultaneously, the trained mechanical feature inference subnetwork itself can serve as a valuable byproduct, used to monitor implicit mechanical states from online data in real time, providing engineers with deeper insights. This represents a significant step in the evolution of digital twins from state synchronization to mechanism synchronization.

[0099] In a preferred embodiment, the training of the mechanical feature inference subnetwork is achieved through the following steps:

[0100] The virtual degradation history data generated by the generator and its corresponding micromechanical state features calculated by the simulation unit 11 are used as real labels to pre-train the mechanical feature inference subnetwork, so that it establishes an initial mapping relationship from sequence features to mechanical state features.

[0101] During the adversarial training process between the discriminator and the generator, the mechanical feature inference subnetwork is simultaneously optimized, and its optimization objectives include:

[0102] The mechanical feature inference subnetwork infers the mechanical state features of the real historical data sequence, and the overall distribution of these features is consistent with the distribution of the microscopic mechanical state features corresponding to the virtual degraded historical data.

[0103] The mechanical feature inference subnetwork infers the mechanical state features from the newly generated virtual degradation history data by the generator, and minimizes the difference between the microscopic mechanical state features corresponding to the newly generated data calculated by the simulation unit 11.

[0104] First, two key concepts need to be clarified. Self-supervised domain adaptation optimization refers to using data generated by the system itself (virtual data and its corresponding real mechanical labels) as supervision signals in the dynamic environment of adversarial training, enabling the network to adapt to the migration process from the virtual data domain to the real data domain. Its goal is to make the network's inference performance on real data match its performance on virtual data. Overall distribution consistency does not require perfect point-to-point consistency, but rather means that the inferred features are similar to the reference distribution in the main statistical properties of the probability distribution (such as mean, variance, and quantiles).

[0105] The training process is implemented in two steps: Step 1: Supervised pre-training. This step aims to provide a high-quality initial state for the inference sub-network. Specifically, in the initial stage of adversarial training or before, the data generation module 1 (generator) is run to generate a large amount of virtual degradation history data. For each generated virtual data segment, the system simultaneously possesses two pieces of information: one is the training feature sequence output by the construction unit 12 (composed of sensor simulated signal features and macroscopic degradation indicators, etc.), and the other is the microscopic mechanical state feature sequence (e.g., equivalent stress value of key points under each simulated impulse) calculated and recorded synchronously by the simulation unit 11 when generating the data, serving as the real label. Using this batch of perfectly paired data (feature sequence - mechanical label sequence), the mechanical feature inference sub-network is trained using traditional supervised learning methods (such as minimizing mean square error loss). The core purpose of this step is to enable the sub-network to learn to accurately reconstruct the known physical state from the sequence closely related to the virtual sensor features, establishing a preliminary mapping relationship based on the simulated world. Step 2: Self-supervised optimization in adversarial training. This step is crucial for training and aims to transfer the subnetwork's ability from understanding virtual data to understanding real data. In the main loop of adversarial training between the generator and discriminator (which includes the inference subnetwork), the parameters of the mechanical feature inference subnetwork are updated synchronously, but its optimization objective differs from pre-training, incorporating two cleverly designed sub-objectives.

[0106] Specifically, sub-objective a is the distribution consistency objective. Its implementation is as follows: During training, a batch of real historical data sequences is periodically sampled and input into the inference sub-network to obtain a batch of inferred mechanical state features; simultaneously, a batch of corresponding real micromechanical state features, calculated by simulation unit 11, is sampled from pre-trained or historically generated virtual data. Then, the overall distribution difference between these two batches of features in high-dimensional space is calculated, for example, using the maximum mean difference (MMD) or calculating the Wasserstein distance through an auxiliary distribution discriminator. The optimization objective is to minimize this distribution difference. This forces the inference sub-network's inference results for real data to converge, in terms of overall statistical regularity, towards the reasonable distribution defined by the digital twin model learned from the virtual data, ensuring that the inference results do not violate fundamental physical laws. Sub-objective b is the self-supervised reconstruction objective. Its implementation is as follows: Each time the generator generates a new batch of virtual degenerate historical data, the system also possesses the feature sequence of this new data and the corresponding real labels of the micromechanical state features. The feature sequence of this new data is also input into the inference sub-network to obtain another set of inferred mechanical features. Then, the difference (e.g., mean squared error) between this set of inferred features and the known true labels is calculated. The optimization objective is to minimize this difference. This objective acts like a calibrator or anchor, continuously and online fine-tuning the inference subnetwork using new virtual data with defined labels that are constantly generated within the system. This prevents the subnetwork from forgetting or deviating from the correct direction while adapting to the distribution of real data, ensuring that the accuracy of its inference ability is maintained and improved during the dynamic process of adversarial training.

[0107] In a preferred embodiment, it further includes:

[0108] The on-site data acquisition and processing module is configured to acquire and process raw signals from the stamping production line to generate the real historical data sequence with time sequence and working condition labels.

[0109] The on-site data acquisition and processing module is configured for:

[0110] The synchronization signals acquired by multiple sensors are aligned using the stamping cycle signal as the timing reference.

[0111] Based on the production event tags obtained from the production line, continuous stamping production is divided into multiple working condition segments with stable working conditions.

[0112] Signal features are extracted from each of the aforementioned working condition segments and arranged according to the stamping time sequence to form the real historical data sequence, wherein each data point is associated with a corresponding stamping number and working condition label.

[0113] Specifically, the stamping cycle signal typically refers to a pulse signal generated when the press slide reaches the top dead center. It is a precise, periodic mechanical event marker, providing a unified physical time anchor for all sensor data. Production event tags refer to information obtained from the production line control system (such as MES, PLC) that describes the context of each production session or batch, such as the specific sheet metal model used (sheet metal identifier), the type and amount of lubricant applied (lubrication status), and the production order number (production batch number). A production segment refers to a continuous period of production time during which all relevant production event tags remain unchanged, indicating stable production conditions.

[0114] The specific implementation of this module includes the following steps:

[0115] First, a multi-source sensor array is deployed and signals are acquired synchronously. High-dynamic pressure sensors are installed in the main cylinder and balance cylinder of the press, triaxial accelerometers are installed in key areas such as the mold base and upper mold base to capture three-dimensional vibrations, and temperature sensors are installed near the mold cavity. All these sensor data acquisition cards are synchronized using the same clock source or hardware triggering to ensure underlying alignment of data timestamps. Simultaneously, production event change signals from the PLC are monitored and captured in real time through a data acquisition system or industrial gateway.

[0116] Secondly, timing alignment is performed. Using the top dead center pulse signal acquired by the hardware as a reference, all sensor signals (pressure curves, vibration waveforms, temperature values) acquired within each stamping cycle are aligned. Specifically, the time point corresponding to each top dead center pulse is identified, and using this as the zero point, all signal data within a specific time window before and after that cycle are extracted to form a single stamping data packet. In this way, signals of different physical quantities and different sampling rates are precisely aligned within the same mechanical cycle, eliminating phase errors caused by sensor response delays and sampling asynchrony.

[0117] Next, the system segments production conditions based on production event tags. The system continuously monitors the production event tag flow. When any tag (such as a sheet metal identifier) ​​changes, it is marked as a production condition switching point. The system categorizes all strokes between two consecutive switching points into the same production condition segment. For example, continuously producing 10,000 strokes of a certain model of automotive door panel lining (sheet metal identifier A, lubricant B) constitutes one production condition segment. When switching to producing another model of part (sheet metal identifier C), a new segment is started. This ensures that the data within each segment is generated under almost identical process conditions, resulting in high data consistency and facilitating the model's learning of degradation patterns under specific production conditions.

[0118] Next, feature extraction and sequence construction are performed. For each aligned stroke data packet within each working condition segment, a set of predefined time-domain, frequency-domain, and time-frequency-domain features are extracted from the signals of each sensor. Time-domain features may include peak value, mean, root mean square, waveform factor, impulse factor, etc.; frequency-domain features can be calculated using Fast Fourier Transform (FFT) to obtain the dominant frequency, centroid frequency, frequency variance, etc.; time-frequency-domain features can be extracted using wavelet transform to extract the energy of a specific frequency band. All features of a stroke are concatenated into a feature vector. Finally, according to the actual order of the stamping (usually the chronological order), the feature vectors corresponding to all strokes within the working condition segment are arranged to form a two-dimensional matrix (time step × feature dimension), which is a real historical data sequence. Each data point (i.e., each row) in this sequence is implicitly or explicitly associated with the cumulative number of stampings since the start of production and the unified working condition label of that segment.

[0119] This implementation improves data quality and consistency, providing a reliable real-world benchmark for subsequent model training and virtual data alignment. Secondly, this structured approach enables the system to distinguish degradation patterns under different operating conditions, laying the foundation for more refined predictions.

[0120] In a preferred embodiment, it further includes:

[0121] The verification and correction module has a pre-set verification rule library based on the failure mechanism of stamping dies. The verification rule library includes geometric constraint rules, energy constraint rules and trend constraint rules.

[0122] The verification and correction module is connected to the data generation module 1 and is configured to:

[0123] After the simulation unit 11 completes a set number of simulated stampings, or after the construction unit 12 generates a piece of virtual degradation history data, the verification rule base is called to verify the intermediate or final results.

[0124] If the verification fails, the input operating parameters of the simulation unit 11 or the uncertainty parameters in the digital twin model are adjusted in reverse according to the triggered rule type, and the simulation is restarted from the starting point corresponding to this verification until the generated virtual degradation history data passes the verification of all relevant rules.

[0125] Specifically, failure mechanisms refer to the physical and chemical processes that lead to the loss of performance in stamping dies, such as wear failure, fatigue failure, and plastic deformation failure. Uncertainty parameters refer to those parameters in the digital twin model that are difficult to determine precisely through theory and usually rely on experimental or empirical calibration, such as the wear coefficient in the Archard wear model and the dispersion parameter in the material fatigue SN curve. The implementation of this module is based on a pre-set verification rule library. The module's workflow is as follows: The system sets checkpoints for the data generation module 1 (especially simulation unit 11). For example, a verification can be triggered every N virtual stamping simulations (e.g., N=1000), or whenever the construction unit 12 prepares a segment of virtual degradation history data of length L. At this time, the verification and correction module is called. The module selects relevant rules from the rule library for application according to the current simulation stage. It reads the intermediate or final results that need to be verified. This may include: the three-dimensional topography map of the cumulative wear depth up to the current checkpoint, the fatigue damage distribution cloud map, historical stress-strain data, the calculated value of cumulative dissipated energy, the wear depth increment sequence, etc. Then, the module applies the selected rules one by one for judgment. For example, applying geometric constraint rules: The minimum allowable cutting edge height (a known constant) defined in the mold design drawings is read and compared with the maximum cutting edge wear depth obtained from the current simulation. Applying energy constraint rules: A real historical data segment corresponding to the current simulation condition is read, a typical pressure curve is extracted from this segment, and its integral is calculated as an approximate work done; simultaneously, the energy consumed by plastic deformation inside the mold (dissipated energy) under the same simulated stroke is calculated from the simulation results. Applying trend constraint rules: The wear depth increment sequence generated by several recent consecutive simulated stampings is analyzed. If all relevant rules pass the verification, the data generation process continues. If a rule fails the verification (e.g., the simulated wear depth has exceeded the design limit), the module will trigger different correction strategies based on the rule type. For failures triggered by geometric constraints, the input parameters of the simulation may be adjusted in reverse, for example, prompting "The current simulated stamping conditions are too harsh," and automatically reducing the simulated sheet thickness or strength parameters before restarting the simulation from the previous checkpoint. For failures triggered by energy constraints, the uncertainty parameters in the digital twin model may be adjusted in reverse, for example, fine-tuning the empirical coefficients of the wear model, and then recalculating the most recent segment of the simulation. This process is repeated until the data for the current checkpoint passes all rule checks. In this way, the module dynamically constrains and guides the simulation process, ensuring that the generated data is not only statistically sound but also physically reliable.

[0126] The technical problem addressed by this implementation is that relying solely on data statistical feature alignment (adversarial training) may occasionally generate virtual data that is physically absurd or engineering impossible (e.g., wear depth exceeding the mold body). This module uses a configurable rule base to perform online verification and intervention on the simulation process, ensuring that the generated virtual degradation history data is reasonable and reliable at both the physical and engineering levels.

[0127] In a preferred embodiment, the geometric constraint rule is: the depth of the simulated cutting edge wear morphology should not exceed the minimum cutting edge height in the mold design drawings;

[0128] The energy constraint rule is that the energy dissipated by a single press calculated from the simulated stress-strain field should be on the same order of magnitude as the approximate work done calculated by integrating the actual pressure curve of the corresponding working condition.

[0129] The trend constraint rule is as follows: under fixed working conditions, in continuous simulated stamping, the sequence of wear depth increments should satisfy a monotonically non-decreasing trend, and its fluctuations should conform to the stochastic statistical model of material wear.

[0130] Specifically, the implementation of geometric constraint rules: The core of these rules is to ensure that mold deformation or wear in the simulation does not exceed the limits of its physical structure. During implementation, the design allowable values ​​for key mold dimensions need to be digitized beforehand. For example, during the system initialization phase, import the mold CAD drawing or directly input "initial cutting edge height H". initial "and "Minimum remaining height H of the cutting edge after wear" min Parameters such as "..." are used. During the simulation, simulation unit 11 outputs the three-dimensional node coordinate changes of the mold surface, from which the current height distribution of the cutting edge line can be calculated, and the minimum value H can be found. current The rule is defined as: if (H initial - H current )>(H initial - H min If the current wear depth exceeds the maximum allowable wear amount, the verification fails. This is a hard constraint that is directly related to the design safety boundary.

[0131] Specific implementation of the energy constraint rule: This rule is used to verify the rationality of simulation and reality at the macroscopic energy level. Implementation consists of two steps: First, estimation of the actual work done. Several typical and complete pressure-slider displacement (or time) curves are selected from real historical data segments corresponding to the working condition. By numerically integrating the pressure curve (∫P ds or ∫P dt), the approximate mechanical work W required to complete one stamping operation is calculated. realThe first step is to calculate the average or range of the energy dissipation. The second step is to simulate and calculate the dissipated energy. At the end of a single stamping operation under the same conditions, the finite element solver can output the internal energy changes, plastic deformation energy, and frictional dissipation energy of the die and sheet metal system. The sum of the main irreversible dissipated energy components is extracted and denoted as W. sim The rule is defined as: W sim With W real They should be on the same order of magnitude (e.g., 0.1W) real <W sim <10W real If W sim Much smaller than W real This might mean the simulation model is too idealistic, neglecting sufficient frictional or plastic work; if it is much greater than W... real This could mean that the material model or parameter settings are inappropriate, resulting in excessive energy consumption during the simulation. This is a soft constraint used to capture significant deviations in magnitude.

[0132] Specific Implementation of the Trend Constraint Rule: This rule is used to verify the rationality of the degradation process evolution. It focuses on the pattern of change of the virtual degradation index sequence (such as wear depth increment Δd) over time (number of stampings). During implementation, for continuous simulated stamping under a fixed working condition, the wear depth increment for each stamping is recorded, forming a sequence {Δd1, Δd2,..., Δdn}. The rule consists of two parts: First, a monotonicity check. Since wear is cumulative, under ideal conditions of uniform material and constant working conditions, the wear increment caused by each stamping should be non-negative. Therefore, it checks whether all Δdi ≥ 0 (or allows for extremely small negative values ​​due to numerical computation noise). Second, a volatility statistical check. The actual wear process is random, and the increment will fluctuate around a mean. The system can predefine or learn a simple stochastic statistical model from initial real data, for example, assuming Δd follows a mean of μ and a variance of σ. 2 The sequence follows a normal distribution. During simulation, the mean and variance of the most recent incremental sequence are periodically checked (e.g., every 1000 simulations) to ensure they match the expected model (or the theoretical range derived from the digital twin model). For example, it is checked whether the sample mean is within a reasonable confidence interval of μ and whether the sample variance is abnormally large. This rule ensures that the evolution trend of the simulation degradation conforms to physical common sense (no decrease) and has reasonable random fluctuation characteristics, avoiding the generation of unrealistic sequences that are too smooth or oscillating violently.

[0133] In a preferred embodiment, the on-site data acquisition and processing module further includes:

[0134] The identification and cleaning unit is configured for:

[0135] Real-time monitoring of changes in the production event tags; when a tag change is detected during a stamping process, the stamping is marked as a cross-condition stamping.

[0136] For each of the marked cross-condition strokes, perform at least one of the following operations:

[0137] Using the precise moment when the label change occurs as the boundary, the sensor signal across the working condition stroke is divided into two sub-strokes, and working condition labels before and after the change are assigned respectively.

[0138] If effective segmentation is not possible, all signal data across the operating conditions will be removed from the sequence of the current operating condition segment; the cleaned data will be used to regenerate or update the real historical data sequence.

[0139] First, there's the real-time identification of cross-condition strokes. The identification and cleaning unit continuously monitors the digital signal stream of production event tags (such as sheet metal codes and lubricant valve opening signals) from the production line control system. Simultaneously, it receives the physical top dead center (TDC) pulse signal from the press, which precisely marks the start time of each stroke cycle. Internally, the unit maintains a state machine whose basic logic is: the time interval between one TDC pulse and the next TDC pulse is defined as a physical stroke. The unit caches the set of condition tags that were active at the previous TDC moment. When a change in the value of any production event tag is detected within a new stroke cycle (e.g., the sheet metal code is updated midway through the stroke), the unit immediately records the precise timestamp (t) of this change event. change Since the label change occurred within the start and end time of the stamping operation, the system marked the stamping operation as a cross-condition stamping operation and recorded the label status before and after the change, as well as the time point t when the change occurred. change .

[0140] Secondly, there is a data cleaning strategy for identified cross-condition strokes. The cleaning unit offers two configurable processing methods, which are typically selected or used in combination based on signal characteristics and engineering requirements.

[0141] The first operation is data segmentation. This operation is suitable for situations where the sensor signal sampling rate is high and can clearly reflect different stages of the stamping process. The specific steps are: using the recorded working condition switching time point t... change Using t as the boundary, all synchronously acquired sensor raw signals (pressure curves, vibration waveforms, etc.) belonging to this cross-condition stroke are segmented. change It's possible that at any given sampling time, a signal interpolation algorithm (such as linear interpolation or spline interpolation) needs to be used at time t. change Virtual signal segmentation points are constructed at each moment. Then, the first half of the signal of that impulse (from the top dead center to t) is...change The signal is attributed to the original operating condition label, forming a sub-stroke; the latter half of the signal (from t) is assigned to the original operating condition label. change (Up to the next top dead center) it belongs to the changed operating condition label, forming another sub-stroke. These two sub-strokes will be sent to the subsequent feature extraction process and classified into the corresponding operating condition segments. This processing method preserves the data to the maximum extent, but requires the signal to be separable before and after the switching point.

[0142] The second operation is data rejection. This conservative strategy is adopted when the operating condition switch is very complex, or when the signal is severely aliased near the switch point, making reliable segmentation impossible. The system will remove all raw sensor data corresponding to the cross-operating condition stroke from the current acquisition stream, and these data will not participate in any subsequent feature extraction and sequence construction. Simultaneously, the system will record this rejection event and may trigger an alarm, indicating an anomaly in the production cycle or control logic. To maintain the continuity of the data sequence, the system will skip the counting of this stroke when constructing the historical data sequence for this operating condition segment.

[0143] Finally, after the above cleaning process, both the segmented sub-stroke data and the clean data retained after removing abnormal strokes will be used to regenerate or dynamically update the real historical data sequence, ensuring that the working condition label associated with each data point in the sequence is pure and accurate.

[0144] This unit aims to address the problem that production condition switching does not always occur during the intervals between stamping operations; it may be triggered or completed within a single stamping stroke (e.g., the production line control system receives and executes a sheet metal change procedure or pressure adjustment command during slide movement). This results in a single stamping stroke's sensor signal simultaneously carrying characteristic information from two different conditions. If this cross-condition stamping stroke data is mixed into subsequent historical data sequences built for a single condition, it becomes a source of noise and interference, contaminating the dataset and severely impacting the model's accuracy in learning degradation patterns under specific conditions. This implementation aims to accurately identify and properly handle such abnormal data points. Its implementation relies on millisecond-level parallel monitoring of the production event tag stream and the stamping mechanical cycle.

[0145] In a preferred embodiment, the prediction module 2 is configured to: synchronously output a confidence assessment value when performing a prediction;

[0146] The credibility assessment value is calculated based on the degree of agreement between the features of the current real-time input data and the overall feature distribution of the virtual degradation historical data and the real historical data used to train the prediction module 2.

[0147] When the credibility assessment value is lower than a preset threshold, the system is configured to automatically trigger the data generation module 1 to generate targeted supplementary virtual degradation historical data, focusing on the operating parameters corresponding to the current real-time data.

[0148] The prediction module 2 uses the newly generated supplementary virtual degradation history data for incremental training.

[0149] Specifically, the first step is the calculation of the credibility assessment value. Prediction module 2 is designed with two outputs: one is a standard health status or remaining life expectancy prediction value (y). pred ), and secondly, a scalar confidence assessment value (c) score (The range is from 0 to 1). Calculate c. score The core idea is to measure the features (x) of the current real-time input data. current The distance or membership degree between the real-time data and the feature distribution of all data seen during the training of prediction module 2 is used. One specific implementation is based on feature space density estimation. In the offline phase, the system extracts features from all virtual degraded historical data and real historical data (i.e., a mixed training set) used to train prediction module 2, forming a high-dimensional feature space point cloud. The density distribution of this feature space can be modeled using kernel density estimation (KDE) or by training a simple single-class classifier (such as a single-class SVM). During online prediction, the real-time data x is used... current After the same feature extraction, the input is fed into this density estimation model. The output probability density value, or the distance to the decision boundary, after normalization, can be used as c. score Higher density or closer distance indicates that the current state is closer to the core area covered by the training data, and the credibility is higher; conversely, it indicates that it is on the edge of the distribution or in a sparse area, and the credibility is low.

[0150] Secondly, there is the low-reliability triggering mechanism. The system presets a threshold (c) threshold (e.g., 0.2). When prediction module 2 makes predictions on real-time data for a single instance or multiple consecutive instances, the calculated c... score Persistently below c threshold If the pattern in the current input data is not found, it is considered a valid low-confidence event. This indicates that the pattern in the current input data is a blind spot or weak point in the system's understanding.

[0151] Next, targeted data generation and incremental training are triggered. Once a low-confidence event is confirmed, the system automatically sends a request to the data generation module 1. The core parameter of the request is the operating condition parameter corresponding to the current real-time data. These parameters can be obtained from the production event tags associated with the real-time data (such as the current plate material and lubrication status) or deduced from signals (such as the estimated stamping tonnage). After receiving the request, the data generation module 1 uses these operating condition parameters as the key input to drive the simulation unit 11 to perform high-intensity virtual simulation focused on the combination of these parameters. For example, centered on the aforementioned low-confidence operating condition parameters, sampling is performed within a small neighborhood of its parameter space to generate a large number (e.g., thousands) of highly correlated supplementary virtual degradation historical data. This newly generated data is specifically targeted at operating conditions where the system exhibits uncertainty, and has strong relevance and supplementation.

[0152] Finally, incremental training is performed. Prediction module 2 is not immediately retrained entirely with the new data; instead, it employs an incremental learning (or online learning) strategy. It mixes newly generated supplementary dummy data with a small portion of the original training data to form a fine-tuned dataset. Then, it is trained for a few epochs based on the existing prediction model weights. This process can be performed periodically or automatically when the system is idle. Through this incremental training, the prediction model quickly absorbs new knowledge about previous blind spots, updates its internal parameters, and thus can make more reliable predictions when encountering similar operating conditions in the future.

[0153] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A stamping die health prediction system based on digital twin and historical data, characterized in that, include: The data generation module (1) is configured to generate virtual degradation history data for stamping dies; including: The simulation unit (11) is built based on the digital twin model of the stamping die and is configured to simulate the degradation index sequence of the die during virtual operation according to the input working parameters. Construction unit (12), connected to the simulation unit (11), is configured to convert the degradation index sequence into a training feature sequence usable by the model to form virtual degradation history data; The data training unit (13) is connected to the construction unit (12) and has a pre-stored real historical data sequence. The data training unit (13) is configured to: use the simulation unit (11) and the construction unit (12) as generators and the real historical data sequence as positive samples to perform adversarial training in order to optimize the parameters of the simulation unit (11) so that the generated virtual degraded historical data conforms to the statistical characteristics of real historical data. The prediction module (2) is configured to predict the health status or remaining lifespan of the mold based on real-time data, and is obtained by mixing the virtual degradation history data generated by the data generation module (1) with the real history data for training. The discriminator in the data training unit (13) is a multi-task discriminator, which includes a mechanical feature inference subnetwork and is configured to execute simultaneously: The first discrimination task is to distinguish whether the input sequence comes from the virtual degenerate historical data of the generator or from the real historical data sequence. The second inference task is to infer the corresponding mechanical state features from the input real historical data sequence through its internal feature extraction network. The optimization objective of the adversarial training includes minimizing the difference between the micromechanical state features in the virtual degraded historical data generated by the generator and the mechanical state features inferred by the mechanical feature inference subnetwork from the corresponding positive sample real historical data sequence. The training of the mechanical feature inference subnetwork is achieved through the following steps: Using the virtual degradation history data generated by the generator and its corresponding micromechanical state features calculated by the simulation unit (11) as real labels, the mechanical feature inference subnetwork is pre-trained to establish an initial mapping relationship from sequence features to mechanical state features; During the adversarial training process between the discriminator and the generator, the mechanical feature inference subnetwork is simultaneously optimized, and its optimization objectives include: The mechanical feature inference subnetwork infers the mechanical state features of the real historical data sequence, and the overall distribution of these features is consistent with the distribution of the microscopic mechanical state features corresponding to the virtual degraded historical data. The mechanical feature inference subnetwork infers the mechanical state features from the newly generated virtual degradation history data by the generator, and minimizes the difference between the mechanical state features corresponding to the newly generated data calculated by the simulation unit (11).

2. The stamping die health prediction system based on digital twin and historical data according to claim 1, characterized in that, The data training unit (13) includes a discriminator, which is used to perform adversarial training with the generator, wherein the real historical data sequence is used as the positive sample of the discriminator; The data training unit (13) is specifically configured for: In the initialization phase, based on the limited real historical data sequence, the initial parameters of the simulation unit (11) are meta-optimized so that the initial virtual degraded historical data generated by it has the smallest difference from the real historical data in the preset distribution metric. During the adversarial training phase, the simulated unit (11) and the construction unit (12) after meta-optimization are used as the initial state of the generator, and adversarial training is performed with the discriminator.

3. The stamping die health prediction system based on digital twin and historical data according to claim 1, characterized in that, The digital twin model in the simulation unit (11) is a parametric finite element model, and its specific configuration is used for: The stress-strain field of the die during the stamping process is simulated, and based on the Achard wear model and the fatigue damage accumulation model, the stress-strain field is mapped to the wear depth increment and fatigue damage increment of the key area. The construction unit (12) is specifically configured for: The wear depth increment and fatigue damage increment are fused with the corresponding simulated stress-strain field characteristics in a multidimensional way to construct a feature sequence that simultaneously contains macroscopic degradation index characteristics and microscopic mechanical state characteristics, which serves as the virtual degradation history data. The optimization objective of the adversarial training includes making the micromechanical state characteristics in the generated virtual degraded historical data consistent in distribution with the mechanical state characteristics inferred from real historical data.

4. The stamping die health prediction system based on digital twin and historical data according to claim 1, characterized in that, Also includes: The on-site data acquisition and processing module is configured to acquire and process raw signals from the stamping production line to generate the real historical data sequence with time sequence and working condition labels. The on-site data acquisition and processing module is configured for: The synchronization signals acquired by multiple sensors are aligned using the stamping cycle signal as the timing reference. Based on the production event tags obtained from the production line, continuous stamping production is divided into multiple working condition segments with stable working conditions. Signal features are extracted from each of the aforementioned working condition segments and arranged according to the stamping time sequence to form the real historical data sequence, wherein each data point is associated with a corresponding stamping number and working condition label.

5. The stamping die health prediction system based on digital twin and historical data according to claim 1, characterized in that, Also includes: The verification and correction module has a pre-set verification rule library based on the failure mechanism of stamping dies. The verification rule library includes geometric constraint rules, energy constraint rules and trend constraint rules. The verification and correction module is connected to the data generation module (1) and is configured for: After the simulation unit (11) completes a set number of simulated stampings, or after the construction unit (12) generates a virtual degradation history data, the verification rule base is called to verify the intermediate or final results. If the verification fails, the input operating parameters of the simulation unit (11) or the uncertainty parameters in the digital twin model are adjusted in reverse according to the triggered rule type, and the simulation is restarted from the starting point corresponding to this verification until the generated virtual degradation history data passes the verification of all relevant rules.

6. The stamping die health prediction system based on digital twin and historical data according to claim 5, characterized in that, The geometric constraint rule is: the depth of the simulated cutting edge wear morphology should not exceed the minimum cutting edge height in the mold design drawings; The energy constraint rule is that the energy dissipated by a single press calculated from the simulated stress-strain field should be on the same order of magnitude as the approximate work done calculated by integrating the actual pressure curve of the corresponding working condition. The trend constraint rule is as follows: under fixed working conditions, in continuous simulated stamping, the sequence of wear depth increments should satisfy a monotonically non-decreasing trend, and its fluctuations should conform to the stochastic statistical model of material wear.

7. The stamping die health prediction system based on digital twin and historical data according to claim 4, characterized in that, The on-site data acquisition and processing module also includes: The identification and cleaning unit is configured for: Real-time monitoring of changes in the production event tags; when a tag change is detected during a stamping process, the stamping is marked as a cross-condition stamping. For each of the marked cross-condition strokes, perform at least one of the following operations: Using the precise moment when the label change occurs as the boundary, the sensor signal across the working condition stroke is divided into two sub-strokes, and working condition labels before and after the change are assigned respectively. If effective segmentation is not possible, all signal data across the operating conditions will be removed from the sequence of the current operating condition segment; the cleaned data will be used to regenerate or update the real historical data sequence.

8. The stamping die health prediction system based on digital twin and historical data according to claim 4, characterized in that, The prediction module (2) is configured to: synchronously output a confidence assessment value when performing a prediction; The credibility assessment value is calculated based on the characteristics of the current real-time input data and the degree of consistency between the overall feature distribution of the virtual degraded historical data and the real historical data used to train the prediction module (2); When the credibility assessment value is lower than the preset threshold, the system is configured to automatically trigger the data generation module (1) to generate targeted supplementary virtual degradation historical data, focusing on the working condition parameters corresponding to the current real-time data. The prediction module (2) uses the newly generated supplementary virtual degradation history data for incremental training.

Images