Automobile forging defect prediction system based on digital twinning

Abstract

This application relates to the field of defect prediction technology, and more particularly to a defect prediction system for automotive forgings based on digital twins. The system includes a model building module, a signal acquisition module, a probability prediction module, and a defect assessment module. The model building module acquires the three-dimensional appearance data of the forging and constructs a virtual model and a Green's function library based on material properties. The signal acquisition module applies narrow-band frequency sweep micro-excitation to the forging at a preset process phase during the forging process and acquires the response signal. The probability prediction module performs time-varying registration of the response signal with the Green's function library and reconstructs the internal defect probability voxel map and defect confidence interval based on an inversion algorithm with sparse constraints. The defect assessment module simulates the defect propagation trend using the virtual model and outputs the defect risk density. This application enables spatial localization and risk quantification assessment of latent defects in the early stages of the forging process, improving forging quality control and service reliability.

Description

Technical Field

[0001] This application relates to the field of defect prediction technology, and in particular to a defect prediction system for automotive forgings based on digital twins. Background Technology

[0002] Current defect detection and quality control of automotive forgings mainly rely on traditional non-destructive testing methods such as ultrasonic testing, radiographic testing, or surface imaging. These methods are widely used in the pre-service inspection of forgings, but their results often have limitations. Specifically: in forgings with complex geometries, especially long cylindrical components like valve tappets, latent defects such as shrinkage cavities, cracks, and inclusions often cannot be accurately identified in the early stages, leading to missed detections and misjudgments; regarding the evolution trend of defects under dynamic service conditions, existing technologies can only provide static detection results, and defects may continue to expand in subsequent service stages, posing potential failure risks; at the quality assessment level, it is difficult to directly correlate inspection data with forging process parameters, resulting in a lack of foresight in forging reliability assessment and insufficient product consistency and lifespan assurance.

[0003] To address the above issues, this application presents a digital twin-based defect prediction system for automotive forgings. Summary of the Invention

[0004] The technical problem this application aims to solve is to address the shortcomings of existing technologies by providing a digital twin-based defect prediction system for automotive forgings. The system includes a model building module, a signal acquisition module, a probability prediction module, and a defect assessment module. The model building module acquires the three-dimensional appearance data of the forging and combines it with material properties to construct a virtual model and a Green's function library. The signal acquisition module applies narrow-band frequency sweep micro-excitation to the forging at a preset process phase during the forging process and acquires the response signal. The probability prediction module performs time-varying registration of the response signal with the Green's function library and reconstructs the internal defect probability voxel map and defect confidence interval based on an inversion algorithm with sparse constraints. The defect assessment module simulates the defect propagation trend using the virtual model and outputs the defect risk density.

[0005] To achieve the above objectives, this application provides the following technical solution:

[0006] A digital twin-based defect prediction system for automotive forgings, comprising:

[0007] The model building module is used to acquire the appearance data of automotive forgings and to build a virtual model and Green's function library based on the appearance data.

[0008] The signal acquisition module is used to perform narrow-band frequency sweep micro-excitation on the automotive forging when the forging process of the automotive forging is in a preset process phase, and to acquire the response signal under the micro-excitation.

[0009] The probability prediction module is used to perform time-varying registration of the response signal with the Green function library, reconstruct the defect probability voxel map inside the automotive forging through a sparse constraint inversion algorithm, and output the defect confidence interval.

[0010] The defect assessment module is used to simulate the defect propagation trend of automotive forgings based on the defect probability voxel map and the virtual model, combined with the defect confidence interval, to obtain the defect risk density.

[0011] The appearance data includes the three-dimensional appearance drawing data of the automotive forging. The model building module includes a virtual model building unit and a function library building unit. The virtual model building unit is used to generate a virtual model based on the appearance data and material properties. The function library building unit is used to establish contact conditions in the virtual model and calculate the disturbance response to generate a Green's function library. The virtual model building unit is configured as follows:

[0012] Based on the three-dimensional appearance drawing data, combined with the preset blank length-to-diameter ratio parameters, mold geometric parameters and forging process parameters, an initial three-dimensional geometric model consistent with the size of the automobile forging is generated.

[0013] The initial three-dimensional geometric model is assigned values ​​based on the material property parameters of the automotive forging to obtain a virtual model. The material property parameters include thermal conductivity, specific heat capacity, coefficient of thermal expansion, yield strength, flow stress function, and damage evolution parameters.

[0014] The function library construction unit is configured as follows:

[0015] The contact area between the blank and the mold is marked in the virtual model, and the parameters of the contact area are set, including the friction coefficient, thermal contact resistance and lubrication condition parameters, which are used to represent the time-varying boundary conditions of the contact area during the forging process.

[0016] The interface unit between the blank and the mold is discretized within the contact area, and a preset disturbance load is applied to the discretized interface unit.

[0017] The displacement response, strain response, and acoustic response of the disturbance load propagating to the external sensing point in the virtual model are calculated using the finite element time-domain solution method. Based on the displacement response, strain response, and acoustic response, the response function is calculated in combination with time-varying boundary conditions.

[0018] The response function is used to establish a multi-index mapping table based on the interface unit position, sensor point number and process phase to obtain the original library of Green's function;

[0019] The original Green's function library is subjected to matrix decomposition and sparse basis vector extraction, and compressed and mapped into a low-dimensional sparse dictionary;

[0020] In the low-dimensional sparse dictionary, a corresponding sub-dictionary is established for different process phases, and the sub-dictionary is bound and stored with the virtual model to obtain the Green function library.

[0021] The process phase refers to a specific stage in the forging process of automotive forgings, which includes the pre-pressing phase, plastic deformation phase, final pressing phase, and constant pressing phase.

[0022] The signal acquisition module includes an excitation unit and a signal acquisition unit. The excitation unit is used to apply narrow-band frequency sweep micro-excitation to the automotive forging at a preset process phase. The signal acquisition unit is used to acquire the response signal through a sensor. The excitation unit is configured as follows:

[0023] At the arrival time of the preset process phase, a phase state vector is constructed based on the measured values ​​of slide displacement, slide speed, mold temperature and contact resistance, wherein the slide is the slide of the forging equipment for the automotive forging.

[0024] The excitation feasible region of the phase state vector is determined by a preset mapping rule, wherein the excitation feasible region includes the injection time window, the upper limit of the excitation amplitude, the candidate interval of the sweep center frequency, and the upper limit of the sweep bandwidth.

[0025] The excitation feasible region is corrected based on the center frequency of the corresponding process phase in the obtained historical forging cycle to obtain the frequency sweep sequence;

[0026] The frequency sweep sequence is superimposed along the coaxial direction of the forming load of the automotive forging and injected into the slide drive channel.

[0027] The excitation unit is configured with correction logic, which includes:

[0028] Using the center frequency as the initial value, and according to the preset screening threshold, the trial frequency pair corresponding to the initial value is selected in the candidate interval of the sweeping center frequency in the excitation feasible domain to obtain the pilot single-frequency excitation.

[0029] The analog response signal corresponding to the pilot single-frequency excitation is obtained by simulation inversion, and the analog response signal is phase-sensitively demodulated to obtain the amplitude index corresponding to the test frequency pair.

[0030] The amplitude index is iterated through extreme value optimization update to calculate the optimized center frequency value, wherein the optimized center frequency value is located within the candidate interval of the swept center frequency;

[0031] Based on the optimized center frequency value and the phase state vector, the sweep bandwidth, frequency stepping strategy, and amplitude envelope are determined, and a sweep sequence is generated.

[0032] During the iterative calculation process, the iteration stops when the rate of change of the contact resistance or the rate of change of the slide velocity exceeds a preset threshold.

[0033] The probability prediction module includes a registration unit and an inversion unit. The registration unit performs time-varying registration of the response signal and the Green's function library to obtain a registered observation vector. The inversion unit performs inversion calculations including sparse constraints based on the observation vector and the virtual model to reconstruct a defect probability voxel map and output a defect confidence interval. The registration unit is configured as follows:

[0034] The response signal is synchronously preprocessed to obtain a preprocessed response, wherein the synchronous preprocessing includes time base alignment based on ram displacement triggering, DC drift removal, bandpass filtering, and full window processing;

[0035] Based on the phase state vector and process phase corresponding to the preprocessed response, the corresponding phase sub-library is obtained from the Green function library, and the Green functions in the phase sub-library are interpolated by time and amplitude to obtain a time-varying Green function set;

[0036] Based on the time-varying Green's function set, the time offset and amplitude scaling factor of each channel are obtained through cross-correlation estimation. The preprocessed response is then time-shifted and amplitude normalized based on the time offset and amplitude scaling factor to obtain the registered observation vector, wherein the channel represents the spatial location of the preprocessed response.

[0037] The inversion unit is configured as follows:

[0038] Define a voxel domain in the coordinate system of the virtual model and generate a three-dimensional voxel mesh, and assign corresponding voxel parameter vectors in the three-dimensional voxel mesh;

[0039] Based on the geometric mapping relationship of the three-dimensional voxel mesh and the time-varying Green's function set, an observation matrix is ​​generated with channels as rows and phase indices as columns, and a weight matrix is ​​calculated based on the noise variance of each channel in the observation vector.

[0040] The observation vector, observation matrix, and weight matrix are input into the weighted least squares inversion model. The observation vector, observation matrix, and weight matrix are sparsely regularized by the weighted least squares inversion model and solved by the alternating direction multiplier method to obtain the voxel parameter estimates. The sparse regularization includes L1 norm constraints on the voxel parameter vector, group sparse constraints obtained by segmenting based on concentric rings and axial direction, and total variational constraints on the discrete gradients of each pair of voxel parameter vectors.

[0041] Based on the voxel parameter estimation, the voxel-level defect probability value is calculated through probability mapping, and the voxel parameter vector is updated and replaced with the voxel-level defect probability value, wherein the probability mapping is obtained through logistic regression.

[0042] Based on the registration residuals obtained during the solution process of the weighted least squares inversion model, the observation noise covariance estimate is calculated. The observation vector is then resampled using the observation noise covariance estimate to obtain the sampling distribution.

[0043] Based on the significance level of the sampling distribution, determine the upper and lower confidence bounds for each voxel, and output the defect probability voxel map and defect confidence interval.

[0044] The mathematical expression of the weighted least squares inversion model is as follows:

[0045] ,

[0046] in, Represents the weight matrix. This indicates finding the minimum value. Let represent the observation vector, and represent the voxel parameter vector. Represents the observation matrix. Represents the phase state vector. This indicates the registration time index. , and The weighting parameter is calculated using the channel noise level and the sweep bandwidth. This represents the weighted norm 2 terms. Voxel parameter vector L1 norm regularization, Indicates sparse regularization of groups. Represents a set of voxel grouping indexes. Indicates the grouping index of a single voxel. Indicates the first The voxel parameter vector of voxel grouping, Indicates the first Group weights for individual element groupings This indicates the total variation regularity. The gradient represents the voxel parameter vector. This represents the upper bound of the voxel parameter vector.

[0047] The defect assessment module is configured as follows:

[0048] Based on the preset service load and boundary condition library, set the time domain analysis step size and sub-step control parameters corresponding to the process phase;

[0049] The defect probability voxel map and defect confidence interval are mapped to the finite element mesh through a mapping operator to obtain the initial defect state set, wherein the initial defect state set includes the void volume fraction and the equivalent initial crack size;

[0050] Starting with the initial temperature field and residual stress field of a standard automotive forging in the corresponding process phase, the finite element mesh is transiently solved under the constraints of the time domain analysis step size and sub-step control parameters to obtain a time history set including stress, strain, temperature and hydrostatic pressure.

[0051] The initial defect state set is updated step by step according to a preset update law based on the time history set, and the defect risk density is output.

[0052] Compared with the prior art, the beneficial effects of this application are:

[0053] This application enables the accurate mapping of signal responses under real working conditions during the forging process by constructing a virtual model and a Green's function library, thereby improving the spatial resolution of defect detection. By applying narrow-band sweep frequency micro-excitation to a preset process phase and combining it with an inversion algorithm containing sparse constraints, the probability voxel map of defects inside the forging can be effectively reconstructed and the defect confidence interval can be output, enabling early identification of latent defects such as shrinkage cavities, cracks, and inclusions. Furthermore, by combining the virtual model to simulate the defect propagation trend, the defect risk density can be obtained, providing a quantitative basis for the service reliability of the forging. Attached Figure Description

[0054] 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:

[0055] Figure 1 This is a schematic diagram illustrating the defect detection problem during the forging process of automotive forgings, as described in an embodiment of this application.

[0056] Figure 2 This is an exemplary application scenario diagram of an embodiment of this application;

[0057] Figure 3 This is a block diagram of a digital twin-based automotive forging defect prediction system according to an embodiment of this application;

[0058] Figure 4 This is a flowchart of a digital twin-based method for predicting defects in automotive forgings, as described in this application. Detailed Implementation

[0059] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0060] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0061] In modern automotive forging production lines, quality control is shifting from end-of-line sampling to in-process visualization. Taking long, slender parts like engine transmission components as an example, the process cycle is tight, the die cavity is deeply enclosed, and loads and temperatures are coupled at high speeds. Conventional offline ultrasound or CT scans are insufficient to provide actionable judgments within the forming window. On the other hand, while digital twins are used for process simulation, without online correction to the actual travel conditions, model predictions can easily become disconnected from the actual process. This creates a typical yet challenging scenario:

[0062] Under conditions of strong impact, strong thermal radiation, and limited field of view, how can we obtain observable measurements sensitive to hidden defects such as internal shrinkage cavities, microcracks, inclusions, and agglomerations within a millisecond time window, and reliably map them to defect levels?

[0063] The solution provided in this application is based on such constraints. By injecting narrowband micro-excitations with extremely small amplitudes and controlled frequencies in the coaxial direction of the forming load, the mold-form material system generates an interpretable dynamic response without disturbing the main forming target. Then, using a virtual model associated with the process phase as a carrier, the response function set used for inversion is pre-calculated and continuously modified, thereby constructing an implicit channel within the closed mold cavity.

[0064] The scope of this application is not limited to a particular forging geometry or a specific type of equipment. Any automotive forging that has a large length-to-diameter ratio, is axisymmetric or near-axisymmetric, and is prone to developing central shrinkage cavities, annular inclusions, or longitudinal microcracks during secondary extrusion, final pressing, or holding pressure stages can be included in the scope of this application. Specific products may include valve lifters, piston pins, and synchronizer rings, or other components with high triaxiality regions deep within the die cavity.

[0065] Applicable scenarios are typically accompanied by the following characteristics in production states:

[0066] Mold closure restricts the observation path of external sensors, and lubrication status and contact thermal resistance change rapidly with the stroke.

[0067] Equipment noise and coupled vibrations result in a low signal-to-noise ratio for passive acoustic signals;

[0068] Microscopic differences in materials between batches can have a perceptible impact on the defect initiation threshold.

[0069] This application does not require fixed partitioning of forgings or unique setting of process routes, nor does it rely on specific brand sensors and press models; as long as a virtual model and response function set can be established under the premise that the stroke phase can be determined, micro-excitation can be injected, and response can be collected, the entire process can be connected.

[0070] Understandably, in order to combine the flexibility of the simulated world with the credibility of the physical world, this application employs two interlocking main lines in its implementation:

[0071] Firstly, a virtual model is constructed and maintained around the process phase, including geometric and material thermodynamic parameters and time-varying contact boundaries, thereby providing a basis for the generation of response functions;

[0072] Secondly, by focusing on time-varying registration and sparse inversion of micro-excitation and multi-channel response, a stable mapping from millisecond-level time-domain signals to voxel-level defect probabilities is formed.

[0073] Both methods perform iterative corrections using the center frequency and amplitude envelope of historical cycles, enabling the model to adapt to each component and converge phase by phase. The final output is not a single-point discrimination, but a three-dimensional defect probability voxel map with confidence boundaries. This is then used to solve for defect evolution within the service load library, generating a risk density field consistent with the voxel coordinates, providing direct quantitative basis for process fine-tuning or component-level release.

[0074] Taking valve lifters as an example, refer to Figure 1 , Figure 1 This is a schematic diagram illustrating the defect detection problem in the forging process of automotive forgings, as provided in an embodiment of this application.

[0075] Figure 1 The diagram illustrates the high-pressure die forging of valve tappets, typically achieved by upper and lower dies on a pre-heated billet. This process involves pre-pressing, plastic deformation, and final pressing, with the material rapidly filling the die cavity and forming the desired structure. Due to the large length-to-diameter ratio of valve tappets, the complex internal metal flow path, and the tendency for high triaxial stress zones to form deep within the die cavity, defects such as shrinkage cavities, cracks, and inclusions are common during the forming process, as shown in the diagram regarding internal shrinkage cavities and cracks.

[0076] Understandably, image sensing or ultrasonic testing is commonly used to inspect valve tappet forgings for quality. However, image sensing relies on external visible light or infrared imaging, and the closed mold cavity and high temperature of the forging result in a limited field of view, making it difficult to identify deep defects. While ultrasonic testing can penetrate the material, the signal-to-noise ratio drops significantly under high-temperature forging conditions, and the complex shape of the forging and the variable contact conditions often lead to missed or misjudged results. Therefore, existing testing methods struggle to accurately identify internal defects in valve tappets during the forging process in real time.

[0077] Figure 1 The right side shows common defect types, including:

[0078] Shrinkage cavities typically occur in areas where metal flow is insufficient, especially in high triaxial stress areas deep within the die cavity during forging. This is caused by an excessively long or incompletely filled metal flow path, resulting in trapped gas inside and the formation of voids.

[0079] Cracks typically appear in areas of uneven cooling or excessive stress during the forging process, possibly due to localized rapid cooling of the material or stress concentration caused by uneven die pressure. Cracks are often accompanied by strong geometric discontinuities, affecting the structural strength and service life of valve lifters.

[0080] Inclusions are substances that remain embedded in the metal during the forging process due to incomplete removal of non-metallic or other impurities. These impurities may originate from the mold, impure raw materials, or incomplete reactions during heating, and are usually located inside the metal, affecting the material's uniformity and mechanical properties.

[0081] refer to Figure 2 , Figure 2 This is an exemplary application scenario diagram provided for an embodiment of this application.

[0082] Figure 2 Taking valve tappets as an example, when the forging process reaches the preset process phase, a narrow-band sweep frequency micro-excitation is applied to the valve tappet. The narrow-band sweep frequency micro-excitation is superimposed on the forming load direction in the form of a low-amplitude, controlled frequency band disturbance, which is used to extract the response characteristics inside the forging without interfering with the main deformation process.

[0083] Furthermore, Figure 2 This diagram illustrates a system where multiple signal sensors are arranged circumferentially and axially on the outer surface of a valve tappet. Sensors at different locations can simultaneously acquire displacement, strain, or acoustic signals, forming multi-channel response data. This multi-channel response data includes the mechanical propagation characteristics and defect disturbance traces of the forging under specific process phases.

[0084] Furthermore, Figure 2The process involves transmitting the acquired multi-channel response signals to a processor. The processor invokes a virtual model and Green's function library corresponding to the current process phase, and normalizes the response signals in terms of time and amplitude using a time-varying registration algorithm, aligning signals from different channels and cycles under a unified reference. Based on the registered observation vectors, the processor further executes an inversion algorithm with sparse constraints, mapping the signal space to the voxel domain of the virtual model to reconstruct a defect probability voxel map. It then provides voxel-level defect probability boundaries using confidence intervals. The processor displays the generated defect probability voxel map in association with the virtual model and outputs the defect detection results. These results not only include the spatial location and probability distribution of defects within the valve lifter but also measure their reliability using confidence intervals. Based on these results, process engineers can obtain real-time information on the distribution of internal defects during forging and adjust process parameters accordingly to reduce the risk of defect propagation and service failure.

[0085] Understandable Figure 2 The signal excitation can be achieved by an electromagnetic exciter, a piezoelectric ceramic actuator or a hydraulic micro-vibration unit, or by an auxiliary drive module installed in the slide drive circuit, so as to inject a narrow-band sweep frequency signal into the superposition path of forming load.

[0086] The signal sensor can be a piezoelectric accelerometer, a resistance strain gauge, a fiber Bragg grating sensor, or an electromagnetic acoustic transducer that operates under high temperature conditions, used to synchronously acquire displacement, strain, or acoustic signals at different locations.

[0087] The processor can be an industrial control computer, an embedded signal processor, a field-programmable logic device, or an industrial control server equipped with a graphics computing unit, used to execute large-scale matrix operations and sparse inversion algorithms. The processor can interact with the virtual model and Green's function library through data interfaces to achieve real-time registration, inversion, and output of defect probability voxel maps of the acquired signals.

[0088] It is important to note that, taking the forging process of valve tappets as an example, the signal sensors are typically positioned on the outer surface of the forging near the neck transition zone, the middle section of the cylinder, and the outer periphery of the end contact area with the die. These locations are highly sensitive to the disturbance characteristics of internal defects on the vibration and strain field during the forming process. The signal sensors can also be arranged on the outer wall of the die or adjacent to the slide block, depending on different process requirements, to achieve indirect measurement of the overall response of the forging. Those skilled in the art can adjust the number and distribution of sensors according to the equipment structure, thermal environment, and signal acquisition accuracy requirements, and are not limited to the above-mentioned forms.

[0089] refer to Figure 3 , Figure 3A block diagram of a digital twin-based automotive forging defect prediction system provided in this application embodiment.

[0090] In one example, this application embodiment provides a digital twin-based defect prediction system for automotive forgings, the system comprising:

[0091] The model building module is used to acquire the appearance data of automotive forgings and to build a virtual model and Green's function library based on the appearance data.

[0092] The signal acquisition module is used to perform narrow-band frequency sweep micro-excitation on the automotive forging when the forging process of the automotive forging is in a preset process phase, and to acquire the response signal under the micro-excitation.

[0093] The probability prediction module is used to perform time-varying registration of the response signal with the Green function library, reconstruct the defect probability voxel map inside the automotive forging through a sparse constraint inversion algorithm, and output the defect confidence interval.

[0094] The defect assessment module is used to simulate the defect propagation trend of automotive forgings based on the defect probability voxel map and the virtual model, combined with the defect confidence interval, to obtain the defect risk density.

[0095] The appearance data includes the three-dimensional appearance drawing data of the automotive forging. The model building module includes a virtual model building unit and a function library building unit. The virtual model building unit is used to generate a virtual model based on the appearance data and material properties. The function library building unit is used to establish contact conditions in the virtual model and calculate the disturbance response to generate a Green function library.

[0096] The signal acquisition module includes an excitation unit and a signal acquisition unit. The excitation unit is used to apply narrow-band sweep frequency micro-excitation to the automotive forging at a preset process phase, and the signal acquisition unit is used to acquire the response signal through a sensor.

[0097] The probability prediction module includes a registration unit and an inversion unit. The registration unit is used to perform time-varying registration of the response signal and the Green's function library to obtain the registered observation vector. The inversion unit is used to perform inversion calculations including sparse constraints based on the observation vector and the virtual model to reconstruct the defect probability voxel map and output the defect confidence interval.

[0098] Next, combined Figure 4 This paper introduces a method for predicting defects in automotive forgings based on digital twins, as provided in the embodiments of this application. Figure 4 The method shown is applied to the digital twin-based automotive forging defect prediction system provided in this application embodiment. The specific steps are as follows:

[0099] S1: Obtain the appearance data of automotive forgings, and construct a virtual model and Green's function library based on the appearance data;

[0100] In this embodiment, the appearance data of automotive forgings, represented by valve tappets, is derived from 3D drawings created during the design phase. Based on the 3D appearance data, combined with the billet's aspect ratio parameters, die geometry parameters, and forging process parameters, an initial 3D geometric model consistent with the target dimensions is first generated. Further, material property parameters, including thermal conductivity, specific heat capacity, coefficient of thermal expansion, yield strength, flow stress function, and damage evolution parameters, are assigned to the initial 3D geometric model, forming a virtual model with physical response characteristics. The contact area between the billet and the die is then marked in the virtual model, and parameters such as friction coefficient, thermal contact resistance, and lubrication conditions are set. A preset perturbation load is applied to the discretized interface elements, and the displacement, strain, and acoustic response of external sensing points are obtained using finite element time-domain solutions, thereby establishing a Green's function library. The virtual model and the Green's function library can dynamically map the actual forging process, providing a stable benchmark for subsequent defect prediction and effectively avoiding prediction biases caused by purely empirical models.

[0101] S2: When the forging process of the automotive forging is in a preset process phase, the automotive forging is subjected to narrow-band frequency sweep micro-excitation, and the response signal under the micro-excitation is collected.

[0102] In this embodiment, the process phase includes at least one of the following: pre-compression phase, plastic deformation phase, final compression phase, and constant compression phase. Micro-excitation is injected by an electromagnetic exciter in the form of a low-amplitude, narrow-band swept frequency signal, superimposed on the slide drive channel along the coaxial direction of the forming load to ensure that internal response characteristics are elicited without affecting the forging process itself. Multiple sets of high-temperature sensors, such as resistance strain gauges, piezoelectric accelerometers, or electromagnetic transducers, are arranged axially and circumferentially on the valve tappet surface, and the acquired response signals can cover the propagation paths at different locations.

[0103] S3: Perform time-varying registration between the response signal and the Green function library, reconstruct the defect probability voxel map inside the automotive forging using a sparse constraint inversion algorithm, and output the defect confidence interval;

[0104] In this embodiment, the acquired response signal is first synchronously preprocessed through time base alignment triggered by ram displacement, DC drift removal, bandpass filtering, and full windowing. Then, the corresponding sub-library of the Green's function library is retrieved based on the phase state vector, and time interpolation and amplitude interpolation are performed to form a time-varying Green's function set. The time offset and amplitude scaling factor of each channel are obtained using cross-correlation estimation, and the response signal is registered to obtain an observation vector under a unified reference. Next, within the voxel domain of the virtual model, an inversion model combining weighted least squares with L1 sparsity, group sparsity, and total variational constraints is established, solved using the alternating direction multiplier method, and outputs a defect probability voxel map. Voxel-level confidence intervals are calculated based on residual resampling or the Laplace approximation.

[0105] S4: Based on the defect probability voxel map and the virtual model, and combined with the defect confidence interval, simulate the defect propagation trend of automotive forgings to obtain the defect risk density;

[0106] In this embodiment, by applying alternating forces, temperature fields, and friction boundary conditions under service conditions, the defect state is gradually updated, forming a multi-scenario evolution by combining probability confidence intervals. After statistically analyzing the simulation results of multiple scenarios, a voxel-level defect risk density field is calculated. This risk density field not only reflects the current location and expansion direction of the defect but also characterizes the impact of uncertainty on risk.

[0107] Before going into the specific steps, the embodiments of this application need to emphasize again:

[0108] For automotive forgings with a slenderness ratio, such as valve tappets, the forming process is subject to typical constraints: when the mold is fully closed, the blank is in a high-temperature and high-pressure environment, the metal flow path becomes complicated due to the geometric constraints of the cavity, and shrinkage cavities, cracks or inclusions are easily formed inside.

[0109] Understandably, conventional inspection methods, such as ultrasonic testing or radiographic imaging, can often only be performed after forging is complete. Due to uncontrollable coupling conditions, low signal-to-noise ratios in high-temperature environments, and multipath attenuation caused by complex structures, early defects are difficult to reliably detect before service. Unlike real defects that gradually expand during service, these detection blind spots often manifest as: ambiguous signal response delays, unstable spatial positioning, or false high-amplitude points generated after multiple energy scattering, which are mistaken for defect signs. Relying on conventional thresholding or filtering methods for differentiation can easily lead to the mistaken exclusion of real defects or the retention of false disturbances, resulting in subsequent quality risks.

[0110] The core processing logic employed in this embodiment does not rely solely on the strength of traditional signals or single waveform characteristics. Instead, it transforms externally acquired response signals into voxel-level perturbation probability distributions by introducing a time-varying Green's function library that is strongly bound to the process phase using a digital twin virtual model. Unlike existing methods, this application does not rely on external single-path imaging to directly observe defects. Instead, it constructs a correction channel that highly corresponds to the actual physical behavior within the process window using controllable inputs of narrowband micro-excitations. In this way, defect manifestations are no longer vague single-point anomalies, but are mapped into probability clusters through sparse constraint inversion, described by confidence intervals. When the probabilities of different voxels exhibit a continuous clustering trend in space, rather than scattered point noise, they can be identified as genuine defects with high confidence; conversely, spurious responses caused by boundary reflections or thermal noise can be eliminated.

[0111] Next, the technical content of constructing a virtual model in the embodiments of this application will be further elaborated.

[0112] In one example, the specific steps of S1 are as follows:

[0113] S1.1: Based on the three-dimensional appearance drawing data, combined with the preset blank length-to-diameter ratio parameters, mold geometric parameters and forging process parameters, an initial three-dimensional geometric model consistent with the size of the automobile forging is generated;

[0114] Specifically, in the process of defect prediction for forged parts, relying directly on two-dimensional structural drawings or physical scan results often fails to accurately reflect the geometric evolution of the forging during the forming process. To ensure that the digital twin model is consistent with the actual process, it is necessary to use three-dimensional appearance drawing data as a basis and overlay the billet length-to-diameter ratio parameters, die geometric parameters, and forging process parameters to generate an initial three-dimensional geometric model whose size and shape conform to the target forging.

[0115] In this embodiment, taking a valve tappet as an example, the 3D appearance drawing data originates from the CAD design stage. This data includes the valve tappet's outer contour line, key cross-sectional dimensions, and the interface position with the mold. The length-to-diameter ratio parameter is used to determine the initial geometric proportions of the blank, thereby controlling the uniformity of the blank's filling in the mold cavity. The mold geometric parameters include the cavity depth of the upper and lower molds, the fillet transition radius, and the end flattening area, etc. These parameters can explicitly constrain the boundaries of the virtual model. The process flow parameters cover the ram stroke and closing speed in steps such as pre-pressing and final pressing, used to reflect the dynamic load characteristics of the forging process. The initial 3D geometric model is not merely a static shape mapping but also possesses the ability to describe the forming path, boundary conditions, and process cycle time.

[0116] Furthermore, during the generation of the initial 3D geometric model, parametric modeling methods can be introduced for different process scenarios. For example, in the neck and head regions of the valve tappet, local mesh refinement technology is used to subdivide the geometric model into smaller units, so as to more accurately capture the location of stress concentration and potential defect initiation in subsequent calculations.

[0117] S1.2: Assign values ​​to the initial three-dimensional geometric model based on the material property parameters of the automotive forging to obtain a virtual model, wherein the material property parameters include thermal conductivity, specific heat capacity, coefficient of thermal expansion, yield strength, flow stress function and damage evolution parameters;

[0118] Specifically, when building a virtual model, geometry alone is insufficient to reflect real physical behavior; it is necessary to assign values ​​to material property parameters in order to simulate the thermodynamic response of forgings under high temperature and high pressure conditions.

[0119] For example, valve lifters are typically made of alloy steel or carburized steel. The thermal conductivity, plastic flow characteristics, and damage evolution of these materials at high temperatures have a direct impact on the formation and propagation of defects.

[0120] In this embodiment, thermal conductivity and specific heat capacity are used to describe the temperature field evolution of the blank during heating and forming, and can predict cold shut defects caused by uneven local heat dissipation; the coefficient of thermal expansion is used to characterize the geometric dimensional changes caused by the temperature gradient, so as to correct the contact state between the mold and the blank; the yield strength and flow stress function together determine the plastic flow behavior of the material at different temperatures and strain rates, so that the virtual model can reproduce the real strain distribution and local necking phenomenon; the damage evolution parameters are used to characterize the initiation and propagation process of microcracks and voids, and by coupling with the virtual model, an evolution trajectory consistent with the actual defect distribution can be formed in the simulation.

[0121] Furthermore, to improve the accuracy of the virtual model, a partitioned assignment method can be adopted during the assignment process. This involves spatially distributing and differentiating the material property parameters based on the stress and thermal history differences in different parts of the valve lifter. For example, a higher coefficient of thermal expansion can be assigned to the head region of the valve lifter to simulate localized thermal stress concentration when the mold cavity closes; different flow stress functions can be assigned to the middle region to reflect the influence of strain rate on material flow. In addition, experimental data or database fitting can be used to bind the material's damage evolution parameters with temperature and strain rate in a multidimensional functional relationship, thereby achieving dynamic response simulation for complex working conditions.

[0122] Next, we will further elaborate on the technical content of constructing the Green function library in the embodiments of this application.

[0123] It is understandable that a Green's function library refers to a set of mappings formed in a virtual model by calculating and storing the response of the billet-die system under specific perturbation conditions, used to describe the time-varying relationship between input excitation and output response. During forging, the contact state between the billet and the die changes continuously with the process phase. Thermal boundary conditions, friction conditions, and material constitutive properties all significantly affect the excitation propagation path. Therefore, relying solely on a single or static response function cannot accurately reflect the actual dynamic behavior. This application constructs a Green's function library to store the perturbation propagation characteristics under different process phases in function form, which serves as the core reference for registration and inversion in subsequent defect prediction. In this way, complex time-varying boundaries are transformed into a callable set of functions, establishing an updatable and iterative correspondence between the virtual model and the real process.

[0124] In one example, the specific steps of S1 also include:

[0125] S1.3: Mark the contact area between the blank and the mold in the virtual model, and set the parameters of the contact area, wherein the parameters include the friction coefficient, thermal contact resistance and lubrication condition parameters, which are used to represent the time-varying boundary conditions of the contact area during the forging process;

[0126] Specifically, in forging operations, the contact area between the billet and the die is a key factor determining the material filling path and the probability of defect generation. Failure to accurately model this area will lead to significant discrepancies between the virtual model and the actual forming process. Therefore, it is necessary to annotate the contact surface between the billet and the die in the virtual model and introduce parameters such as the coefficient of friction, thermal contact resistance, and lubrication conditions to describe the evolution of the contact state. These parameters not only determine the transfer of energy and momentum but also directly affect the uniformity of metal flow and local temperature rise.

[0127] In this embodiment, the coefficient of friction depends on the surface treatment of the mold and the type of lubricant, such as graphite emulsion or glass lubricating film; the thermal contact resistance is used to reflect the resistance to heat transfer between the contact surfaces, and its value is affected by the mold preheating temperature and contact pressure; the lubrication condition parameters are used to distinguish whether the lubrication is evenly distributed and whether the lubrication layer is stable under high temperature and high pressure.

[0128] Furthermore, the parameters of the contact area can be calibrated experimentally. For example, the friction coefficient can be measured using a friction testing machine under different pressures and temperatures, or the transient thermal resistance between the mold and the blank can be estimated using a heat flow sensor. The calibrated parameters are then input into the virtual model, ensuring that the simulation results are highly consistent with the actual process, thereby reducing prediction errors caused by inaccurate boundary conditions.

[0129] S1.4: Discretize the interface unit between the blank and the mold in the contact area, and apply a preset disturbance load to the discretized interface unit;

[0130] Specifically, the physical processes in the contact region involve complex nonlinear behavior, and direct overall calculation would lead to numerical instability. Therefore, it is necessary to divide the contact region into multiple interface elements and discretize them. This ensures that each element has independent physical properties while accurately capturing the changing characteristics of local friction and heat conduction.

[0131] In this embodiment, the perturbation load can take the form of a pulsed force or a narrowband swept frequency signal, with its direction of action coaxial with the forming main load to ensure that the excitation propagates along the same path as the actual mechanical process. The size of the discretized element is determined according to the finite element mesh generation principle, with finer meshes used in stress concentration regions and geometric transition regions to capture more accurate local response characteristics.

[0132] S1.5: Calculate the displacement response, strain response, and acoustic response of the disturbance load propagating to the external sensing point in the virtual model using the finite element time-domain solution method, and calculate the response function based on the displacement response, strain response, and acoustic response, combined with time-varying boundary conditions.

[0133] Specifically, after a disturbance load is applied to the contact area, it propagates in the virtual model in the form of fluctuations or stress waves. Therefore, the finite element time-domain solution method is required to calculate the disturbance propagation process. This method can obtain the displacement, strain, and acoustic signals of sensing points at different locations, thereby forming a response function that characterizes the input-output relationship.

[0134] In this embodiment, the arrangement of the sensing points corresponds to the sensor locations that might be placed on the actual forging, typically in the neck transition area, the middle section, and the circumferential region at the ends. By solving the problem in the finite element time domain, not only can the response waveform in the time domain be obtained, but the propagation characteristics of different frequency bands can also be extracted through spectral analysis. The resulting response function accurately describes the entire process of disturbance transmission from the contact area to the sensing points.

[0135] S1.6: Establish a multi-index mapping table for the response function based on the interface unit position, sensor point number and process phase to obtain the original library of Green's function;

[0136] S1.7: Perform matrix decomposition and sparse basis vector extraction on the original Green's function library, and compress and map it into a low-dimensional sparse dictionary;

[0137] Specifically, the original Green's function library contains an extremely large amount of data, and direct use would lead to excessively low efficiency in registration and inversion calculations. Therefore, it is necessary to perform matrix decomposition and sparse basis vector extraction on the original library to compress the high-dimensional and complex set of functions into a low-dimensional sparse dictionary, so as to retain the main features while reducing storage and computational costs.

[0138] In this embodiment, matrix decomposition methods such as singular value decomposition or principal component analysis can be used to extract the main feature vectors from the original response matrix. Then, representative basis vectors are selected using sparse coding methods to construct a low-dimensional sparse dictionary. The low-dimensional sparse dictionary not only contains the main information of the original response but also effectively filters out noise and redundant parts, thereby improving the stability of subsequent inversion.

[0139] S1.8: Establish corresponding sub-dictionaries for different process phases in the low-dimensional sparse dictionary, and bind and store the sub-dictionaries with the virtual model to obtain the Green function library;

[0140] In this embodiment, the sub-dictionary is built based on the process phase index. Relevant basis vectors in the sparse dictionary are grouped and stored, and then bound to the virtual model during storage. This allows the virtual model to directly access the corresponding response features during invocation. This binding relationship ensures the synchronization between the virtual model and the Green's function library, enabling rapid invocation without additional parameter matching during the prediction process.

[0141] Next, the technical content regarding the process phase in the embodiments of this application will be further elaborated.

[0142] The process phase refers to a specific stage in the forging process of automotive forgings, which includes the pre-pressing phase, plastic deformation phase, final pressing phase, and constant pressing phase.

[0143] It is understandable that the pre-pressure phase refers to the stage when the ram first contacts the billet at the beginning of forging and gradually applies the initial load; the plastic deformation phase refers to the stage when the ram descends further and the billet enters a state of significant plastic flow; the final pressure phase refers to the stage when the die is basically closed and the billet approaches its final geometry; and the constant pressure phase refers to the stage when the pressure is maintained after the ram reaches the end of its stroke.

[0144] Specifically, the pre-compression phase is characterized by small ram displacement and slow speed, with the mold cavity not yet fully closed, and the blank mainly undergoing slight plastic deformation. At this time, contact surface friction and heat conduction begin to play a role, and the friction state determines the direction and uniformity of subsequent metal flow.

[0145] The plastic deformation phase is characterized by an increase in ram velocity, a rapid increase in internal strain and strain rate of the billet, and a nonlinear characteristic of flow stress changing with strain rate at high temperatures. In this case, the narrow-band sweep excitation needs to be adjusted to a higher center frequency to enhance sensitivity to local defects.

[0146] The characteristic of the final pressure phase is that the ram displacement is close to the stroke limit, the blank deformation rate gradually decreases, and the mold cavity filling tends to be completed. At this time, the internal residual voids are prone to forming shrinkage defects due to triaxial stress state.

[0147] The characteristic of a constant pressure phase is that the displacement of the slide remains unchanged, and the inside of the billet is in a static stress state of high temperature and high pressure. At this time, although there is no large plastic flow, the internal stress relaxation and heat conduction process continues.

[0148] Next, the technical content of the narrowband sweep frequency micro-excitation in the embodiments of this application will be further elaborated.

[0149] In one example, the specific steps of S2 are as follows:

[0150] S2.1: At the arrival time of the preset process phase, a phase state vector is constructed based on the measured values ​​of slide displacement, slide speed, mold temperature and contact resistance, wherein the slide is the slide of the forging equipment for the automotive forging.

[0151] Specifically, during the forging process, the thermo-mechanical boundary conditions differ significantly under different process phases. Without accurate state parameters, it is impossible to precisely determine the appropriate timing and amplitude of the excitation. Therefore, it is necessary to collect key parameters, including ram displacement, ram velocity, die temperature, and contact resistance, at the instant the process phase arrives, and combine them into a phase state vector to characterize the working conditions of the forging, ensuring that the excitation application is highly synchronized with the actual forging process.

[0152] In this embodiment, the ram displacement is acquired in real time by a displacement sensor to determine the degree of mold cavity closure; the ram speed is fed back by a servo hydraulic control system to reflect the intensity of plastic flow; the mold temperature is collected by a thermocouple array to reflect changes in the thermal boundary; and the contact resistance is indirectly obtained by a resistance measurement circuit to represent the interface contact state.

[0153] S2.2: Determine the excitation feasible region of the phase state vector through a preset mapping rule, wherein the excitation feasible region includes the injection time window, the upper limit of the excitation amplitude, the candidate interval of the sweep center frequency, and the upper limit of the sweep bandwidth;

[0154] Specifically, the application of the excitation signal must be constrained by process conditions; otherwise, it will interfere with normal forging or cause signal distortion. By combining the phase state vector with a preset mapping rule, the time window, amplitude, and frequency range of the excitation can be dynamically limited, thereby forming the excitation feasible region.

[0155] In this embodiment, the injection time window is determined by the threshold range of the slide displacement and slide velocity to ensure that the excitation is applied within the allowable load fluctuation range; the upper limit of the excitation amplitude is limited by the mold bearing capacity and the nonlinear characteristics of the material to avoid causing additional damage to the forging; the candidate range of the sweep frequency center frequency is derived from the material's natural frequency and propagation characteristics to ensure that the excitation can effectively couple the defect response; the upper limit of the sweep frequency bandwidth is related to the noise tolerance and sensor sensitivity to balance resolution and stability.

[0156] S2.3: The excitation feasible region is corrected according to the center frequency of the corresponding process phase in the obtained historical forging cycle to obtain the frequency sweep sequence;

[0157] Specifically, relying solely on theoretically defined excitation feasible regions may deviate from actual operating conditions, necessitating the introduction of historical forging data for correction. By extracting the center frequency information under the same process phase in the past, the frequency range within the excitation feasible region can be adjusted to more closely approximate actual operating conditions, thereby obtaining a more reasonable frequency sweep sequence.

[0158] In one example, the specific steps of S2.3 are as follows:

[0159] S2.3.1: Using the center frequency as the initial value, select the trial frequency pair corresponding to the initial value in the candidate interval of the sweeping center frequency in the excitation feasible domain according to the preset screening threshold to obtain the pilot single-frequency excitation;

[0160] Specifically, the center frequency obtained from historical forging cycles only reflects the statistically optimal value. Under actual operating conditions, the optimal sensitive frequency will shift due to the influence of phase state, contact boundary, and temperature drift. To avoid energy waste or exceeding the sensitive band by directly performing frequency sweeping at this center frequency, this step sets a pair of symmetrical trial frequencies around the center frequency within the candidate interval. The direction of the response slope under the current operating condition is determined by a single-frequency injection with an extremely short duration and low amplitude. The so-called screening threshold sets the entry conditions for acceptable trial frequencies.

[0161] First, the predicted response intensity must be no less than the noise floor elevation limit;

[0162] Second, the relative offset between the test frequency and the center frequency must be within the allowable band corresponding to the phase, so as to ensure that they are still in the local neighborhood of the same physical mode.

[0163] Third, the energy injection and time window must meet the safety constraints of the incentive feasible domain.

[0164] By testing only two adjacent frequency points that pass the screening threshold, a rapid basis for determining the current sensitive band position can be obtained without interrupting the process.

[0165] In this embodiment, the center frequency is used as the initial value to first determine the upper and lower bounds of the candidate interval, and then a phase-related initial offset ratio is obtained by mapping the phase state vector.

[0166] For example, the pre-compression phase offset ratio is three to five percentage points, the plastic deformation phase is five to eight percentage points, and the final compression phase is four to six percentage points; at the same time, an absolute offset lower limit and an upper limit are set (for example, not less than 40 Hz and not more than 400 Hz), and the more stringent one is taken as the final offset.

[0167] Understandably, the generated symmetrical trial frequencies are located above and below the center frequency. Before generating this pair of frequencies, a rapid prediction of the response intensity of the two frequency points is performed using the phase sub-dictionary corresponding to the virtual model. If the predicted amplitude increase of either frequency point is less than twice the noise floor increase, or if the predicted amplitudes of the two frequency points are extremely unbalanced, the offset is adaptively adjusted: the offset is increased to cross the local flat region when the signal-to-noise ratio is poor, and the offset is decreased to improve the local discrimination sensitivity when the signal-to-noise ratio is good.

[0168] Furthermore, to avoid falling on boundary conditions of rapid drift, a contraction coefficient can be applied to the offset by the rate of change of contact resistance and the rate of change of ram velocity: when either of them is in a high-change state, the offset is contracted by a preset ratio, making the test point closer to the center frequency, thereby reducing the discrimination distortion caused by abrupt boundary changes. The two-point single frequency obtained is the pilot single-frequency excitation. The injection duration is referenced to two to three cycles of this frequency point, and the amplitude is controlled to be less than one percent of the forming load, so that the test can be completed within a limited injection time window without affecting the main forming process.

[0169] S2.3.2: Obtain the analog response signal corresponding to the pilot single-frequency excitation through simulation inversion, perform phase-sensitive demodulation on the analog response signal, and obtain the amplitude index corresponding to the test frequency pair;

[0170] Specifically, frequency validity cannot be directly determined from a single-frequency pilot excitation input alone; a virtual model is required for response calculation. This step involves applying a pilot single-frequency excitation to the virtual model to obtain the corresponding simulated response signal, and then performing phase-sensitive demodulation. The purpose of demodulation is to separate the phase and amplitude features from the original signal, extracting the amplitude index corresponding to the probe frequency, which is used to quantify the frequency validity.

[0171] S2.3.3: Iterate the amplitude index through extreme value optimization update to calculate the optimized center frequency value, wherein the optimized center frequency value is located within the candidate interval of the swept frequency center frequency;

[0172] Specifically, the distribution of the amplitude index in the frequency domain reflects the sensitivity of the defect response. As the excitation frequency gradually approaches the inherent coupling frequency band of the defect, the amplitude index gradually increases and reaches a local maximum at a certain frequency point, which corresponds to the optimal center frequency. If a single trial cannot determine the precise location, an extreme value optimization method needs to be used within the candidate interval to iteratively calculate and narrow the search range, ultimately locking in the optimal center frequency value. This avoids the inefficiency caused by blind scanning across the entire frequency band and improves the matching degree with the actual sensitive points of the defect response.

[0173] In this embodiment, the extreme value optimization adopts a partitioned iterative approach: First, several equally spaced frequency points are selected within the candidate interval, and small-amplitude frequency sweeps or single-frequency excitations are applied accordingly. Amplitude indicators are collected through simulation inversion and sensor signals. Then, based on the trend of amplitude changes, the sub-interval where the maximum value may be located is determined. For example, when the amplitude is high at one end of the interval and gradually decreases, the extreme value point is located near that end. During iteration, the interval is narrowed to this sub-interval, and points are selected again for testing within the new interval. This process is repeated until the amplitude converges to the threshold range. To avoid getting trapped in local spikes, this embodiment uses smoothing and weighted averaging methods to jointly determine the amplitude indicators of adjacent frequency points, thereby ensuring that the optimized value corresponds to a stable maximum value point rather than noise interference.

[0174] S2.3.4: Determine the sweep bandwidth, frequency stepping strategy, and amplitude envelope based on the optimized center frequency value and the phase state vector, and generate the sweep sequence;

[0175] Specifically, after obtaining the optimized center frequency value, the final sweep frequency parameters need to be determined by combining the phase state vector. The sweep bandwidth should cover the main distribution range of the defect response frequency; the frequency stepping strategy determines the resolution and speed of the frequency scan; and the amplitude envelope controls the energy distribution at different frequency points, ensuring that the defect response is excited without interfering with the forging process. Through the comprehensive setting of these three aspects, a sweep frequency sequence adapted to the current process conditions can be generated.

[0176] During the iterative calculation process, the iteration stops when the rate of change of the contact resistance or the rate of change of the slide velocity exceeds a preset threshold.

[0177] S2.4: The sweep frequency sequence is superimposed along the coaxial direction of the forming load of the automotive forging and injected into the slide drive channel.

[0178] Next, the technical content of the response signal processing in the embodiments of this application will be further elaborated.

[0179] Understandably, the core principle of response signal processing lies in converting the raw response signal collected during the forging process into standardized data that corresponds to the virtual model and Green's function library, thereby ensuring the reliability of inversion and registration calculations. In one example, the specific steps of S3 are as follows:

[0180] S3.1: Perform synchronous preprocessing on the response signal to obtain a preprocessed response, wherein the synchronous preprocessing includes time base alignment based on ram displacement triggering, DC drift removal, bandpass filtering, and full window processing;

[0181] Specifically, in actual process environments, acquired signals are affected by equipment vibration, thermal noise, electromagnetic interference, and the periodic fluctuations of the forging load itself. If the raw signals are used directly for defect analysis, problems such as time alignment deviation, amplitude drift, and frequency band aliasing are likely to occur, leading to distorted inversion results. Therefore, a series of preprocessing and registration operations are required to remove irrelevant disturbances and recover the characteristic components that can truly reflect the defect response.

[0182] In this embodiment, the effects of asynchronous loading are first eliminated by aligning the time base triggered by the ram displacement. A trigger threshold is set in the ram displacement sensor signal, such as the instant when the displacement reaches the initial contact point or the speed exceeds the set threshold, as a unified reference time for signal acquisition; then the time axis of each channel response signal is offset and corrected relative to this reference point, so that all signals use the same physical event as the zero moment.

[0183] Furthermore, in actual data acquisition, sensors and front-end amplifier circuits often introduce DC components or slowly varying low-frequency drift, manifesting as an overall signal shift or a slow rise / fall in the baseline over time. Without processing, this can lead to false energy peaks in the Fourier spectrum at low frequencies, masking the true defective response. Solutions include estimating the baseline by performing a moving average on the signal and then subtracting this baseline from the original signal; or using a high-pass filter to cut off components below 1 Hz.

[0184] Furthermore, since the initial narrowband sweep excitation is located within a specific frequency range, upper and lower cutoff frequencies can be set based on the preset sweep center frequency and bandwidth to perform bandpass filtering on the response signal. For example, when the sweep range is 2-10kHz, a bandpass filter range of 1.5-11kHz can be set, which not only fully covers the excitation frequency band but also filters out low-frequency mechanical noise and high-frequency electromagnetic interference unrelated to the excitation. The filter can employ a finite impulse response design to ensure phase linearity, thereby avoiding the introduction of new phase distortion due to filtering.

[0185] Furthermore, the purpose of windowing is to truncate the signal into a time window that matches the calculation step size of the Green's function in the virtual model, ensuring that the length of the time-domain data matches the simulated data. Based on the signal sampling rate and excitation duration, a time period covering the entire main response packet is truncated, while a smoothing window function (such as a Hanning window or a Blackman window) is applied to the window edges to reduce spectral leakage. The windowed signal ensures that the main energy components are fully preserved while avoiding the impact of redundant long tails on computational complexity and matching accuracy.

[0186] S3.2: Based on the phase state vector and process phase corresponding to the preprocessed response, obtain the corresponding phase sub-library from the Green function library, and perform time interpolation and amplitude interpolation on the Green functions in the phase sub-library to obtain a time-varying Green function set;

[0187] Specifically, the Green's function library is pre-generated during the virtual model construction phase, storing response functions under different process phases and boundary conditions. However, in actual forging processes, the operating conditions do not strictly fall on the preset discrete points; for example, die temperature, contact resistance, or ram speed are often in some intermediate state. If a single dictionary from the Green's function library is used directly, it can easily lead to a decrease in matching accuracy, resulting in a deviation in the inversion results.

[0188] In this embodiment, the phase state vector serves as the retrieval key, used to find the nearest Green's function records in the Green's function library. For example, when the mold temperature is 950℃, the slide speed is 1.2 m / s, and the contact resistance is 0.3mΩ, the library may only contain sub-libraries corresponding to temperatures of 900℃ and 1000℃, and sub-libraries corresponding to speeds of 1.0m / s and 1.5m / s. At this time, the Green's functions corresponding to these nearby conditions can be automatically selected, and interpolation can be performed in the time dimension to make the waveform phase consistent with the current slide displacement step size; in the amplitude dimension, weighted interpolation is performed according to the parameter scaling factor to reflect the true energy attenuation level.

[0189] Furthermore, to avoid distortion in the nonlinear region by a single linear interpolation, this embodiment can also use piecewise polynomial interpolation or a high-dimensional interpolation method based on radial basis functions, especially in the phase region of high temperature and high-speed deformation, which can better fit the nonlinear characteristics of the response curve. At the same time, in order to suppress waveform jitter that may occur during the interpolation process, wavelet thresholding denoising is performed after time interpolation to ensure that the generated time-varying Green's function set remains smooth and continuous.

[0190] S3.3: Based on the time-varying Green's function set, the time offset and amplitude scaling factor of each channel are obtained through cross-correlation estimation. The preprocessed response is then time-shifted and amplitude normalized based on the time offset and amplitude scaling factor to obtain the registered observation vector, wherein the channel represents the spatial location of the preprocessed response.

[0191] Specifically, in signal acquisition, even after preprocessing, the response waveforms between different sensing points may still exhibit propagation delays and energy imbalances. This is mainly due to local wave velocity variations caused by internal defects in the forging and differences in coupling conditions. Directly using these unaligned and unnormalized signals will degrade the matrix condition number in subsequent inversion equations, leading to computational instability.

[0192] In this embodiment, the preprocessed response is first correlated with the time-varying Green's function of the corresponding channel within a sliding time window to find the correlation peak position, and the offset of this position relative to time zero is recorded; this offset is the time delay of the signal. Then, the ratio of the correlation peak amplitude to the amplitude of the reference Green's function is compared to obtain the amplitude scaling factor. By shifting the original signal along the time axis by the corresponding offset and dividing the amplitude by the amplitude scaling factor, a standardized waveform consistent with the Green's function can be obtained. After processing all channels, a registered observation vector is formed, whose dimension matches the degrees of freedom of the voxel modeling in the virtual model.

[0193] S3.4: Perform voxel domain modeling on the virtual model, and calculate the defect probability voxel map and defect confidence interval by combining the observation vector;

[0194] Specifically, the spatial distribution of defects inside forgings needs to be modeled using voxel modeling. The advantage of voxel modeling is that it divides the three-dimensional geometry into regular discrete units, each unit corresponding to a parameter to be solved (such as a defect probability value). In this way, the observed signals from sensors can be mapped onto these voxel units through Green's functions, thus establishing an inverse relationship from signal to spatial distribution. Without this voxel-domain modeling, only overall statistics of the signal can be obtained, and the spatial location of defects cannot be achieved.

[0195] In one example, the specific steps of S3.4 are as follows:

[0196] S3.4.1: Define a voxel domain in the coordinate system of the virtual model and generate a three-dimensional voxel mesh, and assign corresponding voxel parameter vectors in the three-dimensional voxel mesh;

[0197] Specifically, this application employs a voxel-based modeling approach, dividing the coordinate space of the virtual model into several cubic units, each unit being a voxel and corresponding to a parameter vector. This allows each location in the space to be uniquely represented by a three-dimensional index, thus creating conditions for subsequent signal-space mapping and sparsity regularization. Without voxel-based modeling, it is impossible to accurately locate the probability distribution of defects in three-dimensional space, and it is also impossible to generate subsequent probability voxel maps.

[0198] In this embodiment, the virtual model coordinate system is first imported from the CAD 3D appearance drawing and calibrated with the process coordinate system of the forging equipment to ensure that the spatial coordinates correspond one-to-one with the actual workpiece.

[0199] Taking valve tappets as an example, in the virtual model coordinate system, a voxel domain covering the entire forging volume is established with the geometric shape of the valve tappet as the boundary. The choice of voxel size needs to balance computational accuracy and computational cost: if the voxel is too large, it may lead to inaccurate defect localization, with multiple defects aggregating into one voxel; if the voxel is too small, it will result in an excessively large observation matrix, causing the inversion calculation to fail to converge. In this embodiment, the voxel side length is determined jointly based on the minimum geometric features of the forging and the sensor resolution.

[0200] Furthermore, a voxel parameter vector is assigned to each voxel to describe whether the voxel has defects and the intensity of those defects. Initially, all voxel parameters are set to zero, representing a defect-free state. During the subsequent inversion process, these parameters are continuously updated based on the observed signals, eventually converging into a probability distribution.

[0201] It is understandable that voxel parameters not only include the probability of defect existence, but can also be extended to multi-dimensional attributes, such as crack direction, pore size, and damage degree. However, in the defect prediction scenario of this application, the core parameter is the defect probability value, so as to directly correspond to the energy propagation model of the Green's function.

[0202] S3.4.2: Based on the geometric mapping relationship of the three-dimensional voxel mesh and the time-varying Green's function set, generate an observation matrix with channels as rows and phase indices as columns, and calculate the weight matrix based on the noise variance of each channel in the observation vector;

[0203] Specifically, the observation matrix essentially describes the contribution relationship of a voxel unit to the response of each sensing channel under the current process phase. It is the result of combining the Green's function and voxel geometry mapping. In other words, each row of the observation matrix corresponds to a sensing channel, and each column corresponds to a voxel unit under a phase index. The matrix elements reflect the influence intensity of the voxel unit on the signal along a specific excitation propagation path.

[0204] In this embodiment, based on the geometric mapping relationship of the three-dimensional voxel mesh, the position of each voxel unit is matched with the propagation path in the Green's function library. Specifically, this includes:

[0205] Using the geometric center of the voxel as a reference point, the spatial distance and propagation direction from it to each sensor location are calculated, and the corresponding response function is found in the time-varying Green's function set. This response function has been interpolated and adjusted in time and amplitude, thus reflecting the true propagation characteristics under the current operating conditions. The peak amplitude or energy density of this response function is used as a matrix element. The coupling strengths of all voxel units with all channels and all phases are combined to form a complete observation matrix.

[0206] Furthermore, during the preprocessing stage, the noise level of each channel was obtained through residual statistics. The reciprocal of the noise variance of each channel was used as the weight coefficient for that channel, constructing a diagonal weight matrix. Channels with low noise and high signal-to-noise ratio (SNR) receive higher weights during the inversion process, while channels with high noise and low SNR are relatively weakened, thereby improving the robustness of the overall solution. In addition, if the noise variance of a channel exceeds a preset threshold, that channel can be marked as a failed channel and set to zero in the weight matrix to prevent it from having a destructive impact on the results.

[0207] S3.4.3: Input the observation vector, observation matrix, and weight matrix into the weighted least squares inversion model. Perform sparse regularization on the observation vector, observation matrix, and weight matrix through the weighted least squares inversion model, and solve it through the alternating direction multiplier method to obtain the voxel parameter estimate. The sparse regularization includes L1 norm constraint on the voxel parameter vector, group sparse constraint obtained by segmenting according to concentric rings and axial direction, and total variational constraint on the discrete gradient of each pair of voxel parameter vectors.

[0208] Specifically, relying solely on weighted least squares fitting of observational data often results in unstable voxel solutions under conditions of strong noise, underdeterminedness, or ill-conditioned states, manifesting as pseudo-ripples, energy diffusion, or large-scale misjudgment regions. Internal defects in forgings typically exhibit spatial characteristics of being few in number, clustered, and having clearly defined boundaries, and are closely related to the geometric commonalities of automotive forgings; for example, valve tappets:

[0209] They are distributed in concentric rings along the circumference of the cylindrical shape and relatively concentrated in several functional sections along the axial direction.

[0210] Understandably, without embedding such prior constraints into the inversion process, it is difficult to suppress noise contamination of the solution and to effectively concentrate limited measurement information on truly suspect voxels. Therefore, when constructing the inversion objective, three complementary priors need to be introduced in addition to the data fitting term:

[0211] One is to impose sparse constraints on the overall voxel parameters, so as to cause the parameters in non-defect areas to automatically shrink to zero, while retaining a small number of high-response voxels;

[0212] Second, based on the group sparsity constraint of concentric rings and axial segments, it is used to encourage voxels in the same ring or the same axial segment to be activated or inhibited together in the form of groups, reflecting the structural prior of forging geometry and forming streamline.

[0213] Third, the total variational constraint on the discrete gradient is used to punish drastic jumps between voxels, preserving the true defect boundary while suppressing the checkerboard effect and high-frequency ringing.

[0214] Because the aforementioned priors are separable and non-smooth, direct solutions encounter the dual difficulties of non-differentiability and large-scale constraints. Therefore, a numerical framework that is separable, parallelizable, and convergently controllable is required. The alternating direction multiplier method can decouple the data terms and the three types of priors into several easily solvable subproblems through variable splitting. These subproblems are updated alternately in each iteration and are coordinated by the multiplier terms, making it suitable for large-scale problems involving multiple channels and phases. Furthermore, the alternating direction multiplier method allows the proximal updates of different priors to be written as closed or semi-closed thresholding operations, resulting in low computational cost and parallelization to voxel blocks or channel clusters, meeting the timeliness requirements of forging operations.

[0215] In this embodiment, the observation matrix is ​​first scaled column-wise and the weights are pre-whitened to make different channels comparable on the same scale, thus preventing a high-gain channel from dominating the direction of the solution. Then, variable splitting is performed: three auxiliary variables are introduced for the voxel parameters, corresponding to overall sparsity, grouped sparsity, and total variational constraints, respectively, and corresponding consistency constraints are set.

[0216] Furthermore, in each iteration, the voxel main variable is first updated under the weighted data terms, that is, a linear subproblem with positive definite structure is solved given the current multipliers and auxiliary variables. The linear subproblem is solved using pre-factoring or conjugate gradient iteration, combined with a preconditioner to suppress ill-conditioned features and improve the convergence speed.

[0217] Furthermore, a soft thresholding update is performed on the overall sparse auxiliary variables, causing a large number of small voxels to shrink rapidly. For the grouped sparse auxiliary variables, a group threshold update is performed according to the group definition of concentric rings and axial segments, causing voxels within the same group to exhibit cooperative activation and inhibition behavior. The definitions of the rings and axial segments are derived from the valve tappet shape and streamlines, and the group boundaries match the geometric transitions. For the total variational auxiliary variables, an anisotropic or isotropic threshold update is performed on the three-dimensional discrete gradient field, preserving defect edges while eliminating blocky noise and checkerboard artifacts. Subsequently, the three multipliers are updated to enhance consistency before proceeding to the next iteration.

[0218] Furthermore, in terms of iterative control, a dual-threshold criterion of original residuals and dual residuals is used to determine shutdown. To avoid the fluctuation between underfitting and overfitting caused by early shutdown, a relative target descent rate is introduced as a supplementary criterion. If the descent rate is less than the preset proportion for several consecutive rounds and the residuals are all within the tolerance, then convergence is determined. The penalty parameter adopts an adaptive strategy: when the original residual is significantly larger than the dual residual, the penalty is appropriately increased to strengthen consistency; when the dual residual is too large, the penalty is appropriately reduced to avoid over-constraint.

[0219] In some optional implementations, this embodiment provides several enhancement strategies to adapt to extreme operating conditions and heterogeneous noise scenarios:

[0220] Firstly, robust losses are introduced into the data items to suppress the dominant role of outliers when transient shocks or contact noises occur in large residual channels.

[0221] Secondly, channel consistency regularization is added outside the weight matrix to suppress or amplify the single-channel update results using the coherence coefficient of the neighboring channels, thus resisting single-point drift.

[0222] Third, it supports a hierarchical structure in the definition of sparse groups, with the outer layer being concentric rings and the inner layer being axial segments or local feature bands, thereby achieving joint constraints from the whole to the local at different scales. It can capture the ring-like properties along the circumference and also distinguish them finely in the key axial segments.

[0223] Fourth, directional weights are used in the total variation to apply different intensities to the gradients along the axial and radial directions to match the anisotropy of material streamlines and heat flow directions, which avoids excessive smoothing of the real crack tip and suppresses radial noise stripes.

[0224] Fifth, an energy-conserving step-size adaptive strategy is introduced in the penalty parameter update to avoid oscillations caused by parameter jumps;

[0225] Sixth, perform sliding window re-optimization under long-term operation, that is, take the observations and solutions of the most recent several cycles as a set, and perform a small-scale refinement solution to correct the systematic deviation caused by the slow drift of temperature and lubrication conditions.

[0226] In other computationally limited implementations, embodiments of this application provide two levels of approximation:

[0227] First-level approximation preserves data items and overall sparsity, TV constraints, and removes group sparsity to obtain a fast feasible solution;

[0228] The second-level approximation refines the high-confidence region voxel blocks only by sparsifying the voxel blocks, thus strictly controlling the computational cost and time overhead.

[0229] S3.4.4: Based on the voxel parameter estimation, calculate the voxel-level defect probability value through probability mapping, and update and replace the voxel parameter vector with the voxel-level defect probability value, wherein the probability mapping is obtained through logistic regression;

[0230] S3.4.5: Based on the registration residuals obtained during the solution process of the weighted least squares inversion model, calculate the observation noise covariance estimate, and then perform residual resampling on the observation vector using the observation noise covariance estimate to obtain the sampling distribution;

[0231] S3.4.6: Determine the upper and lower confidence bounds for each voxel based on the significance level of the sampling distribution, and output the defect probability voxel plot and defect confidence interval;

[0232] The mathematical expression of the weighted least squares inversion model is as follows:

[0233] ,

[0234] in, Represents the weight matrix. This indicates finding the minimum value. Let represent the observation vector, and represent the voxel parameter vector. Represents the observation matrix. Represents the phase state vector. This indicates the registration time index. , and The weighting parameter is calculated using the channel noise level and the sweep bandwidth. This represents the weighted norm 2 terms. Voxel parameter vector L1 norm regularization, Indicates sparse regularization of groups. Represents a set of voxel grouping indexes. Indicates the grouping index of a single voxel. Indicates the first The voxel parameter vector of voxel grouping, Indicates the first Group weights for individual element groupings This indicates the total variation regularity. The gradient represents the voxel parameter vector. This represents the upper bound of the voxel parameter vector.

[0235] In one example, the specific steps of S4 are as follows:

[0236] S4.1: Based on the preset service load and boundary condition library, set the time domain analysis step size and sub-step control parameters corresponding to the process phase;

[0237] S4.2: Map the defect probability voxel map and defect confidence interval to the finite element mesh through the mapping operator to obtain the initial defect state set, wherein the initial defect state set includes the void volume fraction and the equivalent initial crack size;

[0238] S4.3: Taking the initial temperature field and residual stress field of the standard automotive forging in the corresponding process phase as the starting state, the finite element mesh is transiently solved under the constraints of the time domain analysis step size and sub-step control parameters to obtain a time history set including stress, strain, temperature and hydrostatic pressure.

[0239] S4.4: Update the initial defect state set step by step according to the preset update law based on the time history set, and output the defect risk density.

[0240] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A defect prediction system for automotive forgings based on digital twins, characterized in that, The system includes: The model building module is used to acquire the appearance data of automotive forgings and to build a virtual model and Green's function library based on the appearance data. The signal acquisition module is used to perform narrow-band frequency sweep micro-excitation on the automotive forging when the forging process of the automotive forging is in a preset process phase, and to acquire the response signal under the micro-excitation. The probability prediction module is used to perform time-varying registration of the response signal with the Green function library, reconstruct the defect probability voxel map inside the automotive forging through a sparse constraint inversion algorithm, and output the defect confidence interval. The defect assessment module is used to simulate the defect propagation trend of automotive forgings based on the defect probability voxel map and the virtual model, combined with the defect confidence interval, to obtain the defect risk density. The model building module includes a function library building unit, which is configured as follows: The contact area between the blank and the mold is marked in the virtual model, and the parameters of the contact area are set, including the friction coefficient, thermal contact resistance and lubrication condition parameters, which are used to represent the time-varying boundary conditions of the contact area during the forging process. The interface unit between the blank and the mold is discretized within the contact area, and a preset disturbance load is applied to the discretized interface unit. The displacement response, strain response, and acoustic response of the disturbance load propagating to the external sensing point in the virtual model are calculated using the finite element time-domain solution method. Based on the displacement response, strain response, and acoustic response, the response function is calculated in combination with time-varying boundary conditions. The response function is used to establish a multi-index mapping table based on the interface unit position, sensor point number and process phase to obtain the original library of Green's function; The original Green's function library is subjected to matrix decomposition and sparse basis vector extraction, and compressed and mapped into a low-dimensional sparse dictionary; In the low-dimensional sparse dictionary, a corresponding sub-dictionary is established for different process phases, and the sub-dictionary is bound and stored with the virtual model to obtain the Green function library; The probability prediction module includes an inversion unit, which is configured as follows: Define a voxel domain in the coordinate system of the virtual model and generate a three-dimensional voxel mesh, and assign corresponding voxel parameter vectors in the three-dimensional voxel mesh; Based on the geometric mapping relationship of the three-dimensional voxel mesh and the time-varying Green's function set, an observation matrix is ​​generated with channels as rows and phase indices as columns, and a weight matrix is ​​calculated based on the noise variance of each channel in the observation vector. The observation vector, observation matrix, and weight matrix are input into the weighted least squares inversion model. The observation vector, observation matrix, and weight matrix are sparsely regularized by the weighted least squares inversion model and solved by the alternating direction multiplier method to obtain the voxel parameter estimates. The sparse regularization includes L1 norm constraints on the voxel parameter vector, group sparse constraints obtained by segmenting based on concentric rings and axial direction, and total variational constraints on the discrete gradients of each pair of voxel parameter vectors. Based on the voxel parameter estimation, the voxel-level defect probability value is calculated through probability mapping, and the voxel parameter vector is updated and replaced with the voxel-level defect probability value, wherein the probability mapping is obtained through logistic regression. Based on the registration residuals obtained during the solution process of the weighted least squares inversion model, the observation noise covariance estimate is calculated. The observation vector is then resampled using the observation noise covariance estimate to obtain the sampling distribution. Based on the significance level of the sampling distribution, determine the upper and lower confidence bounds for each voxel, and output the defect probability voxel map and defect confidence interval.

2. The automotive forging defect prediction system based on digital twins according to claim 1, characterized in that, The appearance data includes the three-dimensional appearance drawing data of the automotive forging. The model building module includes a virtual model building unit, which is used to generate a virtual model based on the appearance data and material properties. The function library building unit is used to establish contact conditions in the virtual model and calculate the disturbance response to generate a Green's function library. The virtual model building unit is configured as follows: Based on the three-dimensional appearance drawing data, combined with the preset blank length-to-diameter ratio parameters, mold geometric parameters and forging process parameters, an initial three-dimensional geometric model consistent with the size of the automobile forging is generated. The initial three-dimensional geometric model is assigned values ​​based on the material property parameters of the automotive forging to obtain a virtual model. The material property parameters include thermal conductivity, specific heat capacity, coefficient of thermal expansion, yield strength, flow stress function, and damage evolution parameters.

3. The automotive forging defect prediction system based on digital twins according to claim 1, characterized in that, The process phase refers to a specific stage in the forging process of automotive forgings, which includes the pre-pressing phase, plastic deformation phase, final pressing phase, and constant pressing phase.

4. The automotive forging defect prediction system based on digital twins according to claim 1, characterized in that, The signal acquisition module includes an excitation unit and a signal acquisition unit. The excitation unit is used to apply narrow-band frequency sweep micro-excitation to the automotive forging at a preset process phase. The signal acquisition unit is used to acquire the response signal through a sensor. The excitation unit is configured as follows: At the arrival time of the preset process phase, a phase state vector is constructed based on the measured values ​​of slide displacement, slide speed, mold temperature and contact resistance, wherein the slide is the slide of the forging equipment for the automotive forging. The excitation feasible region of the phase state vector is determined by a preset mapping rule, wherein the excitation feasible region includes the injection time window, the upper limit of the excitation amplitude, the candidate interval of the sweep center frequency, and the upper limit of the sweep bandwidth. The excitation feasible region is corrected based on the center frequency of the corresponding process phase in the obtained historical forging cycle to obtain the frequency sweep sequence; The frequency sweep sequence is superimposed along the coaxial direction of the forming load of the automotive forging and injected into the slide drive channel.

5. The automotive forging defect prediction system based on digital twins according to claim 4, characterized in that, The excitation unit is configured with correction logic, which includes: Using the center frequency as the initial value, and according to the preset screening threshold, the trial frequency pair corresponding to the initial value is selected in the candidate interval of the sweeping center frequency in the excitation feasible domain to obtain the pilot single-frequency excitation. The analog response signal corresponding to the pilot single-frequency excitation is obtained by simulation inversion, and the analog response signal is phase-sensitively demodulated to obtain the amplitude index corresponding to the test frequency pair. The amplitude index is iterated through extreme value optimization update to calculate the optimized center frequency value, wherein the optimized center frequency value is located within the candidate interval of the swept center frequency; Based on the optimized center frequency value and the phase state vector, the sweep bandwidth, frequency stepping strategy, and amplitude envelope are determined, and a sweep sequence is generated. During the iterative calculation process, the iteration stops when the rate of change of the contact resistance or the rate of change of the slide velocity exceeds a preset threshold.

6. The automotive forging defect prediction system based on digital twins according to claim 1, characterized in that, The probability prediction module includes a registration unit, which performs time-varying registration of the response signal and the Green's function library to obtain a registered observation vector. The inversion unit performs inversion calculations including sparse constraints based on the observation vector and the virtual model to reconstruct a defect probability voxel map and output a defect confidence interval. The registration unit is configured as follows: The response signal is synchronously preprocessed to obtain a preprocessed response, wherein the synchronous preprocessing includes time base alignment based on ram displacement triggering, DC drift removal, bandpass filtering, and full window processing; Based on the phase state vector and process phase corresponding to the preprocessed response, the corresponding phase sub-library is obtained from the Green function library, and the Green functions in the phase sub-library are interpolated by time and amplitude to obtain a time-varying Green function set; Based on the time-varying Green's function set, the time offset and amplitude scaling factor of each channel are obtained through cross-correlation estimation. The preprocessed response is then time-shifted and amplitude normalized based on the time offset and amplitude scaling factor to obtain the registered observation vector, wherein the channel represents the spatial location of the preprocessed response.

7. The automotive forging defect prediction system based on digital twins according to claim 1, characterized in that, The mathematical expression of the weighted least squares inversion model is as follows: ， in, Represents the weight matrix. This indicates finding the minimum value. Let represent the observation vector, and represent the voxel parameter vector. Represents the observation matrix. Represents the phase state vector. This indicates the registration time index. , and The weighting parameter is calculated using the channel noise level and the sweep bandwidth. This represents the weighted norm 2 terms. Voxel parameter vector L1 norm regularization, Indicates sparse regularization of groups. Represents a set of voxel grouping indexes. Indicates the grouping index of a single voxel. Indicates the first The voxel parameter vector of voxel grouping, Indicates the first Group weights for individual element groupings This indicates the total variation regularity. The gradient represents the voxel parameter vector. This represents the upper bound of the voxel parameter vector.

8. The automotive forging defect prediction system based on digital twins according to claim 1, characterized in that, The defect assessment module is configured as follows: Based on the preset service load and boundary condition library, set the time domain analysis step size and sub-step control parameters corresponding to the process phase; The defect probability voxel map and defect confidence interval are mapped to the finite element mesh through a mapping operator to obtain the initial defect state set, wherein the initial defect state set includes the void volume fraction and the equivalent initial crack size; Starting with the initial temperature field and residual stress field of a standard automotive forging in the corresponding process phase, the finite element mesh is transiently solved under the constraints of the time domain analysis step size and sub-step control parameters to obtain a time history set including stress, strain, temperature and hydrostatic pressure. The initial defect state set is updated step by step according to a preset update law based on the time history set, and the defect risk density is output.

Images