Method for evaluating cooling effect of cooling runner of PET bottle preform mold

Abstract

The invention provides a method for evaluating the cooling effect of a cooling flow channel of a PET bottle preform mold, and the method comprises the steps: embedding an ultrasonic thickness measurement sensor array at a key position of the cooling flow channel, obtaining accurate wall thickness and surface roughness data through the combination of signal filtering and wavelet denoising, and achieving the reconstruction of micro-scale geometrical morphology based on a physical information neural network; according to the method, the thermal-mechanical coupling finite element model is dynamically corrected, thermal stress distribution and risk evolution are predicted through the reduced-order model and the long-short-term memory network, finally, a structured thermal stress evaluation report is output, real-time early warning of the mold failure risk is achieved, and the accuracy and prediction capacity of runner wear detection are improved.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring technology for cooling channels in PET preform molds, and particularly to a method for evaluating the cooling effect of cooling channels in PET preform molds. Background Technology

[0002] In the current field of PET preform mold cooling channel thermal stress prediction and cooling performance evaluation, the mainstream technical approach mainly focuses on thermo-mechanical coupling analysis based on finite element numerical simulation, and on evaluating the cooling effect in a single cycle or short period after acquiring temperature field data using embedded sensors or surface infrared imaging. In mold manufacturing and service management practice, typical solutions often use the theoretical channel geometry and initial material parameters from the design phase as the basis for thermal analysis, and combine them with predetermined thermal boundary conditions to achieve quantitative analysis of temperature and stress during the molding cycle through finite element modeling. These methods have a certain predictive accuracy in the new mold or short-cycle service phase and are well applicable to cooling channel design optimization, short-term fault diagnosis, and process parameter adjustment. However, as the mold service cycle progresses, fluid erosion and thermal fatigue wear inevitably occur on the inner wall of the cooling channel, leading to thinning of the wall thickness, increase in surface roughness, and decrease in heat conduction capacity, which in turn causes an undesirable evolution of local cooling intensity attenuation and stress concentration. Existing mold thermal stress prediction systems largely rely on static geometric models and material parameter configurations established during the initial service life. They lack a real-time state-sensing mechanism that updates with the mold's service life, failing to accurately reflect the actual wear state of the flow channels and its evolutionary impact on temperature and stress field distributions. While some technologies incorporate periodic temperature acquisition and strain monitoring, these are mostly used only for process monitoring and anomaly alarms, failing to establish a dynamic modeling and adaptive simulation framework for a closed-loop "wear-thermodynamic evolution-stress risk." Furthermore, traditional cooling channel wear analysis often employs offline disassembly measurements or periodic ultrasonic testing, resulting in low data acquisition frequency and an inability to cover all critical risk areas, making it difficult to support the need for continuous, high-resolution stress trend prediction. Currently, the thermo-mechanical analysis methods commonly used in the mold industry often assume that the flow channel wall thickness and surface condition remain unchanged throughout the entire service life. They are suitable for new mold sample verification, preliminary cooling strategy selection, and performance evaluation under standard operating conditions. However, they have significant limitations in terms of both prediction accuracy and decision reliability when it comes to fluid heat transfer deterioration caused by wear during long-term service, thermal stress-strain inversion, and remaining life assessment. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method for evaluating the cooling effect of the cooling channel of a PET preform mold.

[0004] The technical solution of this invention is implemented as follows: A method for evaluating the cooling effect of a cooling channel in a PET preform mold, comprising: S1: Deploy a miniature embedded ultrasonic thickness sensor array on the inner wall of the cooling channel of the PET preform mold, periodically collect the original signal and echo characteristic data of the channel wall thickness at multiple spatially distributed sampling points, so as to obtain the time-series observation value of the local geometric state of the cooling channel; S2: Bandpass filtering and wavelet denoising are performed on the acquired raw ultrasonic signals. Based on the reflection characteristics of different material interfaces, the effective echoes on the inner surface of the flow channel are separated, and their flight time and amplitude attenuation characteristics are extracted to generate the clean wall thickness measurement value and surface roughness proxy parameters corresponding to each sampling point. S3: Based on the clean wall thickness measurement values ​​and surface roughness proxy parameters of multiple sampling points, combined with the fluid erosion empirical model and heat transfer boundary condition inversion algorithm, a microscale morphology reconstruction model is constructed. The physical information neural network (PINN) is used to drive the model to generate the actual three-dimensional geometric topology data of the cooling channel under the current service stage. S4: Input the actual three-dimensional geometric topology data into the dynamic geometry-thermodynamic coupling modeling process to replace the initial design geometric model, and update the boundary conditions of the material thermal conductivity and convective heat transfer coefficient according to the surface roughness proxy parameters to form a time-varying thermo-mechanical coupling finite element calculation domain that reflects the wear evolution state. S5: Perform transient temperature field simulation based on the time-varying thermo-mechanical coupled finite element calculation domain, output the unsteady temperature distribution cloud map in each forming cycle, and input the temperature field result as thermal load into the pre-trained reduced-order model (ROM). The reduced-order model is jointly trained by historical full-size simulation data and measured strain data, and is used to quickly calculate the corresponding thermal stress field distribution. S6: Using the thermal stress field distribution data output by the reduced-order model, extract the thermal stress peak sequence at key locations, and combine it with the wall thickness decay sequence at the corresponding time to construct a wear evolution memory bank. Based on the long short-term memory network (LSTM), learn the nonlinear dynamic mapping relationship between the two to generate a thermal stress evolution trend prediction model. S7: Based on the thermal stress evolution trend prediction model, predict the thermal stress field distribution in several future forming cycles, and determine whether there are spatial regions in the prediction results that exceed the material yield strength threshold. If so, mark them as thermal stress risk hot spots. S8: Generates a dynamic thermal stress assessment report containing the current thermal stress field distribution, risk hotspot area identification, and remaining safety period estimate, and outputs it to the process optimization system to trigger cooling strategy adjustments or mold maintenance early warning actions.

[0005] The present invention provides a method for evaluating the cooling effect of a cooling channel in a PET preform mold, which has the following beneficial effects: (1) This invention integrates a miniature embedded ultrasonic thickness sensor array into the inner wall of the cooling channel of a PET preform mold, realizing online and non-invasive sensing of wall thickness changes and inner surface roughness evolution, effectively overcoming the limitations of existing technologies that rely on periodic shutdown inspections or static simulations based on initial design models. Combined with edge computing nodes, real-time acquisition and preliminary processing of local wear information can be completed without interrupting the production cycle, significantly improving the temporal resolution and spatial coverage of condition monitoring. Furthermore, a microscale morphology reconstruction module driven by a physical information neural network (PINN) is introduced, which integrates the fluid erosion empirical model and the heat transfer boundary condition inversion algorithm. It can reconstruct the three-dimensional geometric topology of the cooling channel under the current service state with high fidelity from limited measurement point data, solving the morphology distortion problem caused by measurement blind spots and providing a real and reliable input basis for subsequent simulation analysis. (2) This invention uses a pre-trained reduced-order model (ROM) jointly trained by historical full-size simulation data and measured strain data to replace the traditional complete solution process. While ensuring the accuracy of the thermo-mechanical coupling field distribution prediction, it significantly reduces the simulation time to hundreds of milliseconds, effectively supporting the deployment and implementation of high-frequency periodic evaluation tasks. The reconstructed dynamic geometric model is automatically imported into the simulation engine, and key boundary parameters affected by surface degradation, such as the thermal conductivity and convective heat transfer coefficient of the material, are updated synchronously. This realizes the leap from "static design model" to "dynamic digital twin", significantly enhancing the spatial accuracy and temporal consistency of temperature field and thermal stress field prediction. On this basis, a wear evolution memory library containing the wall thickness decay curve and the evolution trend of thermal stress peak is constructed. The nonlinear correlation law is mined using a long short-term memory network (LSTM) to extrapolate the trend of thermal stress level in several future molding cycles, thereby identifying potential failure risk areas in advance and outputting a dynamic thermal stress assessment report and hot spot warning with spatiotemporal resolution, which greatly enhances the foresight and decision reliability of mold health management. (3) This invention, through a closed-loop architecture of "perception-modeling-simulation-prediction," for the first time deeply integrates the physical degradation mechanism of the mold during its service life into the cooling performance evaluation system, significantly improving the long-term stability and engineering applicability of the evaluation system. At the same time, the entire process does not require manual intervention for modeling updates, has a good degree of automation and scalability, and is suitable for rapid adaptation and batch deployment of multiple mold models. The resulting dynamic thermal stress evaluation results can not only be used to guide the optimization of cooling process parameters and extend the mold life, but also serve as a key input for equipment health status feedback in intelligent manufacturing systems, supporting the formulation of predictive maintenance strategies and promoting the digital transformation of injection molding production towards high precision, high efficiency, and high reliability. Attached Figure Description

[0006] Figure 1This is a flowchart of a method for evaluating the cooling effect of a cooling channel in a PET preform mold according to the present invention; Figure 2 This is a sub-flowchart of a method for evaluating the cooling effect of a PET preform mold cooling channel according to the present invention; Figure 3 This is another sub-flowchart of the method for evaluating the cooling effect of a PET preform mold cooling channel according to the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, this invention provides a method for evaluating the cooling effect of a cooling channel in a PET preform mold, specifically including: S1: Deploy a miniature embedded ultrasonic thickness sensor array on the inner wall of the cooling channel of the PET preform mold, periodically collect the original signal and echo characteristic data of the channel wall thickness at multiple spatially distributed sampling points, so as to obtain the time-series observation value of the local geometric state of the cooling channel; S2: Bandpass filtering and wavelet denoising are performed on the acquired raw ultrasonic signals. Based on the reflection characteristics of different material interfaces, the effective echoes on the inner surface of the flow channel are separated, and their flight time and amplitude attenuation characteristics are extracted to generate the clean wall thickness measurement value and surface roughness proxy parameters corresponding to each sampling point. S3: Based on the clean wall thickness measurement values ​​and surface roughness proxy parameters of multiple sampling points, combined with the fluid erosion empirical model and heat transfer boundary condition inversion algorithm, a microscale morphology reconstruction model is constructed. The physical information neural network (PINN) is used to drive the model to generate the actual three-dimensional geometric topology data of the cooling channel under the current service stage. S4: Input the actual three-dimensional geometric topology data into the dynamic geometry-thermodynamic coupling modeling process to replace the initial design geometric model, and update the boundary conditions of the material thermal conductivity and convective heat transfer coefficient according to the surface roughness proxy parameters to form a time-varying thermo-mechanical coupling finite element calculation domain that reflects the wear evolution state. S5: Perform transient temperature field simulation based on the time-varying thermo-mechanical coupled finite element calculation domain, output the unsteady temperature distribution cloud map in each forming cycle, and input the temperature field result as thermal load into the pre-trained reduced-order model (ROM). The reduced-order model is jointly trained by historical full-size simulation data and measured strain data, and is used to quickly calculate the corresponding thermal stress field distribution. S6: Using the thermal stress field distribution data output by the reduced-order model, extract the thermal stress peak sequence at key locations, and combine it with the wall thickness decay sequence at the corresponding time to construct a wear evolution memory bank. Based on the long short-term memory network (LSTM), learn the nonlinear dynamic mapping relationship between the two to generate a thermal stress evolution trend prediction model. S7: Based on the thermal stress evolution trend prediction model, predict the thermal stress field distribution in several future forming cycles, and determine whether there are spatial regions in the prediction results that exceed the material yield strength threshold. If so, mark them as thermal stress risk hot spots. S8: Generates a dynamic thermal stress assessment report containing the current thermal stress field distribution, risk hotspot area identification, and remaining safety period estimate, and outputs it to the process optimization system to trigger cooling strategy adjustments or mold maintenance early warning actions.

[0010] Step S1: Deploy a miniature embedded ultrasonic thickness sensor array on the inner wall of the cooling channel of the PET preform mold, and periodically collect the original signal and echo characteristic data of the channel wall thickness at multiple spatially distributed sampling points to obtain time-series observations of the local geometric state of the cooling channel. Specifically, this includes: S1.1: Based on the spatial structure parameters of the cooling channel of the PET preform mold and the distribution map of the thermo-mechanical coupling sensitive area, determine the set of key monitoring locations where sensors need to be deployed, use the mold CAD model to perform three-dimensional path planning, and generate a sampling point layout scheme that meets the minimum coverage redundancy and the maximum heat-affected zone coverage to ensure that the collected data is representative and spatially scalable. Based on the spatial structural parameters of the cooling channel of the PET preform mold and the distribution map of the thermo-mechanical coupling sensitive area, the regional sensitivity analysis method (parameters: thermal stress gradient distribution matrix, temperature field mean square error) is used to quantify the force-thermal response characteristics of the mold structure and extract the corresponding set of high-sensitivity locations. Furthermore, by using a three-dimensional geometric feature extraction algorithm (parameters: cooling channel curvature radius, cross-sectional shape factor, path connectivity index), the sensitive location is geometrically projected and mapped to the geometric centerline of the cooling channel, and a set of three-dimensional coordinates of the candidate monitoring area is obtained. Furthermore, a 3D path planning algorithm based on the mold CAD model (parameter: maximum connectivity distance between nodes) is used. mm, maximum spacing between sampling points (mm), to realize the feasibility search of sensor layout paths and generate multiple spatial path schemes covering candidate monitoring areas; Furthermore, a redundancy constraint optimization algorithm (objective function: minimum coverage redundancy, constraint condition: maximum heat-affected zone coverage > 0.9) is used to screen and optimize path schemes and generate an optimal sampling point layout matrix that meets spatial scalability requirements. By using layout matrix visualization and spatial coordinate output processing, the results of the previous step are transformed into a set of key monitoring locations, ensuring that the collected data is representative and spatially scalable, thus achieving the pre-deployment preparation effect for sensor deployment. For example, in a certain PET preform mold, the length of the cooling channel is... mm, radius of curvature in mm to A distribution map of the thermo-mechanical coupling sensitive area within a range of mm is shown at a distance from the gate. The stress concentration factor at the intersection of mm and the parting line is significantly higher than in other regions. The root mean square error of the thermal stress gradient output by the regional sensitivity analysis is... MPa², corresponding to the number of highly sensitive location coordinates extracted. One. The mean cross-sectional shape factor was calculated from the three-dimensional geometric feature extraction. The path connectivity index reached The path planning algorithm limits the maximum connectivity distance between nodes. mm, maximum spacing between sampling points mm, retrieve path schemes that satisfy geometric constraints. The coverage of redundancy optimization calculations reaches [a certain percentage]. Finally selected Each sampling point is used as a key monitoring location, and a layout matrix is ​​generated for visualization verification in the CAD model. The output coordinate set is used in the next step of sensor embedding and deployment. In actual testing, this layout ensures full coverage of the heat-affected zone and supports the efficient execution of subsequent wall thickness time series observations. S1.2: Miniaturized ultrasonic thickness sensor units are embedded at each key monitoring location. The sensor unit integrates a piezoelectric chip, a temperature compensation module and a wireless transmission interface. It is fixed to the substrate material of the inner wall of the cooling channel using a micro-welding process to form a high-temperature resistant and electromagnetic interference resistant embedded sensor network to achieve non-invasive continuous monitoring. The input conditions for embedding miniaturized ultrasonic thickness sensor units at each key monitoring location are the sampling point layout scheme and its three-dimensional spatial coordinate set output from step S1.1, and each coordinate position is used as the installation positioning reference. Employing a precision piezoelectric wafer packaging process (parameters: piezoelectric material type PZT-5H, wafer diameter 6mm, thickness 0.5mm), the sensor achieves high-sensitivity detection of ultrasonic excitation and echo signals from the inner wall of the cooling channel. Furthermore, by integrating a temperature compensation module (parameters: NTC thermistor accuracy ±0.1℃, linearization circuit based on fourth-order polynomial fitting), real-time correction of the piezoelectric wafer resonant frequency drift under high-temperature service environment is achieved, and the compensated excitation frequency setting value is obtained. Furthermore, through an embedded wireless transmission interface (parameters: operating frequency band 2.4GHz, modulation method GFSK, transmission rate 1Mbps), short-range high-speed transmission of sensor thickness measurement data is achieved, and digital signal packets that can be received by edge computing nodes are generated. Furthermore, a micro-welding fixing method (parameters: laser micro-welding energy 0.8J, pulse width 1ms, weld diameter 0.3mm) is adopted to achieve a firm connection between the sensor shell and the substrate material of the cooling channel inner wall, thereby obtaining a high-temperature resistant and vibration-resistant embedded mounting structure; Furthermore, an electromagnetic interference-resistant network topology is constructed between sensor nodes (parameters: shielding braided layer coverage of 95%, grounding impedance of less than 0.1Ω), forming a spatially distributed signal acquisition channel, and generating a network structure diagram with redundant routing. Through the above-mentioned encapsulation, compensation, communication and fixing methods, the installation position parameters of the previous step are transformed into an operational embedded sensor network, realizing the expected technical effect of non-invasive continuous monitoring of cooling channel wall thickness changes and surface state evolution. For example, in a 24-cavity PET preform mold cooling channel layout, a total of 48 key monitoring locations were selected for sensor unit embedding, with a spacing of approximately 20mm between each two sampling points. The sensor piezoelectric wafer uses PZT-5H material, with a nominal resonant frequency of 4MHz. Combined with an NTC thermistor, the real-time temperature fluctuation range of the inner surface of the channel during the injection molding cooling stage is measured from 82℃ at the signal acquisition starting point to the lowest cooling temperature of 32℃. The temperature compensation module calls a fourth-order polynomial compensation function based on the measured temperature value to ensure that the excitation frequency remains stable within the target range. The wireless interface is configured in this scenario as a 2.4GHz band with GFSK modulation, and the actual air interface rate is measured to be 0.96Mbps, meeting the requirement of transmitting all sampled data in a single cycle within 40ms. During the micro-welding fixing process, each sensor shell is welded to the mold steel wall to form three equidistant weld points. The measured laser power is 0.8J, and the shear strength of the weld points is tested to be greater than 200N, ensuring no detachment under high-pressure cyclic vibration during injection molding. The network shielding braided layer uses tin-plated copper wire, with a measured braiding coverage of 95%. The node grounding impedance is below 0.08Ω. Electromagnetic interference testing in the laboratory showed that, under conditions of cooling water pump operation and high-frequency heater activation, the signal bit error rate remained at [value missing]. This level of performance significantly improves signal reliability and monitoring accuracy. S1.3: Configure an edge-triggered acquisition mechanism, start the data acquisition timing based on the injection molding cycle synchronization signal, control each sensor unit to emit high-frequency ultrasonic pulses at the beginning of the cooling stage of each molding cycle, and receive the original echo signal from the mold steel-cooling medium interface to generate an original signal sequence containing flight time, amplitude attenuation and waveform distortion characteristics. S1.4: Perform analog-to-digital conversion and timestamp binding processing on the acquired raw echo signal to convert the analog voltage signal into a digital sampling sequence and associate it with the corresponding spatial coordinates and period number to generate a raw wall thickness observation data package with spatiotemporal labels, which serves as the input condition for subsequent signal purification and feature extraction. S1.5: Transmit the raw wall thickness observation data packets with spatiotemporal tags to the local edge computing node buffer area via an industrial-grade wired / wireless communication bus to complete data collection and preliminary verification, and output a structured raw observation dataset that can be used for the next processing stage, ensuring data integrity and temporal consistency.

[0011] Step S2: The acquired raw ultrasonic signal is subjected to bandpass filtering and wavelet denoising processing. Based on the reflection characteristics of different material interfaces, the effective echo of the inner surface of the flow channel is separated, and its flight time and amplitude attenuation characteristics are extracted to generate the cleaned wall thickness measurement value and surface roughness proxy parameter corresponding to each sampling point. Specifically, this includes: S2.1: Acquire the raw ultrasonic signal periodically collected by the miniature embedded ultrasonic thickness sensor array on the inner wall of the cooling channel of the PET preform mold. This signal contains time-domain echo data from multiple spatially distributed sampling points, which are mixed with high-frequency electronic noise, multipath interference and environmental vibration coupling components. Based on the preset acoustic impedance difference characteristics of the cooling medium-metal interface, determine the theoretical arrival interval of the target echo in the time domain to limit the effective signal analysis window. The input object is the original ultrasonic signal dataset periodically collected by the miniature embedded ultrasonic thickness sensor array on the inner wall of the cooling channel of the PET preform mold. The dataset contains time-domain echo sequences at multiple spatially distributed sampling points. A noise source characteristic analysis method (parameters: sampling frequency 10MHz, analysis time 1ms) is used to estimate the power spectrum of high-frequency electronic noise characteristics and mark the energy concentration interval in the frequency domain to guide the subsequent window selection. Furthermore, by using a signal path modeling method (parameters: sound velocity of mold steel 5900m / s, sound velocity of cooling medium 1500m / s), the propagation delay of multipath interference components is calculated, and delay distribution characteristics are generated to identify the time separation boundary between the target echo and the interference echo. Furthermore, by using a mechanical vibration coupling analysis method (parameter: vibration frequency < 0.5 MHz), the amplitude of low-frequency fluctuations introduced by environmental vibration is detected, and signal segments that do not overlap with the arrival time of the effective echo are screened out. The acoustic impedance difference calculation method is used, based on the acoustic impedance values ​​of the cooling medium-metal interface, respectively. (in For the density of the medium, (for the speed of sound), to achieve the reflection coefficient The calculation, in which The acoustic impedance of the cooling medium, The acoustic impedance of the mold steel is used to determine the theoretical flight time range of the target echo. By using a time-domain windowing method (parameters: window center = median theoretical flight time, window width = ±2σ), the original echo data is cropped to an effective analysis window containing only the target echo, forming a time-domain subsequence with high signal-to-noise ratio separation effect; The above algorithm transforms the original signal result from the previous step into an effective echo analysis window limited to the theoretical arrival interval, achieving time-domain separation of the target echo and interference noise, and providing high-purity initial data for subsequent bandpass filtering and wavelet denoising. For example, in the monitoring of the cooling channel of a PET preform mold, the sensor sampling frequency is configured to 10MHz, and the length of the original echo signal collected at each sampling point is 10,000 points. Power spectrum estimation was used to detect high-frequency electronic noise concentrated above 8MHz. Combined with the acoustic impedance values ​​of the mold steel sound velocity (5900m / s) and the cooling medium sound velocity (1500m / s), the reflection coefficient was calculated as follows: The theoretical flight time of the target echo is between 3.4 μs and 3.6 μs, with a standard deviation σ of 0.05 μs. The original signal is truncated to a time-domain window centered at 3.5 μs with a width of ±0.1 μs, eliminating components with multipath delays >4.0 μs and mechanical vibrations <0.5 MHz. Results show that the effective echo peak amplitude of the truncated time-domain subsequence is approximately three times higher than the original signal, demonstrating significant noise suppression and providing a stable high signal-to-noise ratio input for downstream S2.2 bandpass filtering. S2.2: The original ultrasonic signal is subjected to bandpass filtering within the time domain analysis window. A Butterworth fourth-order digital filter is used, with the passband range set to 1MHz to 5MHz, in order to suppress low-frequency mechanical vibration interference below 1MHz and high-frequency electromagnetic crosstalk components above 5MHz, to obtain an intermediate signal sequence with preliminary signal-to-noise ratio improvement, which serves as the input condition for the next stage of wavelet denoising. S2.3: Based on the intermediate signal sequence, perform Discrete Wavelet Transform (DWT), select the db6 wavelet basis function and implement five-level decomposition, apply an adaptive soft threshold denoising algorithm to the detail coefficients of each level, dynamically adjust the threshold size according to the noise standard deviation estimate, and remove residual random noise; then perform wavelet reconstruction to generate a cleaned ultrasonic echo signal, which significantly enhances the clarity of the effective echo edge and the signal-to-noise ratio. S2.4: Based on the cleaned ultrasonic echo signal, identify the first effective echo peak value corresponding to the metal-coolant interface on the inner wall of the cooling channel, calculate its time of flight (ToF), and combine it with the calibrated value of the propagation speed of ultrasonic waves in the mold steel to calculate the instantaneous wall thickness measurement value at each sampling point; at the same time, analyze the amplitude attenuation of the echo, and use the empirical mapping relationship between amplitude attenuation and surface micro-irregularity to generate a surface roughness proxy parameter as a quantitative indicator reflecting the local wear state; S2.5: The instantaneous wall thickness measurement values ​​and surface roughness proxy parameters corresponding to each sampling point are spatiotemporally aligned and normalized to generate a structured clean wall thickness measurement value sequence and surface roughness proxy parameter vector, which serve as the input dataset for the downstream microscale morphology reconstruction model, ensuring that it has physical consistency and temporal continuity, and is used to drive the dynamic geometric topology reconstruction process.

[0012] like Figure 2As shown, step S3 involves: based on the cleaned wall thickness measurements and surface roughness proxy parameters from multiple sampling points, combining a fluid erosion empirical model and a heat transfer boundary condition inversion algorithm, constructing a microscale morphology reconstruction model, and using a Physical Information Neural Network (PINN) to drive this model to generate the actual three-dimensional geometric topology data of the cooling channel under the current service stage. Specifically, this includes: S3.1: Based on the clean wall thickness measurement and surface roughness proxy parameters of each sampling point output from the previous steps, combined with the mold cooling medium flow rate, temperature time series records and material hardness distribution data, the cumulative erosion depth evolution curve of each sampling area is calculated using the fluid erosion empirical model, and a local wear estimation sequence containing spatiotemporal decay characteristics is generated as the data-driven input for microscale geometric degradation modeling. Based on the clean wall thickness measurement values ​​and surface roughness proxy parameters of each sampling point output from the previous steps, a multivariate data binding method (parameters: spatial coordinate index, time label, wall thickness value, roughness value) is adopted to achieve multi-source association with mold cooling medium flow rate, temperature time sequence records and material hardness distribution data; An empirical model of fluid erosion was adopted (parameter: erosion coefficient). Using the medium density ρ, average medium velocity v, and material hardness H, the initial values ​​of the erosion depth per unit time at each sampling point are calculated, and the cumulative erosion curve over time is obtained. This erosion model satisfies the following relationship:

[0013] in, The depth of erosion per unit time. This is the empirical erosion coefficient. The density of the cooling medium, For local flow velocity, The velocity index is the flow rate index. The material hardness grade; Furthermore, by using the time series integration method (parameters: integration step size Δt, total number of cycles T), the cumulative erosion depth evolution curve of each sampling area within the actual service cycle is calculated, and a local wear estimation sequence containing spatiotemporal attenuation characteristics is obtained. Furthermore, by using a spatial mapping algorithm (parameter: interpolation weights of the three-dimensional coordinate system), the spatial correlation and full-field expansion of wear at different sampling points are realized, resulting in a spatiotemporal wear distribution matrix covering all monitoring areas; By smoothing and filtering the spatiotemporal decay data (parameter: three-dimensional Gaussian kernel σ), the discrete wear estimation sequence from the previous step is transformed into a continuous function field suitable for microscale geometric degradation modeling, thus realizing the differentiability of the input data in both spatial and temporal dimensions. By using a fluid erosion empirical model and a time-series integration processing method, the measurement and surface state parameter results from the previous step are transformed into a local wear estimation sequence with spatiotemporal continuity and physical consistency, thus realizing the data-driven input for microscale geometric degradation modeling. For example, in an evaluation scenario for a PET preform cooling mold, the density of the cooling medium is known to be set as follows: kg / m³, the local flow velocity was measured to be m / s, material hardness value HV, empirical erosion coefficient Flow rate index Within the service life, the length of each cycle is... min, let the integration step size Δt be... min, the total number of evaluation periods is Substituting into the formula, we can obtain the erosion depth per unit time:

[0014] The calculation result is mm / min. Using the time series integration method, in the total... The cumulative erosion depth obtained within the minimum service time is approximately mm, and generate time-series wear curves. Three-dimensional interpolation is applied to all monitoring points, followed by Gaussian smoothing (σ=0.5 mm) to obtain a spatially continuous wear distribution matrix. This matrix is ​​input into the microscale morphology degradation modeling module. Numerical verification shows that this input significantly improves the accuracy of geometric degradation field reconstruction and exhibits stable physical consistency in subsequent PINN-driven dynamic morphology prediction. S3.2: Based on the difference between the estimated local wear amount sequence and the original design geometric reference surface, an initial microscale morphology degradation field is constructed. The spatial covariance of the non-sampling area is estimated by Kriging interpolation, and a preliminary full-field three-dimensional geometric deviation distribution grid is generated to form a low-frequency morphology base reflecting the macroscopic wear trend. S3.3: Based on the heat transfer boundary condition inversion algorithm, the surface roughness proxy parameter at each sampling point is mapped to the equivalent heat transfer area reduction factor and the interface contact thermal resistance enhancement coefficient. The influence weight of the local temperature gradient is extracted through thermal response sensitivity analysis, and a spatially distributed thermal influence potential field is generated as the physical constraint condition for high-frequency detail reconstruction of the morphology. Based on the surface roughness proxy parameters of each sampling point obtained by S2 processing, the heat transfer boundary condition inversion algorithm (constraints: instantaneous recording of inlet temperature and flow rate of cooling medium, and reference field of thermal conductivity of mold steel) is used to map the roughness proxy parameters into an equivalent heat transfer area reduction factor, and obtain the preliminary heat transfer reduction field corresponding to the low-frequency morphology substrate. Furthermore, by using the same heat transfer boundary condition inversion algorithm (coupling parameters: roughness proxy parameter, solid-liquid interface acoustic impedance difference calculation results), the spatial mapping of the interface contact thermal resistance enhancement coefficient is realized, the thermal resistance enhancement distribution affected by wear-induced roughness increase is obtained, and a thermal conduction degradation characteristic field complementary to the heat transfer area reduction field is formed. Furthermore, the thermal response sensitivity analysis method (parameters: transient finite element temperature field simulation results, local boundary disturbance amplitude set to ±5%) is used to extract the influence weights of the above thermal resistance enhancement coefficient and heat transfer area reduction factor on the local temperature gradient, and obtain the spatially distributed thermal conduction sensitivity weight matrix. Furthermore, by weighting the thermal conductivity sensitivity matrix with the thermal resistance enhancement distribution and the heat transfer area reduction field element by element, a thermal influence potential field containing high-frequency local thermal perturbation information is generated to constrain the physical consistency of PINN in the process of microscale surface undulation structure reconstruction. By generating and processing the potential field with thermal influence, the roughness proxy parameter of the previous step is transformed into a spatially continuous and quantifiable thermal conduction constraint field, thereby realizing the embedding of thermophysical constraints into the high-frequency detail reconstruction process of dynamic geometric topology, and improving the thermodynamic adaptability and prediction reliability of geometric reconstruction. For example, in the cooling channel of a PET preform mold that has undergone 150,000 cycles of operation, the number of sensor sampling points is 32, the roughness proxy parameter ranges from 0.15 to 0.62, the equivalent heat transfer area reduction factor calculated by the heat transfer boundary condition inversion algorithm ranges from 0.02 to 0.11, and the interface contact thermal resistance enhancement coefficient ranges from 0.0008 to 0.0045 K·m² / W. In the thermal response sensitivity analysis, the disturbance amplitude is set to ±5%, and the number of mesh nodes in the finite element model is... The calculated local temperature gradient ranged from 15.6 to 38.7 K / mm, with corresponding sensitivity weight matrix element values ​​ranging from 0.45 to 1.08. By weighting the sensitivity weight matrix element-wise with the heat transfer area reduction field and the thermal resistance enhancement field, the resulting thermal influence potential field exhibited a spatial distribution trend of higher values ​​near the inlet and lower values ​​further away from the outlet, with a maximum potential value of 1.21 and a minimum value of 0.47. Using this potential field as a constraint input to PINN significantly improved the reconstruction precision of the high-frequency surface undulation structure, and substantially reduced the fitting residuals of the temperature and stress fields in subsequent finite element thermo-mechanical coupling simulations. This enabled accurate capture of the wear evolution effect and improved the stability of the model's long-term predictions. S3.4: Construct a Physical Information Neural Network (PINN), whose loss function is composed of a weighted sum of data fitting terms, residual terms of the mass conservation partial differential equation, and physical constraint terms of the heat conduction equation; input the low-frequency morphology substrate as the initial condition into the network, and introduce the thermal influence potential field as the boundary adjustment signal to drive the network to optimize the microscale surface undulation structure under the premise of satisfying physical laws, and output high-resolution real-time three-dimensional geometric topology data. Based on the combined input conditions of low-frequency morphology substrate and thermal influence potential field, a physical information neural network (PINN) modeling method is adopted (network structure: multi-layer fully connected network, activation function: tanh, physical constraint weight coefficient ratio is data fitting term: mass conservation residual term: heat conduction equation residual term = 2:1:3) to achieve microscale geometric detail optimization that takes into account both data-driven and physical law constraints. Furthermore, through an adaptive weight update mechanism (parameters: learning rate 0.001, weight update cycle 10 iteration cycles), the weight ratio of the three terms of the loss function is dynamically adjusted during gradient backpropagation, so as to achieve a balance between the data fitting accuracy and the physical equation residuals in different training stages and obtain a gradually converging network parameter set. Furthermore, by utilizing the thermally influenced potential field as a boundary adjustment signal, and correlating the potential field value with the geometric node position through a boundary condition embedding layer, the temperature gradient constraint term of the node under the boundary conditions of the heat transfer equation is calculated, and the following loss function is used to construct the formula:

[0015] in, The output value of the loss function. , , These are the dynamic weights for the three types of loss terms. This is the mean square error term for the wall thickness and roughness fitting. This represents the mean square value of the residuals in the heat-fluid mass conservation equation. This is the sum of squared residuals from the steady-state or transient heat conduction equations; Furthermore, in the physical constraint layer of PINN, the spatial coordinates of the low-frequency topography base grid nodes are combined with the temperature gradient weights of the thermal influence potential field to form a constraint vector, and the second derivative term in the heat transfer equation is calculated using an automatic differentiation engine. , , To construct a spatially continuous heat transfer residual field, thereby achieving the joint satisfaction of geometric morphology detail optimization and thermal constraints; Through the above network training and boundary condition embedding, the low-frequency degraded morphology substrate of the previous step is transformed into high-resolution geometric topology data that combines real wear topology and thermal boundary adaptability, thereby achieving the technical effect of fine reconstruction and dynamic updating of the microscale undulation structure of the inner wall of the cooling channel. For example, in the geometry reconstruction of the cooling channel of a PET preform mold with over 15,000 service cycles, the low-frequency topography substrate mesh resolution was set to 0.5 mm. The thermal influence potential field was obtained by mapping surface roughness surrogate parameters, with temperature gradient weights ranging from 0.2 to 0.8. The PINN network had a depth of 8 layers, with 256 neurons per layer, using the tanh activation function. The Adam optimizer learning rate decreased from 0.001 to 0.0001. The initial weights of the loss function were set to... Adjusted to 5000 iterations To enhance physical constraint convergence, the accuracy of the second derivative of the temperature field at each node in the automatic differential calculation of the heat transfer equation residuals is maintained at [a certain level]. The final high-resolution geometric topology model output requires no additional correction in the finite element simulation preprocessing adaptation and can be directly used for transient thermo-mechanical coupling solution. Verification results show that this geometric model can significantly improve the accuracy of local stress peak location in thermal stress prediction and maintain good long-term stability. S3.5: Perform boundary smoothing and finite element compatibility verification on the output high-resolution three-dimensional geometric topology data, remove sharp protrusions or excessively recessed areas that do not meet the manufacturing process limits, and encapsulate it into a standard CAD intermediate format file for subsequent dynamic geometry-thermodynamic coupling modeling process to ensure that the model can be parsed by the finite element engine and used for transient simulation.

[0016] like Figure 3 As shown, step S4 involves inputting the actual three-dimensional geometric topology data into the dynamic geometry-thermodynamic coupling modeling process, replacing the initial design geometric model, and updating the boundary conditions of the material's thermal conductivity and convective heat transfer coefficient according to the surface roughness proxy parameters, thus forming a time-varying thermo-mechanical coupling finite element computational domain that reflects the wear evolution state. Specifically, this includes: S4.1: Obtain the actual three-dimensional geometric topology data output by the microscale topology reconstruction model driven by the Physical Information Neural Network (PINN), perform boundary representation (B-rep) reconstruction processing based on the three-dimensional geometric topology data, and generate a dynamic geometric entity model with complete topological relationships as the physical domain benchmark that evolves over time in the thermo-mechanical coupling simulation. S4.2: Based on the surface roughness proxy parameters generated in the previous steps, the local effective thermal conductivity attenuation factor is calculated using an empirical rough surface heat transfer correction model (such as Coleman-Kays correlation). The original thermal conductivity of the mold material is spatially mapped and corrected to generate a non-uniform thermal conductivity field that varies with position, so as to reflect the degradation of thermal conductivity performance of the inner wall of the cooling channel due to wear. S4.3: Based on the same surface roughness surrogate parameter, combined with the turbulent near-wall flow model (such as the Launder-Spalding wall function), calculate the local Nusselt number correction value, and then invert to obtain the spatially distributed convective heat transfer coefficient field, generating a dynamically updated third type of boundary condition input set to characterize the actual heat transfer intensity change between the cooling medium and the worn flow channel wall. Based on the surface roughness proxy parameter input set, a turbulent near-wall flow model is adopted (model parameters: Launder-Spalding wall function coefficient is 0.41, and the turbulent shear velocity calibration value is derived from the measured flow velocity at the inlet of the cooling medium) to achieve a quantitative characterization of the flow heat transfer mechanism of roughness in the near-wall region. By using the velocity profile formula of the wall function, the roughness height is mapped to an equivalent wall shape coefficient, and further combined with the local Reynolds number. The spatial distribution of the data was analyzed, and the initial values ​​of the dimensionless Nusselt number at different monitoring locations were calculated. ; By using the roughness correction correlation (correlation parameter: equivalent roughness height) Dynamic viscosity thermal conductivity ), calculate the corrected Nusselt value The formula is as follows:

[0017] in This is the roughness influence coefficient. For the equivalent roughness height, This refers to the local hydraulic diameter; Furthermore, the revised Nusselt numbers Combined with the local thermal conductivity field, the local convective heat transfer coefficient is calculated using the classical convective heat transfer formula. :

[0018] Using a spatial interpolation method (algorithm parameters: radial basis function interpolation, kernel parameter set to 1.5 times the hydraulic diameter of the flow channel), the discrete monitoring locations are... The values ​​are mapped to a continuous spatial distribution field, forming a dynamically updated convective heat transfer coefficient field; The convective heat transfer coefficient field obtained by the above interpolation is encapsulated as a third type of boundary condition input set to realize the physical quantitative expression of the change in heat transfer intensity between the cooling medium and the worn flow channel wall. For example, for a mold cooling channel after 3000 service cycles, the input surface roughness proxy parameter vector ranges from 0.25mm to 0.45mm, corresponding to an equivalent roughness height. The thickness ranges from 0.3mm to 0.5mm, the measured flow rate of the cooling medium is 2.5m / s, the inlet temperature is 25℃, and the thermal conductivity of the mold material is... Take 43 W / (m·K), local hydraulic diameter The initial Nusselt number was calculated using the Launder-Spalding model and is 8 mm. The roughness influence coefficient is used at various locations ranging from 105 to 117. After correcting to 2.3, we obtain the corrected Nusselt number. It falls within the range of 112 to 130. Further substitution... The convective heat transfer coefficient is calculated from the values, thermal conductivity, and local hydraulic diameter. The range is 602 to 697 W / (m²·K). The three-dimensional convective heat transfer coefficient field generated by radial basis function interpolation exhibits high concentrations near the inlet region and slight attenuation in the middle section of the flow channel. This dynamic boundary condition input set significantly improves the sensitivity of temperature field prediction to local heat transfer intensity in subsequent finite element simulations, realizing the dynamic response capability of thermal stress calculation to the wear state of the flow channel. S4.4: Import the dynamic geometric solid model into the finite element preprocessing engine, divide it into a tetrahedral-prism hybrid mesh adapted to the current geometry, set a boundary layer mesh in the near-wall region to capture the temperature gradient, and assign the non-uniform thermal conductivity field and the dynamic convection heat transfer field as material properties and boundary conditions in a spatial matching manner to construct a time-varying heat conduction control equation system with physical consistency. S4.5: Based on the assigned time-varying heat conduction control equation system, define the time step sequence and initial conditions for transient solution, and encapsulate them to form a callable time-varying thermo-mechanical coupled finite element computational domain. This computational domain can respond to geometric and physical property inputs under different service cycles, providing support for subsequent execution of unsteady temperature field simulations.

[0019] Step S5: Based on the time-varying thermo-mechanical coupled finite element computational domain, perform transient temperature field simulation, output the unsteady-state temperature distribution cloud map within each forming cycle, and input the temperature field result as a thermal load into a pre-trained reduced-order model (ROM). This reduced-order model is jointly trained from historical full-size simulation data and measured strain data, and is used to quickly calculate the corresponding thermal stress field distribution. Specifically, it includes: S5.1: Based on the time-varying thermo-mechanical coupled finite element calculation domain, initialize the initial boundary conditions of the transient heat conduction equation, wherein the initial temperature field is set as the ambient temperature before mold start-up, and the boundary heat flow conditions are dynamically applied according to the inlet temperature, flow rate and convective heat transfer coefficient of the cooling medium during the injection molding cycle, so as to construct an unsteady heat conduction solution system that conforms to the actual service conditions. S5.2: The unsteady heat conduction solution system is discretized by time step. The transient control equations are transformed into a series of linear algebraic equations using the implicit Euler method. Combined with the spatial domain finite element mesh generation results, a sparse stiffness matrix and load vector are generated for each time step to support iterative solution of the temperature field step by step. S5.3: Call the parallel solver to efficiently solve the linear algebraic equations, obtain the temperature response value of each node at each time step, and then reconstruct the unsteady temperature distribution cloud map of the mold cooling channel area during the entire molding cycle, as the complete thermal load input dataset required for subsequent thermal stress analysis. S5.4: Slice the unsteady temperature distribution cloud map by time series and spatially align it to the input dimension of the pre-trained reduced-order model (ROM), wherein the reduced-order model is jointly trained based on the temperature-stress mapping sample set generated by historical full-size finite element simulation and measured strain data, and uses principal component analysis (PCA) to extract modal basis vectors, and fits the nonlinear mapping relationship between modal coefficients and thermal load through a neural network; The dataset of unsteady temperature distribution cloud map output by transient temperature field simulation is used as the processing object. The temperature vector of the whole field node and the corresponding spatial coordinate index at each time step are read to realize the temporal slice separation and single frame extraction of the temperature field. A spatial alignment mapping algorithm (parameters: finite element mesh node mapping matrix, ROM input mesh topology reference) is used to re-interpolate the node temperature values ​​of each time slice to the input node position of the pre-trained reduced-order model (ROM) to achieve spatial data matching between different mesh topologies; Furthermore, modal basis vectors are extracted from the historical full-size finite element simulation temperature-stress sample set using the principal component analysis (PCA) method (parameters: covariance matrix calculation mode = sample covariance, eigenvector selection criterion = cumulative variance contribution rate ≥ 95%). This process includes centering and standardizing the sample matrix, calculating the covariance matrix, solving for eigenvalues ​​and eigenvectors, and truncating the top K principal component basis vectors according to their contribution rates. Furthermore, the spatially aligned temperature vector is projected onto the reduced-dimensional space using the following PCA mapping formula:

[0020] in, This is the reduced modal coefficient vector. This is the spatially aligned temperature vector matrix. The modal basis vector matrix extracted by PCA; Furthermore, a neural network fitting structure is established (parameters: input layer dimension = K, number of hidden layers = 3, activation function = ReLU, output layer dimension = number of stress mode coefficients), using the thermal load mode coefficients in historical samples as input and the thermal stress mode coefficients as output, and optimizing the network weights to approximate the nonlinear mapping relationship; Through the above mapping and fitting processes, the temperature field data that has been time-sliced ​​and spatially aligned is transformed into thermal load mode coefficients that meet the ROM input requirements, thereby achieving the preprocessing goal of rapid calculation of thermal stress field. For example, in the finite element simulation of a PET preform mold, the transient temperature field output includes 120 time steps, with 45,000 mesh nodes per step. When using spatial alignment mapping, the node temperature values ​​are matched to the 5,000-node reference mesh of the ROM through cubic spline interpolation. PCA dimensionality reduction is performed on historical simulation samples (sample size = 3,000), and the cumulative variance contribution rate reaches 97%. The top 8 principal components are selected to form... Matrix. Spatially align the temperature vector at the current moment. (Dimension = 5000) multiplied by (5000×8) yields the modal coefficients. (Dimension=8), input to a four-layer fully connected neural network model configured as 8-64-64-32-output, the mean square error of the network on the validation set is significantly reduced, and the thermal stress field distribution reconstructed by the output stress mode coefficients after inverse PCA projection is highly consistent with the full-size finite element calculation results, achieving millisecond-level inference calculation speed; S5.5: Based on the reduced-order model (ROM), perform fast thermal stress field inference calculation on the received temperature field slices, and output the spatial continuous thermal stress distribution results at the corresponding time, including the equivalent stress (von Mises stress) cloud map and the principal stress components on the critical path, to achieve an efficient replacement from complete thermal simulation to millisecond-level response, and ensure the real-time requirements of dynamic thermal stress prediction. The input condition for the pre-trained reduced-order model (ROM) is an unsteady temperature field dataset obtained by S5.4 after time-series slicing and spatial alignment. Its structure is arranged sequentially along the finite element mesh nodes and meets the modal basis vector mapping dimension requirements. A fast regression method for modal coefficients (parameters: principal component basis vector set, nonlinear mapping weight matrix) is used to project the input temperature field slices onto the low-dimensional modal space extracted by PCA, generating the corresponding temperature modal coefficient vectors. Furthermore, through a neural network inference method (parameters: number of hidden layer nodes, activation function type, and training weights), a nonlinear transformation mapping from temperature mode coefficient vector to stress mode coefficient vector is achieved, and the time step coefficient sequence of each principal stress mode is obtained; Furthermore, by using the modal reconstruction operation method (parameter: set of stress mode basis vectors), the full-field reconstruction operation of the stress mode coefficient sequence in the finite element space is realized, generating the spatial continuous thermal stress distribution matrix corresponding to each time step. This process is completed in millisecond-level calculation time. Furthermore, a stress tensor decomposition algorithm (parameters: von Mises equivalent stress calculation formula, principal stress component analytical formula) is adopted to transform the stress tensor of each node in the full-field thermal stress distribution matrix into an equivalent stress scalar field and principal stress component curves on the critical path.

[0021] in, , , These represent the three components of the principal stress. This represents the von Mises equivalent stress value; Through the above-mentioned mode conversion and stress tensor analysis, the temperature field slicing results of the previous step are transformed into spatially continuous thermal stress cloud map data and critical path principal stress components, realizing the real-time technical effect of dynamic thermal stress prediction. For example, during the 200th molding cycle of a PET preform mold, the input is a 50-dimensional temperature field slice vector processed by PCA. The neural network configuration of the reduced-order model has 128 hidden layer nodes, the activation function type is ReLU, and the training weights have undergone 5000 batches of iterative optimization. During the inference phase, the temperature modal coefficient vector is mapped to generate a stress modal coefficient vector, also 50-dimensional. After modal basis vector reconstruction, the output is a full-field thermal stress distribution matrix with a total of 120,000 finite element mesh nodes. Based on the stress tensor decomposition calculation formula, the region corresponding to the maximum value in the output equivalent stress cloud map is located near the wall layer of the cooling channel inlet, and the principal stress component curve shows a significant increase in the first principal stress peak value in this region. This execution effect completes the entire calculation process within milliseconds, significantly improving the accuracy and speed of real-time thermal stress prediction within the cycle.

[0022] Step S6: Using the thermal stress field distribution data output by the reduced-order model, extract the thermal stress peak sequence at key locations, and combine it with the wall thickness decay sequence at the corresponding time to construct a wear evolution memory bank. Based on a Long Short-Term Memory (LSTM) network, learn the nonlinear dynamic mapping relationship between the two to generate a thermal stress evolution trend prediction model. Specifically, this includes: S6.1: Obtain the complete thermal stress field distribution data of each molding cycle output by the reduced-order model, filter out the set of mesh nodes corresponding to the intersection of the gate, parting surface and cooling channel based on the preset key structural area mask matrix, extract the extreme values ​​of the equivalent stress value sequence in the set of nodes, calculate the local thermal stress peak value in each cycle, and generate a time series data sequence of thermal stress peak value with timestamp. The spatial continuous thermal stress distribution results corresponding to each molding cycle output by the reduced-order model (ROM) are used as the input data source. The results are stored in the node data matrix under the finite element mesh topology in the form of equivalent stress cloud map, which includes the spatial coordinates and equivalent stress value attributes of each node. A key structural region mask matrix (parameters: gate region index set, parting surface region index set, cooling channel intersection region index set) is used to perform region filtering operations on the above node data matrix to extract the node set of the target sensitive structural region; Furthermore, the local peak values ​​of the equivalent stress sequence are calculated at the node set of each cycle using an extreme value detection algorithm (parameters: von Mises stress value vector, threshold strategy = full cycle maximum value determination) to ensure that the extracted values ​​correspond to the moments when plastic deformation risks may occur. Furthermore, through the timestamp binding process (parameters: cycle number, system clock synchronization value), the local peak stress value calculated in each cycle is appended with a time tag to form a time sequence record of thermal stress peak values ​​sorted in ascending order of cycle number; Furthermore, the thermal stress peak data with bound timestamps is standardized and stored as a single-dimensional time series vector through time-series structuring processing (parameters: time series length M, sampling interval Δt), providing a strict time-series index basis for subsequent wall thickness attenuation data alignment; Through the above mask screening and peak extraction processing, the spatial continuous stress distribution results of the reduced-order model are transformed into time series data of thermal stress peak values ​​for key structural regions, realizing the accurate quantification of thermal load extreme values ​​within the service life of the mold, and providing highly sensitive dynamic response characteristics for the construction of wear evolution memory library; For example, under the condition that the service cycle of a PET preform cooling mold is 120s, the reduced-order model outputs an equivalent stress cloud map with a total of 50,000 nodes, of which the key structural region mask matrix index set contains 3,500 nodes corresponding to the gate, parting surface, and intersection area. An extreme value detection algorithm is applied to the von Mises stress value vectors of the 3,500 nodes selected in each cycle. When the parameter setting threshold strategy is set to local maximum value determination, approximately 15 peak nodes can be obtained per cycle, with the peak stress range being [missing information]. MPa. After appending a period number and a system synchronization timestamp to the peak stress value, a time-series record of the thermal stress peak value, ordered by period and with a length of 500, is generated, with a sampling interval of [missing information]. The time series length M is set to 500, and a single-dimensional vector is ultimately formed and stored in the stress peak database of the evaluation system. The thermal stress peak records obtained in this scenario show a significant nonlinear correlation when the LSTM model is subsequently aligned with the wall thickness attenuation data, effectively improving the reliability of future periodic thermal stress trend prediction. S6.2: Synchronously retrieve the clean wall thickness measurement values ​​of each sampling point generated in S2, calculate the attenuation of the wall thickness relative to the initial design value for the sensor position that matches the key structural area in S6.1, and form a wall thickness attenuation time series data sequence that is periodically aligned with the thermal stress peak sequence, as an input feature reflecting the geometric degradation of the mold cooling channel; S6.3: Based on the time series data of peak thermal stress and the time series data of wall thickness decay, construct multivariate time series data pairs and store them in the wear evolution memory storage area; the storage area adopts a circular buffer queue structure to manage the historical data window, ensuring that the training samples always cover the most recent N effective production cycles to reflect the dynamic characteristics of the current wear stage; S6.4: Utilize multivariate time series data pairs in the wear evolution memory to perform online incremental training of the LSTM neural network: take the wall thickness decay sequence as the input sequence and the thermal stress peak sequence as the target output sequence, optimize the network weight parameters through the backpropagation algorithm, establish a nonlinear time-series mapping function from the geometric degradation state to the thermal stress response level, and generate a thermal stress evolution trend prediction model with dynamic adaptability. Based on multivariate time series data pairs stored in the wear evolution memory bank, a Long Short-Term Memory (LSTM) network architecture is adopted (parameters: the number of input layer nodes corresponds to the dimension of the wall thickness decay sequence, the number of output layer nodes corresponds to the dimension of the thermal stress peak sequence, and the number of hidden layer units is adaptively set according to the sample length) to achieve training of the dynamic mapping relationship between geometric degradation state and thermal stress response level. Furthermore, the gradient of the memory cell state of the LSTM is calculated using the backpropagation time algorithm (BPTT, parameter: the truncation step size is set to the training window length N), generating a set of partial derivatives of the weight matrix and bias vector, which are used to optimize the network parameters. Furthermore, the gradient at each time step is updated using the stochastic gradient descent (SGD) method (parameter: learning rate η is adaptively adjusted by Adam), which enables the network to perform online incremental training when new samples arrive, while preserving the stability of historical weights. Furthermore, a mean squared error loss function is introduced to quantify the prediction deviation from the wall thickness attenuation sequence to the thermal stress peak sequence, and is calculated using the following expression:

[0023] in, This is the sample loss value. This represents the true peak thermal stress. For LSTM predictions, This represents the current number of training samples; Furthermore, by performing gradient backpropagation on the loss function, the resulting gradient is used to update the input weights, hidden layer recursive weights, and output layer weights, thereby improving the prediction accuracy of the network during its service life evolution. Through the above online incremental training processing method, the multivariate time series data of the previous step is transformed into a thermal stress evolution trend prediction model that can dynamically adapt to the geometric degradation of the mold, thereby achieving the stability and real-time performance of cooling performance evaluation under long-term service environment. For example, in a PET preform mold service life monitoring scenario, the wall thickness decay sequence (each sequence length is 10, unit is mm decay) of the most recent N=50 effective production cycles and the thermal stress peak sequence (unit is MPa) are selected to form input-output sample pairs. The LSTM network is set with 64 hidden layer units, 10 input layer nodes, 1 output layer node, and a truncation step size of 50. An Adam optimizer with an initial learning rate η=0.001 and a batch size of 1 is used for incremental training. During online training, the aforementioned mean squared error loss function is introduced to calculate the prediction bias of the current batch. For example, for the k-th cycle, the measured peak value... =120MPa, predicted value =115MPa, then the loss contribution of this sample is =0.5102. Through gradient updates, the prediction error of the network in subsequent cycles is significantly reduced, and the mean square error between the output thermal stress peak extrapolation curve and the measured value in the latest test cycle converges to within 0.2, realizing high fitting and real-time response capability of the thermal stress prediction results to the actual working conditions under the condition of continuous evolution of wear state; S6.5: Perform a one-step advance prediction verification on the trained LSTM model. Input the wall thickness attenuation observations of the latest M cycles, generate the predicted value of the thermal stress peak for the next cycle, and compare it with the output of the actual reduced-order model to calculate the mean square error index. If the error exceeds the preset threshold, the model retraining mechanism is triggered to maintain the accuracy and robustness of the thermal stress evolution trend prediction model.

[0024] Step S7: Based on the thermal stress evolution trend prediction model, predict the thermal stress field distribution over several future forming cycles, and determine whether there are spatial regions in the prediction results that exceed the material yield strength threshold. If so, mark them as thermal stress risk hotspots. Specifically, this includes: S7.1: A thermal stress evolution trend prediction model based on the output of the preceding steps. This model is constructed by learning the nonlinear dynamic mapping relationship between the wall thickness decay sequence and the thermal stress peak sequence in the wear evolution memory bank using a long short-term memory network (LSTM). The latest wall thickness decay observation value and corresponding thermal stress field distribution data at the key locations in the current service stage are obtained as initial input conditions. The time extrapolation capability of the LSTM model is used to perform multi-step forward propagation calculations to generate the predicted thermal stress peak sequence for each key location in the next N forming cycles. The thermal stress evolution trend prediction model is based on a long short-term memory network (LSTM). The input conditions are the latest wall thickness attenuation observations and corresponding thermal stress field distribution data at key locations during the current service phase. A time series normalization method (parameter: the normalization interval between the maximum and minimum values ​​is set to [0,1]) is used to perform physical quantity dimension mapping consistency processing on the wall thickness attenuation observation sequence and the thermal stress peak sequence, so as to achieve feature fidelity of the input data under a unified scale; Furthermore, a multi-step prediction input sample tensor is constructed using the sequence time window slicing method (parameters: window length M is the most recent M forming cycles, step size is 1 cycle), thereby realizing the structured organization of temporal features in the batch training and inference stages; Furthermore, by utilizing the multi-layer stacked structure of the LSTM network (parameters: input layer dimension equal to the length of the double feature sequence, number of hidden layer neurons is 128, activation function is tanh), multi-step forward propagation computation is performed to generate thermal stress prediction output for future cycles while considering long-term dependence. Furthermore, by regressing the linear transformation of the output layer, the predicted sequence of thermal stress peak values ​​at each key location within the next N forming cycles is obtained, and the predicted values ​​are automatically bound to timestamps and spatial location indices to form the input benchmark for subsequent spatial field interpolation. By using the chain algorithm described above, the wear-stress time series relationship model from the previous step is transformed into position-by-position thermal stress peak prediction data for the next N periods, thereby quantifying the thermal load change trend under the extrapolation of the service life. For example, in a PET preform mold cooling channel life assessment case, the key monitoring location is a node near the gate. The wall thickness decay observation values ​​were recorded as a monotonically increasing sequence of 0.05mm, 0.06mm, and 0.07mm in the last 10 cycles, with corresponding equivalent thermal stress peak values ​​of 120MPa, 122MPa, and 125MPa. The decay sequence was mapped to the [0,1] range using a maximum-minimum normalization method, and an input tensor was constructed with a time window length M=5. The number of hidden layer neurons in the LSTM prediction network was set to 128, the batch size to 32, and the prediction step size N to 3 cycles. The normalized decay sequence and thermal stress sequence were input into the network for forward inference calculation. The network output the thermal stress peak values ​​for the next three cycles as 128.4MPa, 131.7MPa, and 134.9MPa, respectively, with corresponding timestamps appended. The output has been verified to have an average absolute error of less than 2 MPa compared with the actual simulation results, which meets the evaluation accuracy requirements. Furthermore, spatial interpolation can be used in subsequent steps to generate a complete thermal stress field distribution cloud map, thereby significantly improving the reliability of risk hotspot prediction. S7.2: The predicted sequence of thermal stress peaks at each key location is associated with its spatial coordinate information, and a complete spatial distribution cloud map of thermal stress field for each future target period is reconstructed based on the finite element mesh topology interpolation. The reconstruction process uses the radial basis function (RBF) interpolation algorithm to perform spatial smoothing on the discrete point prediction results to generate continuous two-dimensional / three-dimensional thermal stress field distribution data, which serves as the input field for subsequent risk criterion analysis. S7.3: Obtain the temperature-related yield strength database of the metal materials used in the PET preform mold. This database records the yield strength threshold of the mold steel at different service temperatures. Based on the temperature distribution cloud map of each cycle output by the transient temperature field simulation, extract the local temperature value of each spatial location at the corresponding moment of the thermal stress field, and look up the table to obtain the real-time yield strength threshold at that temperature, forming a dynamic strength reference field that varies with time and space. S7.4: Perform pixel-by-pixel comparison calculations between the reconstructed complete thermal stress field distribution cloud map and the dynamic strength reference field, and execute spatial Boolean logic judgment: if the predicted thermal stress value of a certain spatial location in a future period exceeds its corresponding dynamic yield strength threshold, then the location is determined to have entered the plastic deformation risk zone; generate a binary thermal stress over-limit identification matrix to mark all over-limit spatial regions and their occurrence periods, as the preliminary detection results of thermal stress risk hotspots; S7.5: Perform connected component analysis and morphological closure processing on the thermal stress over-limit identification matrix, merge adjacent isolated risk points and eliminate noise interference, and extract risk hotspot areas with significant geometric features; output a risk hotspot feature set containing spatial coordinates, expanded area, maximum over-limit amplitude and first occurrence period, which is used for subsequent assessment report generation and early warning action triggering.

[0025] Step S8: Generate a dynamic thermal stress assessment report containing the current thermal stress field distribution, risk hotspot area identification, and remaining safety period estimate, and output it to the process optimization system to trigger cooling strategy adjustments or mold maintenance early warning actions. Specifically, this includes: S8.1: Based on the thermal stress field distribution data of the current molding cycle output by the reduced-order model and the thermal stress risk hotspot areas marked in S7, obtain the spatially gridded thermal stress tensor matrix and the corresponding node coordinate set, use the isosurface extraction algorithm to generate a three-dimensional thermal stress distribution cloud map, and apply high-brightness color coding to the areas exceeding the material yield strength threshold to generate a visual thermal stress risk hotspot layer, so as to intuitively present the potential failure locations in the mold structure. S8.2: Based on the thermal stress evolution trend prediction model generated by LSTM, the peak thermal stress variation curves of multiple future forming cycles are output. Combined with the current number of running cycles and historical wear decay rate, the expected time window when the thermal stress reaches the material yield limit is calculated, and the remaining safe cycle estimate is generated. This estimate is processed by piecewise linear interpolation and confidence interval boundary correction to form a statistically reliable life prediction output. S8.3: The generated three-dimensional thermal stress distribution cloud map, thermal stress risk hotspot layer and remaining safety period estimate are fused with multi-source information, and a predefined report template engine is called to encapsulate it into a structured dynamic thermal stress assessment report according to the standard communication protocol format of the industrial edge system. The report includes timestamp, mold identification code, key location number and its corresponding thermal stress state classification label. S8.4: The structured dynamic thermal stress assessment report is pushed to the host computer process optimization system via the OPC UA protocol, and simultaneously stored in the time-series database for archiving on the local edge computing node; if the report contains risk hotspot areas or the remaining safety cycle is lower than a preset threshold, the priority marking mechanism is automatically activated, and a high-priority event notification frame is sent to the manufacturing execution system (MES). S8.5: Based on the received dynamic thermal stress assessment report, perform rule matching operations in the process optimization system: if there is a thermal stress risk hotspot area, trigger the optimization suggestion of PID adjustment strategy for cooling channel flow; if the remaining safety cycle is lower than the maintenance threshold, generate a mold preventive maintenance work order and push it to the equipment management terminal to achieve a closed-loop response from assessment to control.

[0026] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0027] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for evaluating the cooling effect of a cooling channel in a PET preform mold, characterized in that, Includes the following steps: S1: Periodically collect the original signal and echo characteristic data of the channel wall thickness at multiple spatially distributed sampling points on the inner wall of the cooling channel of the PET preform mold, and obtain the time-series observation value of the local geometric state of the cooling channel; S2: Bandpass filtering and wavelet denoising are performed on the time series observations. Based on the reflection characteristics of different material interfaces, the effective echoes on the inner surface of the flow channel are separated, and their flight time and amplitude attenuation characteristics are extracted to generate the clean wall thickness measurement value and surface roughness proxy parameter corresponding to each sampling point. S3: Based on the clean wall thickness measurement and surface roughness proxy parameters of multiple sampling points, combined with the fluid erosion empirical model and heat transfer boundary condition inversion algorithm, a microscale morphology reconstruction model is constructed. The microscale morphology reconstruction model is driven by the physical information neural network to generate the actual three-dimensional geometric topology data of the cooling channel under the current service stage. S4: Input the actual three-dimensional geometric topology data into the dynamic geometry-thermodynamic coupling modeling process to replace the initial design geometric model, and update the boundary conditions of the material thermal conductivity and convective heat transfer coefficient according to the surface roughness proxy parameters to form a time-varying thermo-mechanical coupling finite element calculation domain. S5: Perform transient temperature field simulation based on the time-varying thermo-mechanical coupled finite element computational domain, output the unsteady temperature distribution cloud map in each molding cycle, and input the unsteady temperature distribution cloud map as thermal load into the pre-trained reduced-order model to calculate the corresponding thermal stress field distribution.

2. The method for evaluating the cooling effect of a cooling channel in a PET preform mold according to claim 1, characterized in that, Following step S5, the following is also included: S6: Using the thermal stress field distribution data, extract the thermal stress peak sequence at key locations, and combine it with the wall thickness decay sequence at the corresponding time to construct a wear evolution memory bank. Based on the long short-term memory network, learn the nonlinear dynamic mapping relationship between the two to generate a thermal stress evolution trend prediction model. S7: Based on the thermal stress evolution trend prediction model, extrapolate the thermal stress field distribution in several future molding cycles to predict whether there are spatial regions in the prediction results that exceed the material yield strength threshold. If so, mark them as thermal stress risk hotspot regions. S8: Generates a dynamic thermal stress assessment report containing the current thermal stress field distribution, risk hotspot area identification, and remaining safety period estimate, and outputs it to the process optimization system to trigger cooling strategy adjustments or mold maintenance early warning actions.

3. The method for evaluating the cooling effect of a cooling channel in a PET preform mold according to claim 1, characterized in that, Step S1 specifically includes: Based on the spatial structure parameters of the cooling channel of the PET preform mold and the distribution map of the thermo-mechanical coupling sensitive area, the set of key monitoring locations where sensors need to be deployed is determined. The mold CAD model is used for three-dimensional path planning to generate a sampling point layout scheme that meets the minimum coverage redundancy and the maximum heat-affected zone coverage. Miniaturized ultrasonic thickness sensor units are embedded at key monitoring locations and fixed to the substrate material of the inner wall of the cooling channel using micro-welding technology to form an embedded sensing network. The edge-triggered acquisition mechanism is configured to start the data acquisition timing based on the injection molding cycle synchronization signal, control each sensor unit to emit high-frequency ultrasonic pulses at the beginning of the cooling stage of each molding cycle, and receive the original echo signal from the mold steel-cooling medium interface to generate the original signal sequence. The original signal sequence is subjected to analog-to-digital conversion and timestamp binding processing to convert the analog voltage signal into a digital sampling sequence and associate it with the corresponding spatial coordinates and period number to generate an original wall thickness observation data package with spatiotemporal tags. The original wall thickness observation data packet with spatiotemporal tags is transmitted to the local edge computing node cache area to complete data collection and preliminary verification.

4. The method for evaluating the cooling effect of a PET preform mold cooling channel according to claim 3, characterized in that, The miniaturized ultrasonic thickness sensor unit integrates a piezoelectric chip, a temperature compensation module, and a wireless transmission interface.

5. The method for evaluating the cooling effect of a cooling channel in a PET preform mold according to claim 1, characterized in that, Step S2 specifically includes: The original ultrasonic signals periodically collected by a miniature embedded ultrasonic thickness sensor array on the inner wall of the cooling channel of a PET preform mold are obtained. Based on the preset acoustic impedance difference characteristics of the cooling medium-metal interface, the theoretical arrival interval of the target echo in the time domain is determined to limit the effective signal analysis window. The original ultrasound signal is subjected to bandpass filtering within the time domain analysis window to obtain an intermediate signal sequence with preliminary signal-to-noise ratio improvement; Discrete wavelet transform is performed based on the intermediate signal sequence. The db6 wavelet basis function is selected and a five-level decomposition is implemented. An adaptive soft threshold denoising algorithm is applied to the detail coefficients of each level. The threshold value is dynamically adjusted according to the noise standard deviation estimate. Wavelet reconstruction is then performed to generate a cleaned ultrasonic echo signal. Based on the cleaned ultrasonic echo signal, the first effective echo peak value corresponding to the metal-coolant interface of the inner wall of the cooling channel is identified, its flight time is calculated, and combined with the calibrated value of the propagation speed of ultrasonic waves in the mold steel, the instantaneous wall thickness measurement value at each sampling point is calculated. At the same time, the amplitude attenuation degree of the first effective echo peak value is analyzed, and the surface roughness proxy parameter is generated by using the empirical mapping relationship between amplitude attenuation and surface micro-irregularity. The instantaneous wall thickness measurement values ​​corresponding to each sampling point and the surface roughness proxy parameters are spatiotemporally aligned and normalized to generate a structured clean wall thickness measurement value sequence and a surface roughness proxy parameter vector.

6. The method for evaluating the cooling effect of a cooling channel in a PET preform mold according to claim 1, characterized in that, Step S3 specifically includes: Based on the clean wall thickness measurement and surface roughness proxy parameters of each sampling point, combined with the mold cooling medium flow rate, temperature time series records and material hardness distribution data, the cumulative erosion depth evolution curve of each sampling area is calculated using the fluid erosion empirical model to generate a local wear estimation sequence. Based on the difference between the estimated local wear amount sequence and the original design geometric reference surface, an initial microscale morphology degradation field is constructed. The spatial covariance of the non-sampling area is estimated by Kriging interpolation, generating a preliminary full-field three-dimensional geometric deviation distribution grid to form a low-frequency morphology base. Based on the heat transfer boundary condition inversion algorithm, the surface roughness proxy parameter at each sampling point is mapped to the equivalent heat transfer area reduction factor and the interface contact thermal resistance enhancement coefficient. The influence weight of the parameter on the local temperature gradient is extracted through thermal response sensitivity analysis to generate a spatially distributed thermal influence potential field. A physical information neural network is constructed, and the low-frequency morphological substrate is used as the initial condition input into the physical information neural network. At the same time, the thermal influence potential field is introduced as the boundary adjustment signal to drive the network to optimize the microscale surface undulation structure under the premise of satisfying physical laws and output high-resolution real-time three-dimensional geometric topology data. The real-time three-dimensional geometric topology data is subjected to boundary smoothing and finite element adaptability verification to remove sharp protrusions or excessively recessed areas that do not meet the manufacturing process limits, and then packaged into a standard CAD intermediate format file.

7. The method for evaluating the cooling effect of a PET preform mold cooling channel according to claim 6, characterized in that, The loss function in the physical information neural network is composed of a weighted sum of the data fitting term, the residual term of the mass conservation partial differential equation, and the physical constraint term of the heat conduction equation.

8. The method for evaluating the cooling effect of a PET preform mold cooling channel according to claim 1, characterized in that, Step S4 specifically includes: The actual three-dimensional geometric topology data output by the microscale topology reconstruction model driven by the physical information neural network is obtained, and the boundary representation reconstruction processing is performed based on the actual three-dimensional geometric topology data to generate a dynamic geometric entity model. Based on the surface roughness proxy parameter, the local effective thermal conductivity attenuation factor is calculated using the empirical rough surface heat transfer correction model, and the original thermal conductivity of the mold material is spatially mapped and corrected to generate a non-uniform thermal conductivity field. Based on the surface roughness proxy parameters, combined with the turbulent near-wall flow model, the local Nusselt number correction value is calculated, and then the spatially distributed convective heat transfer coefficient field is obtained by inversion, generating a dynamically updated third type of boundary condition input set. The dynamic geometric solid model is imported into the finite element preprocessing engine, and a tetrahedral-prism hybrid mesh adapted to the current geometric shape is divided. A boundary layer mesh is set in the near-wall region to capture the temperature gradient. The non-uniform thermal conductivity field and the dynamic convective heat transfer field are used as material properties and boundary conditions for spatial matching and assignment, and a time-varying heat conduction control equation system is constructed. Based on the aforementioned time-varying heat conduction control equation system, a time step sequence and initial conditions for transient solution are defined and encapsulated to form a callable time-varying thermo-mechanical coupled finite element computational domain.

9. The method for evaluating the cooling effect of a cooling channel in a PET preform mold according to claim 1, characterized in that, Step S5 specifically includes: Based on the time-varying thermo-mechanical coupled finite element computational domain, the initial boundary conditions of the transient heat conduction equation are initialized, wherein the initial temperature field is set as the ambient temperature before mold start-up, and the boundary heat flow conditions are dynamically applied according to the inlet temperature, flow rate and convective heat transfer coefficient of the cooling medium during the injection molding cycle, thereby constructing an unsteady heat conduction solution system. The unsteady heat conduction solution system is discretized by time step. The transient control equations are transformed into a system of linear algebraic equations using the implicit Euler method. Combined with the spatial domain finite element mesh generation results, the sparse stiffness matrix and load vector at each time step are generated. The parallel solver is called to efficiently solve the linear algebraic equations, obtain the temperature response value of each node at each time step, and then reconstruct the unsteady temperature distribution cloud map of the mold cooling channel region throughout the entire molding cycle. The unsteady temperature distribution cloud map is sliced ​​according to the time series and spatially aligned to the input dimension of the pre-trained reduced-order model. Principal component analysis is used to extract the modal basis vectors, and a neural network is used to fit the nonlinear mapping relationship between the modal coefficients and the thermal load. Based on the reduced-order model, fast thermal stress field inference calculations are performed on the received temperature field slices, and the spatial continuous thermal stress distribution results at the corresponding time are output.

10. The method for evaluating the cooling effect of a cooling channel in a PET preform mold according to claim 9, characterized in that, The reduced-order model is jointly trained based on a temperature-stress mapping sample set generated from historical full-size finite element simulations and measured strain data.

Images