A temperature field digital twin closed loop assimilation and process feedback optimization method for composite material curing process

Abstract

The application provides a temperature field digital twin closed loop assimilation and process feedback optimization method for a composite material curing process, and relates to the technical field of autoclave temperature control. A staged reduced temperature field model and a time series prediction model are constructed through a temperature field multi-source database to realize rapid online prediction of the full-field temperature field of the component in a low-dimensional modal space; real-time temperature data of the thermocouple of the part is collected during the curing process, observation residuals between the observation values of the thermocouple measuring points and the predicted measuring point temperatures are established, and the prior temperature field state is modified by weighted closed loop assimilation combined with the physical consistency discrimination and the reliability weight calculation of the measuring points, so that the posterior state is fed back to the next time prediction process to improve the online estimation accuracy and robustness of the full-field temperature field of the component. And based on the assimilated posterior full-field temperature field, a process parameter dynamic optimization strategy can be extracted to form a current tank online feedback closed loop and a tank-to-tank model iterative update closed loop.

Description

Technical Field

[0001] This invention relates to the field of autoclave temperature control technology, and in particular to a method for temperature field digital twin closed-loop assimilation and process feedback optimization in the curing process of composite materials. Background Technology

[0002] Carbon fiber reinforced resin matrix composite components have been widely used in high-end manufacturing fields such as aerospace, rail transportation, and energy equipment due to their advantages such as high specific strength, high specific stiffness, fatigue resistance, and strong designability. Autoclave curing is the core process for achieving stable molding quality in these components. During the curing process, the resin matrix inside the component undergoes a curing reaction accompanied by an exothermic effect. Complex unsteady-state heat transfer coupling exists between the mold, auxiliary materials, the component, and the autoclave environment. The temperature field distribution directly affects resin flow, degree of curing evolution, residual stress formation, and final molding quality. For large, thick-walled, variable-thickness, or reinforced composite components, due to factors such as complex geometry, differences in layup, significant mold thermal inertia, and uneven local heat transfer conditions, different areas of the component are prone to temperature lag, local overheating, local underheating, and temperature unevenness. Relying solely on the autoclave air temperature or a few thermocouple measuring points is insufficient to reflect the temperature distribution across the entire component in real time, and it is also impossible to accurately predict potential quality risks.

[0003] Existing technologies employ various methods for simulating, predicting, and assimilating the autoclave curing process of composite material components. However, these methods still have many shortcomings. For example, while high-fidelity numerical simulation methods can meticulously describe the temperature field distribution, they suffer from large computational scale, long processing time, and strong parameter dependence, making them difficult to directly apply for online monitoring and rapid process feedback. Furthermore, when there are deviations between actual operating conditions and simulation settings, prediction accuracy is difficult to guarantee. Thermocouple monitoring methods are mature in engineering applications, but limited by the number of measurement points, they can only reflect temperature changes at a limited number of locations and cannot directly obtain the entire temperature field. Moreover, abnormal measurement points can easily lead to distorted state judgments. Although reduced-order models and surrogate models have relatively fast computation speeds, they are mostly open-loop prediction modes. Relying on the coverage of training data, it is prone to systematic errors when facing deviations in process, boundary, and material parameters of actual batches, and lacks a real-time observation and correction mechanism. Existing data assimilation methods mostly only make single corrections to the temperature field at the current moment, without feeding back the posterior state to the subsequent prediction process, and cannot form a continuously evolving closed-loop assimilation mechanism. At the same time, existing digital twin and process control schemes are mostly isolated from each other, lacking a complete technical system that deeply couples sparse observation assimilation, full-field temperature reconstruction, quality risk assessment, and process feedback optimization. It is difficult to simultaneously achieve online closed-loop control of the current batch and model iterative optimization across batches, and cannot fully meet the high-precision and intelligent quality control requirements of the curing process of large composite material components.

[0004] Therefore, it is necessary to propose a closed-loop assimilation and process feedback optimization method for the temperature field digital twin in the composite material curing process to solve or at least alleviate the above-mentioned defects. Summary of the Invention

[0005] The main objective of this invention is to provide a digital twin closed-loop assimilation and process feedback optimization method for the temperature field in the curing process of composite materials, in order to solve the problems in the prior art where data assimilation lacks a continuous closed-loop feedback mechanism, and process control lacks a basis for the entire field temperature, making it difficult to simultaneously achieve online optimization of the current batch and iterative improvement across batches.

[0006] To achieve the above objectives, this invention provides a method for temperature field digital twin closed-loop assimilation and process feedback optimization in the curing process of composite materials, comprising the following steps: S1. Establish a multi-source temperature field database based on the curing process of the composite material component in the current batch. The multi-source temperature field database includes the posterior modal coefficients of the previous moment and the current curing characteristic data. The current curing characteristic data includes the current process parameters, the current curing stage, and time information. S2, the curing process of composite material components is divided into multiple process stages, and the temperature field data of each process stage is input into the staged reduced-order temperature field model for order reduction processing in combination with the multi-source temperature field database to obtain a low-dimensional temperature field modal coefficient sequence. S3, Construct a time-series prediction model based on the low-dimensional temperature field modal coefficient sequence; S4, For the current moment, input the posterior modal coefficients of the previous moment and the current solidification feature data into the time series prediction model to predict the prior modal coefficients of the current moment, and determine the predicted temperature of the thermocouple measuring point of the part based on the prior modal coefficients of the current moment. S5. Calculate the reliability of the measurement point at the current moment based on the observed temperature and the predicted temperature of the thermocouple measurement point of the component. Correct the prior modal coefficients at the current moment based on the reliability of the measurement point at the current moment to obtain the posterior modal coefficients at the current moment. When the next moment arrives, update the posterior modal coefficients at the current moment to the posterior modal coefficients at the previous moment in the multi-source temperature field database, and return to step S4 to perform the process. Repeat steps S4 to S5. S6, when the current number of cycles reaches the preset number threshold, the temperature field is reconstructed based on the posterior modal coefficients at the current moment to obtain the posterior full-field temperature field. The quality risk assessment of the current curing process is performed based on the posterior full-field temperature field to output the quality risk assessment result, and the process optimization instruction for the current batch is generated according to the quality risk assessment result. S7, execute the process optimization instruction, update the temperature field multi-source database based on the multi-source data of the current batch when the curing process of the current batch ends, initialize the current cycle number to zero when the curing process of the next batch begins, determine the first predicted time of the next batch as the current time, and return to execute step S4.

[0007] Preferably, step S2 specifically includes the following steps: S21, Based on the autoclave curing process curve of composite material components, the curing process is divided into multiple process stages; S22, generate a temperature field snapshot matrix based on the temperature field data of each process stage; S23, Based on the temperature field snapshot matrix, a staged reduced-order temperature field model is constructed. Based on the staged reduced-order temperature field model, the high-dimensional temperature field data is reduced in order to obtain a low-dimensional temperature field modal coefficient sequence, expressed as follows: in, This represents a sequence of modal coefficients for a low-dimensional temperature field. Represents the average temperature field; Represents the reduced-order basis function matrix; Indicates at time The modal coefficients.

[0008] Preferably, the prior mode coefficients at the current time in step S4 are expressed as follows: in, Represents the prior modal coefficients at the current moment; Represents the posterior modal coefficients after assimilation correction at the previous time step; Indicates the process parameters for the current tank batch; Indicates the current solidification stage; Indicates the current time; Indicates the time step; The parameter is The time series prediction model.

[0009] Preferably, step S5 specifically includes the following steps: S51, determine the arrangement position of thermocouple measuring points of each component, and establish the observation mapping relationship between thermocouple measuring points of each component and the low-dimensional temperature field modal coefficient, so as to obtain the predicted temperature of thermocouple measuring points of the component based on the observation mapping relationship. S52, collect the observed temperature of each thermocouple measuring point arranged on the surface or inside of the composite material component parts, and calculate the observation residual based on the predicted temperature and the thermocouple observed temperature; S53, calculate the confidence level of the measurement point, and determine whether there is a reliable observation based on the comparison result between the confidence level of the measurement point and the preset threshold; if so, calculate the assimilation gain based on the confidence level of the measurement point, and combine the assimilation gain and the observation residual to correct the prior mode coefficient at the current time to obtain the posterior mode coefficient at the current time. S54, when the next moment arrives, update the posterior modal coefficients of the current moment to the posterior modal coefficients of the previous moment in the temperature field multi-source database, and return to execute step S4, so as to cyclically execute steps S4 to S5 to realize the closed-loop update of the temperature field digital twin.

[0010] Preferably, step S6 specifically includes the following steps: S61, the posterior full-field temperature field is obtained by reconstructing the posterior modal coefficients at the current moment after the loop is completed; S62, extract temperature features for quality monitoring and risk assessment based on the posterior full-field temperature field; S63, perform a quality risk assessment on the current curing process based on the temperature characteristics to classify the risk level; S64, determine whether the current risk level is within the acceptable range; if yes, generate a process optimization instruction to maintain the current process parameters; if no, identify the risk type and generate a process optimization instruction for the current moment.

[0011] Preferably, the predicted temperature of the thermocouple measuring point of the component in step S51 is... Expressed as: in, This represents the temperature vector at the thermocouple measuring point, representing the average temperature field. This represents the observation matrix.

[0012] Preferably, the calculation of the reliability of the measurement point in step S53 includes: The reliability of the measurement points is calculated using a weighted summation method. Its expression is: in, Indicates the first Each measuring point at time... The Class consistency score; Indicates the first The weights corresponding to the class consistency score; Indicates the number of consistency evaluation items; .

[0013] Preferably, the correction process in step S53 specifically includes the following steps: S531, based on the reliability of the measurement points The observation noise matrix is ​​dynamically adjusted, and its expression is as follows: in, Indicates the first Each measuring point at time... Confidence correction for observation noise; Indicates the first The basic observation noise corresponding to each measuring point; Indicates the first Each measuring point at time... Credibility weight; This prevents small positive numbers with a denominator of zero; S532, calculate the confidence-weighted assimilation gain, which is expressed as: in, This represents the assimilation gain after considering the reliability of the measurement points; This represents the prior state error covariance matrix at the current moment; Represents the observation matrix; Represents the observation matrix transpose; This represents the observation noise matrix after confidence correction; This represents finding the inverse of a matrix. S533, the prior mode coefficients are corrected based on the aforementioned confidence-weighted assimilation gain, expressed as follows: in, Represents the posterior modal coefficients after assimilation correction at the current time. , This represents the observation residual at the current moment; This indicates the real-time temperature of the thermocouple at the current moment. This indicates the thermocouple's predicted temperature at the current moment.

[0014] Preferably, the calculation of the reliability of the measurement point in step S53 further includes: The reliability of the measurement points is calculated by comprehensively considering multiple consistency evaluation items. Its expression is: in, Indicates the first Temperature range consistency score for each measuring point; Indicates the first Time continuity score for each measurement point; Indicates the first Consistency score of heating rate at each measuring point; Indicates the first Spatial consistency score of each measurement point; Indicates the first Consistency score of process stages at each measuring point; Indicates the first Consistency score of model residuals at each measurement point; This represents the comprehensive evaluation function.

[0015] Preferably, the process optimization instruction at the current moment in step S64 is specifically expressed as follows: in, This indicates the process optimization instruction at the current moment; This represents the posterior global temperature field at the current moment; This represents the set of temperature features extracted from the posterior full-field temperature field. This indicates the risk assessment result at the current moment; Indicates process window constraints; Indicates equipment operating constraints; This represents the process optimization decision function.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a closed-loop assimilation and process feedback optimization method for the temperature field digital twin in a composite material curing process. It constructs a multi-source temperature field database to provide a comprehensive data foundation for temperature field modeling and process optimization. A phased reduction-order modeling approach compresses the high-dimensional temperature field into low-dimensional modal coefficients and combines this with a time-series prediction model for prior prediction. This significantly reduces the online computational load while preserving the core spatial distribution characteristics of the temperature field, enabling the overall temperature field prediction to meet the real-time requirements of the autoclave curing process. By introducing a measurement point reliability assessment mechanism to perform physical consistency judgment and weighting processing on real-time thermocouple observation data, it effectively suppresses the interference of abnormally low-reliability measurement points on the assimilation results, while retaining effective observation information from actual operating condition deviations, thus improving the robustness and estimation accuracy of the assimilation process. Based on a closed-loop recursive mechanism of "prior prediction—reliability-weighted assimilation—posterior state feedback to the next moment's prediction," the temperature field digital twin can be dynamically updated in real time. The continuous self-calibration of observations effectively adapts to various engineering disturbances such as process parameter mismatches and boundary heat transfer deviations, significantly improving the online estimation accuracy and tracking capability of the overall temperature field. Furthermore, based on the assimilated posterior overall temperature field, quality risk assessments are conducted, overcoming the limitations of traditional single-point temperature measurement. This allows for the comprehensive identification of potential quality risks such as localized overheating, thermal hysteresis, and temperature unevenness, achieving an upgrade from measurement-point-level judgment to overall temperature state judgment. Simultaneously, the risk assessment results are combined to output targeted process optimization decisions and drive adjustments in the autoclave, realizing online closed-loop feedback from temperature field perception to process control. This effectively improves the temperature uniformity and curing quality of the curing process. Moreover, through autoclave data archiving and database iterative update mechanisms, the current autoclave's operational and quality data can be fed back into model optimization and subsequent process recommendations, forming a cross-autoclave learning loop. This provides a complete technical path for continuous improvement of curing processes and iterative enhancement of model accuracy in mass production scenarios. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the overall method in one embodiment of the present invention; Figure 2 This is a schematic diagram showing the data comparison of the experimental process in one embodiment of the present invention.

[0019] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0023] Furthermore, the use of terms such as "first" and "second" in this invention is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature. Additionally, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. When the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed by this invention.

[0024] Please see the appendix Figure 1 ~Appendix Figure 2 The present invention provides a method for temperature field digital twin closed-loop assimilation and process feedback optimization in a composite material curing process, comprising the following steps: S1. Establish a multi-source temperature field database based on the current composite material component curing process. The multi-source temperature field database includes the posterior modal coefficients of the previous time step and the current curing characteristic data. The current curing characteristic data includes the current process parameters, the current curing stage, and time information. By integrating relevant data on the composite material curing process from multiple sources, the multi-source temperature field database is established. During implementation, at least one or more of the following are collected: experimental data, simulation data, historical batch data, process parameter data, material and boundary parameter data, equipment operation data, and quality inspection data. Among them, experimental data includes thermocouple test data (measured temperature curves) of parts embedded in the process test, process test data (curing degree test data under different process parameters), and quality inspection data of the molded component; simulation data includes component temperature field data and curing degree field results obtained from high-fidelity numerical simulation, covering samples under different heating rates, holding temperatures, holding times, and boundary heat transfer conditions; historical batch data includes autoclave process execution curves, thermocouple monitoring data, equipment operation parameters, and molding quality traceability records of past batches. All data uses a unified spatial coordinate system, timestamp format, and parameter units, and is categorized and stored according to component type, material system, and operating condition, forming a standardized temperature field sample set. This provides a comprehensive data foundation for subsequent phased reduction modeling, thermocouple placement optimization, online closed-loop assimilation, and process feedback optimization. It ensures the sample coverage of both the reduction and prediction models and provides a benchmark for subsequent model iteration and process optimization, solving the problems of insufficient samples from a single data source and the difficulty in reusing scattered data.

[0025] S2, the curing process of composite material components is divided into multiple process stages. The temperature field data of each process stage is input into a staged reduced-order temperature field model using the multi-source temperature field database for order reduction processing, resulting in a low-dimensional temperature field modal coefficient sequence. The high-dimensional temperature field is compressed through staged order reduction to obtain low-dimensional temperature field modal coefficients. In practice, based on the temperature change characteristics of the composite material autoclave curing process, the complete curing process is first divided into multiple process stages with similar heat transfer and reaction laws. Then, for the temperature field samples of each process stage, an intrinsic orthogonal decomposition and other order reduction algorithms are used to extract the dominant spatial modes of the temperature field in that stage, establishing a linear mapping relationship between the overall temperature vector and the low-dimensional modal coefficients. Since the heating, holding, and cooling stages of the curing process correspond to three different thermal response laws—transient heat transfer, curing heat release and heat transfer equilibrium, and cooling heat transfer—staged processing avoids the problems of modal redundancy and insufficient fitting accuracy across stages in a single reduced-order model. By compressing the full-field temperature state of tens of thousands to hundreds of thousands of dimensions into low-dimensional modal coefficients of less than tens of dimensions, the complexity of online calculation is greatly reduced, enabling the prediction and correction of the full-field temperature field to meet the real-time requirements of the autoclave curing process.

[0026] It is worth mentioning that the first prediction time of the current batch is determined as the current time, and the initial posterior modal coefficients are determined based on the initial temperature field and the temperature field modal basis corresponding to the initial curing stage of the current batch. The initial posterior modal coefficients are used as the posterior modal coefficients of the previous time when step S4 is executed for the first time.

[0027] S3. A time-series prediction model is constructed based on the low-dimensional temperature field modal coefficient sequence. The low-dimensional temperature field modal coefficient sequences of each process stage are used as output labels, and the historical modal coefficients, process parameters, curing stage identifiers, and time information at corresponding times are used as input features. A network structure with time-series modeling capabilities is employed to build the model framework. The time-series prediction model can use an LSTM network, or it can use a GRU, Transformer, temporal convolutional network, state-space model, neural differential equation model, or other models with time-series prediction capabilities. This establishes a time-series evolution mapping relationship of low-dimensional states, enabling the model to quickly deduce the temperature field modal coefficients of subsequent times based on historical states. Temperature field prior prediction can be completed without running high-fidelity simulations online, ensuring the real-time performance of online computation.

[0028] S4, for the current moment, the posterior modal coefficients of the previous moment and the current curing feature data are input into the time-series prediction model to predict the prior modal coefficients of the current moment. Based on the prior modal coefficients of the current moment, the predicted temperature of the thermocouple measuring point of the part is determined. The prior modal coefficients of the current moment are deduced through the time-series prediction model, which is implemented in two stages: offline training and online prediction. The offline stage uses the low-dimensional modal coefficient sequence in the multi-source temperature field database as a label, and uses historical modal coefficients, corresponding process parameters, curing stage identifiers, and time information. Using input features, a time-series prediction model is trained to learn the evolution of low-dimensional states over time and under process conditions. During online operation, the assimilated and corrected posterior modal coefficients from the previous time step, the current batch process parameters, the current curing stage, and time information are input into the trained time-series prediction model, outputting the prior modal coefficients for the current time step. Then, through a pre-constructed observation mapping relationship, the predicted temperature of each component's thermocouple measuring point is calculated using the observation matrix and prior modal coefficients, laying the groundwork for subsequent calculations of measuring point reliability and adapting to the requirements of online real-time operation. This achieves rapid prior prediction of the temperature field in a low-dimensional state space, obtaining the full-field temperature prediction result for the current time step without requiring online high-fidelity finite element simulation, thus improving computational efficiency. Simultaneously, it provides a baseline prior state for subsequent observation assimilation and correction, ensuring the continuous evolution capability of the temperature field digital twin.

[0029] S5: Calculate the measurement point reliability at the current moment based on the observed temperature and predicted temperature of the thermocouple measurement point of the component. Correct the prior modal coefficients at the current moment based on the measurement point reliability to obtain the posterior modal coefficients at the current moment. When the next moment arrives, update the posterior modal coefficients at the current moment to the posterior modal coefficients at the previous moment in the multi-source temperature field database, and return to step S4. Repeat steps S4 to S5. After correction using a reliability-weighted closed-loop assimilation mechanism, obtain the posterior modal coefficients at the current moment, and feed the posterior state back to the prediction process at the next moment (i.e., when the next moment arrives, use it as an input feature to return to step S4, repeating steps S4 to S5). This achieves closed-loop updating of the temperature field digital twin. Unlike traditional single-step post-processing corrections that only correct the current moment's result without changing the subsequent prediction trajectory, this method propagates the correction information from each observation along the time axis, continuously bringing the prediction trajectory closer to the actual operating conditions. This step utilizes real observation data to continuously offset prediction errors caused by process parameter mismatch, boundary heat transfer deviation, and material parameter fluctuations, thereby improving the temperature field model's adaptability to engineering disturbances and forming a true closed-loop update mechanism for the temperature field digital twin.

[0030] S6, when the current cycle count reaches a preset threshold, the temperature field is reconstructed based on the posterior modal coefficients at the current moment to obtain the posterior full-field temperature field. Based on the posterior full-field temperature field, a quality risk assessment is performed on the current curing process to output the quality risk assessment result, and a process optimization instruction for the current batch is generated based on the quality risk assessment result. By reconstructing the full-field temperature field based on the posterior modal coefficients and carrying out quality risk assessment and process optimization decisions, the limitations of traditional single-point temperature measurement are broken, realizing an upgrade from point-level monitoring to full-field temperature state assessment. It can identify potential quality risks in areas not covered by thermocouples, and at the same time, the risk assessment results are directly converted into executable process optimization instructions, opening up a complete link from temperature field perception to process control.

[0031] S7, execute the process optimization command. At the end of the current batch's curing process, update the temperature field multi-source database based on the current batch's multi-source data. At the start of the next batch's curing process, initialize the current cycle count to zero, determine the first predicted time of the next batch as the current time, and return to step S4. A two-layer closed-loop system is formed through online cyclic control and cross-batch data iteration. During implementation, the generated process optimization decisions are sent to the autoclave control system or pushed to process personnel for execution. Simultaneously, thermocouple observation data is continuously collected according to the preset sampling cycle, and assimilation correction, risk assessment, and process optimization are repeatedly executed. The process involves completing an online loop of "observation-confidence assessment-assimilation correction-posterior temperature field reconstruction-risk assessment-process optimization-process execution-new observation" in each sampling cycle, achieving real-time feedback closed-loop control for the current batch. After the current batch is cured, the process parameters, original observation data, measurement point confidence records, prior and posterior temperature field results, observation residual statistics, risk assessment records, process adjustment records, and post-molding quality inspection results of this batch are fully archived and updated to the temperature field multi-source database for subsequent incremental training of the model, observation noise calibration, risk threshold optimization, and process decision rule iteration.

[0032] Based on archived data, the reduced-order model, time-series prediction model, measurement point reliability assessment model, and process optimization strategy can be iteratively updated. For example, the training dataset can be expanded using newly added tank batch data to retrain or incrementally update the time-series prediction model; the observation noise matrix can be corrected based on thermocouple residual statistics; the risk assessment threshold can be optimized based on quality inspection results; and the process optimization decision rules can be adjusted based on process execution performance. The cross-tank batch learning closed loop can be represented as: tank batch data archiving → database update → model iteration → next tank batch parameter recommendation. Specifically, tank batch data archiving refers to storing the current tank batch observation data, assimilation results, evaluation results, and process execution results; database update refers to writing newly added tank batch data into the temperature field multi-source database; model iteration refers to updating the reduced-order model, time-series prediction model, measurement point reliability assessment model, or process optimization strategy; and next tank batch parameter recommendation refers to providing process parameter suggestions for subsequent curing processes based on the updated model and historical tank batch experience. By simultaneously constructing an online control closed loop for the current batch and a learning closed loop across batches, it not only ensures real-time control and quality stability of the curing process in a single batch, but also continuously improves the model accuracy and process optimization capabilities through the continuous accumulation of production data, adapting to the continuous process improvement needs in batch production scenarios.

[0033] In a preferred embodiment of the present invention, step S2 specifically includes the following steps: S21, based on the autoclave curing process curve of the composite component, the curing process is divided into multiple stages. During implementation, the preset curing process curve is read, the inflection point of the temperature slope change and the start and end times of the temperature plateau are identified, and combined with the characteristic stages of the curing reaction, the complete curing process is divided into independent stages such as the first heating stage, the first holding stage, the second heating stage, the second holding stage, and the cooling stage. Each stage corresponds to a set of temperature field evolution data with similar thermal response laws, and the spatial distribution evolution laws of the temperature field in different stages are significantly different. This eliminates the negative impact of the differences in thermal response laws between different process stages on the accuracy of a single order reduction model, providing a foundation for subsequently constructing a more adaptable order reduction model in stages, and effectively improving the fitting accuracy of the order reduction expression to the temperature field of each stage.

[0034] S22, a temperature field snapshot matrix is ​​generated based on the temperature field data of each process stage. During implementation, within each process stage, the full-field temperature vector at the corresponding time interval is extracted. The dimension of each temperature vector is equal to the total number of nodes in the component's finite element model, and each element in the vector corresponds to the temperature value of a spatial node. The temperature vectors from each sampling time are concatenated in chronological order as column vectors to form a two-dimensional temperature field snapshot matrix. The number of rows in the matrix represents the total number of nodes, and the number of columns represents the number of temperature field snapshots within that stage. Therefore, for the... For each process stage, the temperature field data within that stage is organized into a temperature field snapshot matrix, the expression of which is: in, Indicates the first Temperature field snapshot matrix for each process stage; Indicates the first The process stage is in the first The full-field temperature vector at each sampling time; Indicates the first [number]th ... Each sampling time; This indicates the number of temperature field snapshots within this stage. Before processing, the temperature field data needs to be preprocessed to remove abnormally distorted snapshots, align the time base, and standardize the temperature units to ensure data consistency within the matrix. Organizing the scattered time-series temperature field data into a standardized matrix format fully adapts to the input format requirements of reduction algorithms such as intrinsic orthogonal decomposition, ensuring the standardization and executability of subsequent reduction calculations.

[0035] S23, Based on the temperature field snapshot matrix, a staged reduced-order temperature field model is constructed. Based on the staged reduced-order temperature field model, the high-dimensional temperature field data is reduced in order to obtain a low-dimensional temperature field modal coefficient sequence, expressed as follows: in, This represents a sequence of modal coefficients for a low-dimensional temperature field. Represents the average temperature field; Represents the reduced-order basis function matrix; Indicates at time The modal coefficients are obtained; a staged reduced-order temperature field model is constructed based on the temperature field snapshot matrix, establishing a linear mapping relationship between high-dimensional temperature data and low-dimensional modal coefficients. During implementation, the temperature field snapshot matrix of each process stage is first mean-removed to calculate the average temperature field of that stage, and then a centered temperature deviation matrix is ​​obtained. The intrinsic orthogonal basis of the deviation matrix is ​​solved by singular value decomposition. The modes are sorted according to the singular values ​​from largest to smallest, and the energy proportion of each mode is calculated. The first few modes whose cumulative energy proportion reaches a preset threshold are selected as the reduced-order basis function matrix. Finally, the overall temperature vector is represented as a linear combination of the average temperature field, the reduced-order basis functions, and the modal coefficients. The average temperature field represents the reference temperature distribution of that stage, the reduced-order basis functions represent the spatial perturbation mode of the temperature field, and the modal coefficients are the time-varying weights of the corresponding modes. This can characterize most of the spatial distribution characteristics of the temperature field, greatly compressing the dimensionality of the state variables while ensuring computational accuracy, and providing a mathematical basis for rapid temperature field reconstruction and state correction in online stages.

[0036] In a preferred embodiment of the present invention, the prior mode coefficients at the current time in step S4 are expressed as follows: in, Represents the prior modal coefficients at the current moment; Represents the posterior modal coefficients after assimilation correction at the previous time step; Indicates the process parameters for the current tank batch; Indicates the current solidification stage; Indicates the current time; Indicates the time step; The parameter is A time-series prediction model is proposed. During implementation, a network structure with long temporal memory is used to construct the model. Input features include the posterior modal coefficient vector from the previous time step, the process parameter vector from the current time step, the unique thermal encoding of the curing stage, and the cumulative curing time. The output is the prior modal coefficient vector from the current time step. Offline training uses continuous modal coefficient sequences from historical simulations or batches, with mean squared error as the loss function for supervised training to improve the model's temporal stability. During online execution, the posterior state from the previous time step is directly substituted into the model to recursively obtain the prior modal coefficients from the current time step. This implementation method can directly utilize the latest state after assimilation and correction for recursive prediction, allowing the correction information of the posterior state to be naturally integrated into the subsequent prediction process, ensuring the close fit between the prior prediction results and actual operating conditions. Furthermore, the recursive prediction method has a short computational chain and fast inference speed, meeting the requirements for online real-time operation.

[0037] In a preferred embodiment of the present invention, step S5 specifically includes the following steps: S51, determine the location of thermocouple measuring points for each component, and establish an observation mapping relationship between the thermocouple measuring points of each component and the low-dimensional temperature field modal coefficients, so as to obtain the predicted temperature of the thermocouple measuring points based on the observation mapping relationship; optimize the point layout by combining the component geometry, temperature field change amplitude, key mass areas, on-site wiring feasibility and measuring point redundancy, and prioritize the selection of measuring points that are highly sensitive to the observation of each order of reduced mode and are located in key mass locations such as thick-walled areas and stringer junction areas; after the measuring point locations are determined, construct a selection matrix to extract the measuring point temperature from the total field temperature, the number of rows of the matrix equals the number of measuring points, the number of columns equals the number of nodes in the total field, the element corresponding to the measuring point is 1, and the other positions are 0, and then multiply the selection matrix with the reduced-order basis function matrix to derive the observation matrix, and establish a mapping relationship for calculating the predicted temperature of the measuring points from the modal coefficients. By establishing a direct correlation between a limited number of thermocouple measuring points on the components and the low-dimensional state of the entire temperature field, the observation data of sparse measuring points can be used as an effective input for assimilation correction, solving the problem that traditional measuring points can only be used for single-point monitoring and cannot support full-field temperature correction.

[0038] In detail, since the prior temperature field can be reconstructed from the reduced-order model, the predicted temperature at the thermocouple measuring point is... It can be represented as follows: in, This represents the prior mode coefficients at the current moment.

[0039] After rearranging the above relationships, we get the following: in, This represents the temperature vector at the thermocouple measuring point, representing the average temperature field. Represents the observation matrix; This represents the prior mode coefficients at the current moment.

[0040] In the above relationship, the average measurement point temperature vector can be expressed as follows: The observation matrix can be represented as follows: in, This represents the measurement point selection matrix. Represents the average temperature field. Represents the reduced-order basis function matrix. Represents the observation matrix. Through the observation matrix... This allows for the establishment of a correspondence between low-dimensional modal coefficients and the predicted temperature of thermocouple measuring points on components, enabling a limited number of thermocouple measuring points on components to participate in the correction of the overall temperature field of the component.

[0041] S52, collect the observed temperatures of thermocouple measuring points on the surface or inside the composite material component parts, and calculate the observation residuals based on the predicted temperatures and the observed temperatures; collect real-time temperature data of each component's thermocouples according to a preset sampling interval, and perform preprocessing such as timestamp alignment, unit unification, and range outlier marking to ensure that the measured data strictly corresponds to the predicted time; then, subtract the measured temperature from the corresponding predicted temperature according to the measuring point number to obtain the observation residuals of each measuring point and form a residual vector. The sign of the residual represents the relative level of the measured temperature to the predicted temperature, and the amplitude of the residual represents the magnitude of the deviation. This quantifies the deviation between the prior temperature field and the actual component temperature, providing a driving force for subsequent state correction, enabling assimilation correction to specifically compensate for model prediction errors. The larger the deviation, the stronger the driving force for correction. The thermocouple observation vector at the current moment can be expressed as follows: in, This represents the thermocouple observation vector at the current moment; Indicates the first Thermocouple measuring points of the component at time The observed temperature; Indicates the number of thermocouple measuring points on the component; superscript Represents the transpose of a vector or matrix.

[0042] S53, calculate the reliability of the measurement point, and determine whether there is a reliable observation based on the comparison result between the reliability of the measurement point and the preset threshold; if so, calculate the assimilation gain based on the reliability of the measurement point, and combine the assimilation gain and the observation residual to correct the prior mode coefficient at the current time to obtain the posterior mode coefficient at the current time; evaluate the consistency of the observation data of each thermocouple measurement point from multiple dimensions such as physical laws, process characteristics and model deviation, and calculate the reliability weight with a value range between 0 and 1; preset the reliability threshold, compare the reliability weight of each measurement point with the threshold, and include the measurement points with a reliability higher than the threshold into the effective observation set; if the effective observation set is not empty, calculate the assimilation gain based on the residual of the effective observation, the corresponding observation matrix sub-block and the observation noise after reliability correction, and perform weighted correction on the prior mode coefficient at the current time to obtain the posterior mode coefficient; it is worth noting that if there is no reliable observation at the current time, or all measurement points are missing, the prior mode coefficient at the current time can be directly used as the posterior mode coefficient, and enter the no-observation rolling mode. This avoids interference from abnormally low-confidence measurement points on the overall assimilation results, while retaining effective observation information reflecting the deviation of the actual working conditions, thus improving the robustness and accuracy of the assimilation correction process; the observation-free rolling processing method ensures the continuous operation capability of the digital twin under abnormal scenarios such as measurement point failure and data interruption, avoiding system operation interruption.

[0043] S54, at the next time step, the posterior modal coefficients of the current time step are updated to the posterior modal coefficients of the previous time step in the multi-source database of the temperature field, and the process returns to step S4. Steps S4 to S5 are executed cyclically to achieve closed-loop update of the temperature field digital twin. After calculating the posterior modal coefficients of the current time step, they are used as the latest input state of the time series prediction model, replacing the original open-loop prediction state. Combined with the process parameters, curing stage, and time information at the next time step, the prior modal coefficients at the next time step are calculated, completing a closed loop of one recursive cycle. Each time observation assimilation is completed, the initial state of the prediction is updated, so that the correction effect of a single observation can continuously affect the prediction trajectory of all subsequent time steps. This forms a continuous recursive mechanism of "prediction-correction-re-prediction", breaking the error accumulation of open-loop prediction and solving the deficiency of single-step post-processing correction that only improves the current time step and cannot correct subsequent prediction trajectories. It truly realizes the closed-loop evolution and continuous self-correction of the temperature field digital twin, which is particularly advantageous in engineering scenarios with large observation intervals.

[0044] In a preferred embodiment of the present invention, step S6 specifically includes the following steps: S61, based on the posterior modal coefficients at the current moment after the loop is completed, the posterior full-field temperature field is reconstructed; first, according to the current process stage (after the loop is completed), the average temperature field vector and the reduced-order basis function matrix corresponding to the stage are called; the posterior modal coefficient vector and the reduced-order basis function matrix are multiplied by matrix to obtain the temperature perturbation field; then the temperature perturbation field is superimposed on the average temperature field to obtain the temperature values ​​of all spatial nodes, that is, the assimilated and corrected posterior full-field temperature field.

[0045] S62, extract temperature features for quality monitoring and risk assessment based on the posterior full-field temperature field; S63, based on the temperature characteristics, a quality risk assessment is performed on the current curing process to classify risk levels; a node-by-node statistical and regional analysis is conducted on the post-hoc overall temperature field to extract multiple types of temperature characteristics. Combined with expert knowledge, process window constraints, and quality control standards, a quality risk assessment is performed on the current curing process. Quality risks may include over-temperature risk, under-temperature risk, local thermal hysteresis risk, temperature non-uniformity risk, insufficient insulation risk, incomplete curing risk, and thermal gradient risks that may lead to abnormal residual stress. Overall statistical characteristics include the highest temperature, lowest temperature, maximum temperature difference, average temperature, and temperature standard deviation; regional characteristics include the average temperature, highest temperature, and lowest temperature of key quality areas; dynamic rate characteristics include the average heating rate and heating rate of key areas; local anomalies include leading point temperature, lagging point temperature, location and area of ​​local thermal hysteresis areas, and location and area of ​​local over-temperature areas; deviation characteristics include temperature deviation from the set process curve and time difference from the target temperature.

[0046] For example, rule-based judgments can be made based on preset thresholds. When the maximum temperature of the posterior temperature field exceeds the upper limit of the allowable temperature for the current process stage, it is determined that there is an over-temperature risk. The judgment relationship is as follows: in, This represents the maximum temperature in the posterior temperature field at the current moment; Indicates the current process stage The upper limit of the allowable temperature.

[0047] When the minimum temperature of the post-test temperature field is lower than the lower limit of the allowable temperature for the current process stage, it is determined that there is a risk of undertemperature. The judgment relationship is as follows: in, This represents the minimum temperature in the posterior temperature field at the current moment; Indicates the current process stage The lower limit of the permissible temperature.

[0048] When the overall temperature difference exceeds the maximum allowable temperature difference for the current process stage, it is determined that there is a risk of temperature unevenness. The judgment relationship is as follows: in, This represents the posterior temperature difference across the entire field at the current moment. Indicates the current process stage The maximum allowable temperature difference. This transforms massive amounts of temperature data across all nodes into structured quality characteristic indicators, comprehensively covering quality-related factors such as temperature level, uniformity, dynamic rate, and local anomalies. This avoids the one-sidedness of a single indicator and provides multi-dimensional basis for subsequent standardized risk assessments and process decisions.

[0049] In another implementation, the risk level can be comprehensively judged based on rule models, fuzzy logic models, expert systems, machine learning classifiers, or optimization models. Risk levels can be divided into normal, low, medium, and high risk, or other levels can be set according to project needs.

[0050] S64, determine whether the current risk level is within the acceptable range; if yes, generate a process optimization instruction to maintain the current process parameters; if no, identify the risk type and generate a process optimization instruction for the current moment; compare the assessed risk level with the preset acceptable standard; if the risk level is normal or at the level of concern, it is determined to be within the acceptable range, and a process optimization instruction to maintain the current process parameters is generated. After the instruction is executed, the autoclave continues to operate according to the original process curve; if the risk level is a warning or severe warning level, it is determined to be within the unacceptable range, and further identify the specific dominant risk type, and match the preset process control strategy library to generate optimization instructions. For example: when the posterior temperature field shows that the temperature of a local lag area is lower than the process target temperature for a long time, and the current stage is heat preservation, an optimization suggestion to extend the heat preservation time can be output; when the posterior temperature field shows that there is an over-temperature risk in a local area, an optimization suggestion to reduce the heating rate, reduce the target temperature, or enter manual confirmation can be output; when the posterior temperature field shows that the overall heating is lagging and there is no local over-temperature risk, a suggestion to appropriately increase the heating rate or adjust the heating power can be output. This step enables direct linkage between risk assessment results and process control strategies, with clear and traceable decision-making logic. It further transforms the digital twin results of the temperature field into executable or referable process optimization strategies, and the output process instructions can be directly sent to the autoclave control system or guide process personnel to operate, ensuring that the curing process is always within a controllable quality range.

[0051] Furthermore, the predicted temperature of the thermocouple measuring point of the component in step S51... Expressed as: in, This represents the temperature vector at the thermocouple measuring point, representing the average temperature field. The observation matrix is ​​represented by the following steps: First, the average temperature field at each thermocouple measuring point is calculated for the corresponding process stage, forming an average measuring point temperature vector. Then, the observation matrix is ​​obtained by multiplying the measuring point selection matrix with the reduced-order basis function matrix. The dimension of the observation matrix is ​​the number of measuring points multiplied by the modal order. During online prediction, the prior modal coefficients at the current moment are multiplied by the observation matrix to obtain the measuring point temperature deviation corresponding to the modal perturbation. This deviation is then superimposed with the average measuring point temperature vector to obtain the predicted temperature at each measuring point. By establishing a linear mapping relationship in this way, the predicted temperature at the measuring points can be quickly calculated without reconstructing the entire temperature field, further reducing the online computation load and improving the efficiency of the assimilation step. This provides a direct computational basis for the subsequent calculation of observation residuals and assimilation correction.

[0052] Further, the calculation of the reliability of the measurement point in step S53 includes: The reliability of the measurement points is calculated using a weighted summation method. Its expression is: in, Indicates the first Each measuring point at time... The Class consistency score; Indicates the first The weights corresponding to the class consistency score; Indicates the number of consistency evaluation items; The weights for each consistency score can satisfy the following normalization constraint, and their formula is: The implementation method involves pre-setting weight coefficients for various evaluation items, such as temperature range, temporal continuity, heating rate, and spatial consistency, with the sum of all weight coefficients being 1 to satisfy normalization constraints. The consistency score for each measuring point on each evaluation item is calculated separately; a higher score indicates greater reliability of the measuring point data under that dimension. Finally, each score is multiplied by its corresponding weight and summed to obtain the reliability weight value of the measuring point at the current moment. The temperature range consistency score is calculated based on whether the measuring point temperature falls within the reasonable temperature range of the current process stage; exceeding this range lowers the score. The temporal continuity score is calculated based on whether the temperature change between adjacent moments exceeds a reasonable rate; a larger jump in temperature results in a lower score. This implementation method has clear calculation logic, is easy to implement in engineering, and can flexibly adjust weights according to the importance of different evaluation items, adapting to the reliability assessment needs of measuring points under different components and process scenarios.

[0053] Furthermore, the correction process in step S53 specifically includes the following steps: S531, based on the reliability of the measurement points The observation noise matrix is ​​dynamically adjusted, and its expression is as follows: in, Indicates the first Each measuring point at time... Confidence correction for observation noise; Indicates the first The basic observation noise corresponding to each measurement point (i.e., the basic observation noise covariance matrix R of the nth measurement point) line, number (column diagonal elements); Indicates the first Each measuring point at time... Credibility weight; This indicates a small positive number that prevents the denominator from being zero; its value can be 0. In implementation, a baseline observation noise value is set for each measuring point, representing the measurement noise level when the measuring point is working normally. The baseline observation noise is then divided by the sum of the measuring point's confidence level and a small positive number to obtain the diagonal elements of the confidence-corrected observation noise matrix. The principle behind this implementation is that observation noise represents the uncertainty of the observation value; the greater the noise, the less reliable the observation, and the smaller the corresponding correction weight under the assimilation framework. When the confidence level of a measuring point decreases, its observation noise is amplified, automatically reducing the proportion of that measuring point in the assimilation correction. This method achieves continuous weighting of confidence levels, rather than a hard 0-1 elimination, which reduces the negative impact of low-confidence measuring points and avoids the loss of observation information that may result from measuring point elimination, making the assimilation correction process smoother.

[0054] S532, calculate the confidence-weighted assimilation gain, which is expressed as: in, This represents the assimilation gain after considering the reliability of the measurement points; This represents the prior state error covariance matrix at the current moment; Represents the observation matrix; Represents the observation matrix transpose; This represents the observation noise matrix after confidence correction. , This indicates a diagonal matrix construction operation, where each element within the parentheses is sequentially treated as a main diagonal element of the matrix, and the remaining off-diagonal elements are set to zero. Indicates the number of thermocouple measuring points; This involves matrix inversion. First, the prior state error covariance matrix at the current time step is obtained. This matrix is ​​derived from the posterior error covariance at the previous time step through state transition propagation and represents the uncertainty of the prior mode coefficients. Then, combined with the observation matrix and the observation noise matrix after confidence correction, the weighted assimilation gain matrix is ​​calculated according to the Kalman gain formula. The physical meaning of the assimilation gain is to balance the uncertainty of the prior state and the uncertainty of the observation, allocating their weights in the state correction: the more reliable the observation and the smaller the observation noise, the larger the gain and the stronger the contribution of the observation to the state correction; the more accurate the prior state and the smaller the prior error, the smaller the gain and the smaller the correction magnitude. By incorporating the confidence of the measurement point into the gain calculation through observation noise, the core logic of weighted assimilation is realized, ensuring that the magnitude of the correction matches the confidence level of the observation.

[0055] S533, the prior mode coefficients are corrected based on the aforementioned confidence-weighted assimilation gain, expressed as follows: in, Represents the posterior modal coefficients after assimilation correction at the current time. , This represents the observation residual at the current moment; This indicates the real-time temperature of the thermocouple at the current moment. The current thermocouple-predicted temperature is represented by the weighted assimilation gain matrix. The observation residual vector is multiplied to obtain the correction amount for each modal coefficient; positive residuals correspond to positive corrections, and negative residuals correspond to negative corrections. The correction amount is then added to the prior modal coefficients to obtain the posterior modal coefficients at the current moment. The entire correction process is completed in a low-dimensional modal space, and the computational cost is only related to the modal order and the number of measurement points, far less than the assimilation operation in the full-field dimension. This implementation method efficiently corrects the prior state, bringing the low-dimensional state closer to the actual operating condition, and then reconstructs a more accurate full-field temperature field through order reduction mapping. Simultaneously, the extremely low computational cost ensures the feasibility of online real-time operation.

[0056] Furthermore, the calculation of the reliability of the measurement point in step S53 also includes: The reliability of the measurement points is calculated by comprehensively considering multiple consistency evaluation items. Its expression is: in, Indicates the first Temperature range consistency score for each measuring point; Indicates the first Time continuity score for each measurement point; Indicates the first Consistency score of heating rate at each measuring point; Indicates the first Spatial consistency score of each measurement point; Indicates the first Consistency score of process stages at each measuring point; Indicates the first Consistency score of model residuals at each measurement point; The comprehensive evaluation function is used to calculate six consistency scores for each measuring point: temperature range consistency (verifying whether the measuring point temperature is within the physically reasonable range of the current process stage); time continuity consistency (verifying the smoothness of temperature changes over time and identifying single-point jumps); heating rate consistency (verifying whether the temperature change rate matches the process characteristics of the current stage); spatial consistency (verifying whether the temperature deviation between this measuring point and adjacent measuring points is within a reasonable range and identifying local anomalies); process stage consistency (verifying whether the temperature change trend matches the heating, holding, or cooling actions of the current stage); and model residual consistency (verifying whether the residual conforms to the historical statistical distribution and identifying systematic biases). These six scores are then fused through a preset comprehensive evaluation function to obtain the final measuring point reliability result. This multi-dimensional cross-validation method can more accurately distinguish between single-point sensor failures and actual overall operating condition deviations, avoiding the accidental deletion of effective observations and significantly improving the comprehensiveness and accuracy of reliability assessment.

[0057] Furthermore, the process optimization decision in step S64, which issues the process optimization command at the current moment, is specifically expressed as follows: in, This indicates the process optimization instruction at the current moment; This represents the posterior global temperature field at the current moment; This represents the set of temperature features extracted from the posterior full-field temperature field. This indicates the risk assessment result at the current moment; Indicates process window constraints; Indicates equipment operating constraints; This represents the process optimization decision function. It takes the current posterior global temperature field, the set of temperature features extracted from the posterior global temperature field, and the current risk assessment result as state inputs, while also incorporating constraints: process window constraints, including the maximum and minimum allowable curing temperatures, maximum heating / cooling rates, and minimum curing requirements, ensuring the optimized solution does not exceed the material's process window; and equipment operation constraints, including the autoclave's maximum heating power, maximum fan speed, temperature regulation accuracy, and pressure regulation range, ensuring the output instructions are within the equipment's capabilities. By solving for the optimal process adjustment within the constraints using the process optimization decision function, the optimal process optimization instructions for the current moment are generated, simultaneously satisfying quality requirements, material limitations, and equipment capabilities. This results in stronger engineering feasibility and avoids theoretically optimal but impractical solutions.

[0058] To facilitate understanding by those skilled in the art, experimental records of this application and two other prior art solutions are listed below: The experimental subject was an independent test condition during the composite material autoclave curing process. Three schemes were compared in the experiment: Scheme A: AR-LSTM open-loop prediction scheme, which only uses the autoregressive LSTM temperature field prediction model to output the temperature field without introducing thermocouple measured data for correction. This scheme is used to characterize the open-loop prediction error without assimilation. Scheme B: Kalman post-processing correction scheme, which uses thermocouple observation data to perform Kalman assimilation correction on the predicted temperature field at the current time, but the corrected posterior state is not fed back to the prediction model at the next time step. This scheme is used to characterize the effect of "correction at the current time step, but no closed-loop feedback". Scheme C: The closed-loop feedback assimilation scheme of this application, forming a closed-loop digital twin process of the temperature field: "prediction—observation—difference innovation—assimilation correction—posterior feedback".

[0059] The positive and negative process parameter mismatch experiment is used to simulate the situation where the input process parameters are inconsistent with the actual operating conditions during the curing process. In the experiment, only the first six process parameters of the AR-LSTM are perturbed, the assimilation interval is fixed at 5, and the actual temperature field, actual POD coefficient, and stage division are still derived from the training file.

[0060] Assume the actual process parameters are: in, , For the heating rate, , To maintain the temperature, , This refers to the heat preservation time.

[0061] The mismatch levels set in the experiment included: This means simultaneously examining the impact of positive and negative deviations of process parameters on temperature field prediction.

[0062] In the positive and negative process parameter mismatch experiment, a total of 27 valid comparison scenarios were formed. The results show that Scheme C has a better closed-loop correction effect than Scheme B in most scenarios, as shown in Table 1 below: Table 1 Experimental results under different process parameters It can be seen that as the process mismatch becomes more severe, the errors between Scheme A and Scheme B increase significantly. In contrast, the error of Scheme C, the closed-loop feedback assimilation scheme of this invention, is still suppressed to a certain extent. This result indicates that when the process parameter mismatch is large, the advantages of closed-loop feedback assimilation are more obvious. Please refer to the appendix for details. Figure 2It is a comparison chart of the post-hoc MAE of three schemes A / B / C under the mismatch of positive and negative process parameters.

[0063] In engineering scenarios, there is a certain delay between the data collected by the sensors and the received data. The more sensors there are, the more the delay needs to be matched. Therefore, it is impossible for sensor data to be transmitted in every moment for data assimilation. To adapt to such engineering scenarios, different assimilation intervals of 1 / 5 / 10 were selected to test the robustness of the method. The experimental results are shown in Table 2 below: Table 2 Experimental results at different assimilation intervals It can be seen that when the assimilation interval is 1, Scheme B has a relatively low error because it can use the observation data for post-processing correction at every moment, while Scheme C shows a smaller absolute improvement compared to Scheme B. However, when the assimilation interval increases to 5 or 10, Scheme B struggles to continuously correct the predicted trajectory when there are no observations, while Scheme C can transmit historical observation information to the subsequent prediction process through posterior state feedback, thus significantly enhancing its advantages. This result supports the necessity of "closed-loop feedback" in this invention: the advantage of Scheme C is not only reflected in the presence of observations, but also in engineering scenarios where observations are discontinuous or the sensor sampling interval is large.

[0064] Experiments on the changes in auxiliary material properties and boundary heat transfer conditions are used to simulate situations where the convective heat transfer conditions on the component surface, the thermal properties of auxiliary materials, the heat transfer state of the mold, or the local flow field conditions are inconsistent with the nominal model during the actual curing process.

[0065] In the experiment, a baseline operating condition was selected, and then the heat transfer boundary conditions were changed based on the baseline operating condition to generate a heat transfer-biased operating condition and a heat transfer-biased operating condition. None of the three schemes A, B, and C used the baseline operating condition as a priori. The heat transfer-biased and heat transfer-biased operating conditions were assimilated as the sensor observation inputs B / C in the real engineering environment.

[0066] This experiment is closer to the actual engineering situation: the boundary heat transfer state of real components may be different from the benchmark simulation model, but local temperature information can only be obtained through observation of a limited number of thermocouples.

[0067] The specific experimental results are shown in Table 3 below: Table 3. Experimental results of boundary heat transfer mismatch The results show that in the group with stronger heat transfer, scheme A has a larger open-loop prediction error of 2.061K. When the assimilation interval is 5, the error of scheme B is 1.725K, while that of scheme C decreases to 0.381K, representing an improvement of 77.887%. In the group with weaker heat transfer, the initial open-loop error is relatively small, with scheme A having an error of 0.459K. When the assimilation interval is 5, the error of scheme B is 0.398K, while that of scheme C decreases to 0.170K, representing an improvement of 57.265%. When the assimilation interval is 1, scheme C is slightly worse than scheme B by 0.004K in the group with weaker heat transfer, indicating that when there is observational correction at every time step and the initial error is small, the posterior advantage brought by closed-loop feedback is not necessarily significant. However, this difference is very small and does not affect the overall conclusion. When the assimilation interval is increased to 5 or 10, Scheme C is significantly better than Scheme B in both the heat transfer strong and heat transfer weak groups, indicating that the closed-loop feedback mechanism can propagate the correction of historical observations to the non-assimilation time.

[0068] Therefore, the above experiments can support the core technical effect of the present invention: the present invention enables the digital twin model of the composite material autoclave curing temperature field to have online self-correction capability through the closed-loop mechanism of "thermocouple observation residual - low-dimensional assimilation correction - posterior state feedback", which can adapt to engineering disturbances such as process parameter mismatch, changes in auxiliary material properties and changes in boundary heat transfer conditions.

[0069] Furthermore, a verification experiment was conducted to optimize the curing process parameters. This experiment was used to demonstrate that after obtaining information on the component's temperature field and curing state, key indicators such as temperature uniformity, curing equilibrium, and curing efficiency can be further extracted, and a better curing process parameter scheme can be generated based on a multi-objective optimization method, thereby supporting the technical effectiveness of the process optimization decision module in this application.

[0070] The experiment focuses on the curing process of thermosetting prepreg composite components, optimizing process parameters based on a two-stage heating and holding curing regime. This experiment serves as an offline or quasi-online verification method for the process optimization decision module of this invention, illustrating the feasibility of generating process parameter optimization schemes based on the posterior full-field temperature field and its derived quality characteristics. It does not limit this invention to necessarily employing this specific surrogate model or optimization algorithm. The optimization design variables include the first-stage heating rate, the first-stage holding temperature, the first-stage holding time, the second-stage heating rate, the second-stage holding temperature, and the second-stage holding time.

[0071] The curing process parameters before and after optimization are shown in Table 4 below: Table 4 Optimized curing process parameters The original curing regime employed a single-stage heating and holding strategy, directly raising the temperature to 180°C at a rate of 2.0 K / min and then holding it at that temperature for 12320 s. This process path was relatively simple, but it was prone to significant transient temperature difference accumulation during the heating phase. The optimized curing regime, however, adopted a coupled approach of segmented heating and intermediate plateau holding. In the first stage, the temperature was raised to 127°C at 1.7 K / min and held for 1966 s, followed by a second stage where the temperature was raised to 183°C at 2.0 K / min and held for 12582 s. The optimization results demonstrate the following process adjustment characteristics: First, the initial heating rate was reduced from 2.0 K / min to 1.7 K / min, which helps reduce the accumulation of temperature gradient during the heating stage; Second, the addition of a 127℃ intermediate platform insulation section allows for more sufficient diffusion and equalization time of the internal temperature of the component, which helps reduce the risk of local thermal lag and temperature unevenness; Third, the target temperature for the second stage was adjusted from 180℃ to 183℃, which helps compensate for the reduced reaction rate caused by the initial gentle heating and ensures sufficient curing in the later stage; Fourth, the insulation time for the second stage was adjusted from 12320s to 12582s, which helps ensure that all areas of the component meet the requirements for sufficient curing under the condition of prioritizing quality.

[0072] The above-mentioned adjustment direction is consistent with the process optimization strategies such as "heating rate adjustment, holding time correction, and target temperature correction" described in this invention, which demonstrates that the process optimization decision module of this invention has the ability to generate practical and interpretable process adjustment schemes.

[0073] The key performance indicators before and after optimization are compared in Table 5 below: Table 5 Comparison of Key Optimization Indicators As can be seen from the table, the optimized scheme significantly improved the temperature uniformity and curing equilibrium during the curing process. The maximum temperature difference before optimization was 24.7℃, while after optimization, as verified by numerical simulation, it was 15.0℃, representing a reduction of approximately 39.3%. This result indicates that the optimized segmented heating and intermediate platform insulation strategy can effectively reduce the non-uniformity of the internal temperature field of the component, which is beneficial for reducing the risk of thermally induced residual stress and localized temperature anomalies.

[0074] Regarding the uniformity of curing degree, the curing degree difference was 0.26 before optimization and 0.16 after optimization, as verified by numerical simulation, representing a decrease of approximately 38.5%. This result indicates that the optimized curing regime can improve the synchronicity of curing reactions in different regions and reduce the risk of local under-curing or uneven curing.

[0075] Regarding curing efficiency, the time required to reach the curing threshold was 9280s before optimization and 9570s after optimization, showing a slight increase in curing time. This indicates that the optimization scheme does not simply pursue the shortest curing cycle, but rather, under the principle of quality priority, it achieves a significant improvement in temperature uniformity and curing equilibrium by moderately extending the curing process. This result aligns with the multi-objective trade-off between quality and efficiency in the curing process of composite components.

[0076] Furthermore, the surrogate model exhibits low relative errors compared to high-fidelity numerical simulations at the recommended optimization schemes, with relative errors of 3.8% for temperature difference, 1.8% for curing degree difference, and 1.7% for curing time. These results indicate that the surrogate evaluation model can effectively support the rapid screening of candidate process schemes and can be combined with high-fidelity simulations or actual furnace data for final verification.

[0077] Based on the above results, the following conclusions can be drawn.

[0078] First, the multi-objective optimization method based on temperature field and curing state characteristics can transform the quality requirements of composite material curing process into calculable, comparable, and screenable process optimization objectives. Second, the surrogate evaluation model can quickly predict the key quality responses corresponding to candidate process parameters while ensuring a certain level of accuracy, thereby improving the efficiency of process optimization decision-making. Third, the multi-objective optimization and comprehensive selection method can balance temperature uniformity, curing equilibrium, and curing efficiency to generate recommended process parameter schemes with engineering interpretability. Fourth, the optimized process, by reducing the initial heating rate, introducing intermediate platform insulation, and appropriately adjusting the target temperature and insulation time, reduces the maximum temperature difference by approximately 39.3% and the curing degree difference by approximately 38.5%, indicating that process optimization decision-making can significantly improve the temperature uniformity and curing equilibrium during component curing. Fifth, this experiment serves as a verification method for the process optimization decision-making module of this invention, demonstrating that this invention can not only obtain the posterior full-field temperature field through closed-loop assimilation but also further transform the posterior temperature field and quality risk assessment results into executable or referable process parameter optimization suggestions, thereby forming a feedback closed loop of "temperature field digital twin - risk assessment - process optimization - process execution".

[0079] It should be noted that the above experiments are verification results under preferred embodiments of the present invention, used to illustrate the technical effects of the present invention, and do not constitute a limitation on the scope of protection of the present invention. Those skilled in the art can make adaptive adjustments to the model structure, assimilation algorithm, optimization algorithm, and process parameter range according to the specific component type, autoclave equipment, number of sensors, material system, and curing regime.

[0080] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for temperature field digital twin closed-loop assimilation and process feedback optimization in the curing process of composite materials, characterized in that, Includes the following steps: S1. Establish a multi-source temperature field database based on the curing process of the composite material component in the current batch. The multi-source temperature field database includes the posterior modal coefficients of the previous moment and the current curing characteristic data. The current curing characteristic data includes the current process parameters, the current curing stage, and time information. S2, the curing process of composite material components is divided into multiple process stages, and the temperature field data of each process stage is input into the staged reduced-order temperature field model for order reduction processing in combination with the multi-source temperature field database to obtain a low-dimensional temperature field modal coefficient sequence. S3, Construct a time-series prediction model based on the low-dimensional temperature field modal coefficient sequence; S4, For the current moment, input the posterior modal coefficients of the previous moment and the current solidification feature data into the time series prediction model to predict the prior modal coefficients of the current moment, and determine the predicted temperature of the thermocouple measuring point of the part based on the prior modal coefficients of the current moment. S5. Calculate the reliability of the measurement point at the current moment based on the observed temperature and the predicted temperature of the thermocouple measurement point of the component. Correct the prior modal coefficients at the current moment based on the reliability of the measurement point at the current moment to obtain the posterior modal coefficients at the current moment. When the next moment arrives, update the posterior modal coefficients at the current moment to the posterior modal coefficients at the previous moment in the multi-source temperature field database, and return to step S4 to perform the process. Repeat steps S4 to S5. S6, when the current number of cycles reaches the preset number threshold, the temperature field is reconstructed based on the posterior modal coefficients at the current moment to obtain the posterior full-field temperature field. The quality risk assessment of the current curing process is performed based on the posterior full-field temperature field to output the quality risk assessment result, and the process optimization instruction for the current batch is generated according to the quality risk assessment result. S7, execute the process optimization instruction, update the temperature field multi-source database based on the multi-source data of the current batch when the curing process of the current batch ends, initialize the current cycle number to zero when the curing process of the next batch begins, determine the first predicted time of the next batch as the current time, and return to execute step S4.

2. The method for temperature field digital twin closed-loop assimilation and process feedback optimization in the composite material curing process according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21, Based on the autoclave curing process curve of composite material components, the curing process is divided into multiple process stages; S22, generate a temperature field snapshot matrix based on the temperature field data of each process stage; S23, Based on the temperature field snapshot matrix, a staged reduced-order temperature field model is constructed. Based on the staged reduced-order temperature field model, the high-dimensional temperature field data is reduced in order to obtain a low-dimensional temperature field modal coefficient sequence, expressed as follows: in, This represents a sequence of modal coefficients for a low-dimensional temperature field. Represents the average temperature field; Represents the reduced-order basis function matrix; Indicates at time The modal coefficients.

3. The method for temperature field digital twin closed-loop assimilation and process feedback optimization in the composite material curing process according to claim 1, characterized in that, The prior mode coefficients at the current time in step S4 are expressed as follows: in, Represents the prior modal coefficients at the current moment; Represents the posterior modal coefficients after assimilation correction at the previous time step; Indicates the process parameters for the current tank batch; Indicates the current solidification stage; Indicates the current time; Indicates the time step; The parameter is The time series prediction model.

4. The method for temperature field digital twin closed-loop assimilation and process feedback optimization in the composite material curing process according to claim 3, characterized in that, Step S5 specifically includes the following steps: S51, determine the arrangement position of thermocouple measuring points of each component, and establish the observation mapping relationship between thermocouple measuring points of each component and the low-dimensional temperature field modal coefficient, so as to obtain the predicted temperature of thermocouple measuring points of the component based on the observation mapping relationship. S52, collect the observed temperature of each thermocouple measuring point arranged on the surface or inside of the composite material component parts, and calculate the observation residual based on the predicted temperature and the observed temperature; S53, calculate the reliability of the measurement point, and determine whether there is a reliable observation based on the comparison result between the reliability of the measurement point and the preset threshold; If so, the assimilation gain is calculated based on the confidence level of the measurement point, and the prior mode coefficients at the current time are corrected by combining the assimilation gain and the observation residuals to obtain the posterior mode coefficients at the current time. S54, when the next moment arrives, update the posterior modal coefficients of the current moment to the posterior modal coefficients of the previous moment in the temperature field multi-source database, and return to execute step S4, so as to cyclically execute steps S4 to S5 to realize the closed-loop update of the temperature field digital twin.

5. The method for temperature field digital twin closed-loop assimilation and process feedback optimization in the composite material curing process according to claim 4, characterized in that, Step S6 specifically includes the following steps: S61, the posterior full-field temperature field is obtained by reconstructing the posterior modal coefficients at the current moment after the loop is completed; S62, extract temperature features for quality monitoring and risk assessment based on the posterior full-field temperature field; S63, perform a quality risk assessment on the current curing process based on the temperature characteristics to classify the risk level; S64, determine whether the current risk level is within an acceptable range; if so, generate a process optimization instruction to maintain the current process parameters; If not, identify the risk type and generate the process optimization instruction for the current moment.

6. The method for temperature field digital twin closed-loop assimilation and process feedback optimization in the composite material curing process according to claim 4, characterized in that, The predicted temperature of the thermocouple measuring point of the component in step S51 Expressed as: in, This represents the temperature vector at the thermocouple measuring point, representing the average temperature field. This represents the observation matrix.

7. The method for temperature field digital twin closed-loop assimilation and process feedback optimization in the composite material curing process according to claim 4, characterized in that, The calculation of the reliability of the measurement point in step S53 includes: The reliability of the measurement points is calculated using a weighted summation method. Its expression is: in, Indicates the first Each measuring point at time... The Class consistency score; Indicates the first The weights corresponding to the class consistency score; Indicates the number of consistency evaluation items; .

8. The method for temperature field digital twin closed-loop assimilation and process feedback optimization in the composite material curing process according to claim 7, characterized in that, The correction process in step S53 specifically includes the following steps: S531, based on the reliability of the measurement points The observation noise matrix is ​​dynamically adjusted, and its expression is as follows: in, Indicates the first Each measuring point at time... Confidence correction for observation noise; Indicates the first The basic observation noise corresponding to each measuring point; Indicates the first Each measuring point at time... Credibility weight; This prevents small positive numbers with a denominator of zero; S532, calculate the confidence-weighted assimilation gain, which is expressed as: in, This represents the assimilation gain after considering the reliability of the measurement points; This represents the prior state error covariance matrix at the current moment; Represents the observation matrix; Represents the observation matrix transpose; This represents the observation noise matrix after confidence correction; This represents finding the inverse of a matrix. S533, the prior mode coefficients are corrected based on the aforementioned confidence-weighted assimilation gain, expressed as follows: in, Represents the posterior modal coefficients after assimilation correction at the current time. , This represents the observation residual at the current moment; This indicates the real-time temperature of the thermocouple at the current moment. This indicates the thermocouple's predicted temperature at the current moment.

9. The method for temperature field digital twin closed-loop assimilation and process feedback optimization in the composite material curing process according to claim 4, characterized in that, The calculation of the reliability of the measurement point in step S53 also includes: The reliability of the measurement points is calculated by comprehensively considering multiple consistency evaluation items. Its expression is: in, Show the first Temperature range consistency score for each measuring point; Indicates the first Time continuity score for each measurement point; Indicates the first Consistency score of heating rate at each measuring point; Indicates the first Spatial consistency score of each measurement point; Indicates the first Consistency score of process stages at each measuring point; Indicates the first Consistency score of model residuals at each measurement point; This represents the comprehensive evaluation function.

10. The method for temperature field digital twin closed-loop assimilation and process feedback optimization in the composite material curing process according to claim 5, characterized in that, The process optimization instruction at the current moment in step S64 is specifically expressed as follows: in, This indicates the process optimization instruction at the current moment; This represents the posterior global temperature field at the current moment; This represents the set of temperature features extracted from the posterior full-field temperature field. This indicates the risk assessment result at the current moment; Indicates process window constraints; Indicates equipment operating constraints; This represents the process optimization decision function.

Images