A method and system for predicting the aging performance and lifespan of picture frame materials based on transfer learning

By constructing a transfer learning-based method for predicting the aging performance and lifespan of eyeglass frame materials, and using aging experiment data to establish aging memory features and a digital damage twin model, the accuracy and adaptability issues of predicting the aging performance of eyeglass frame materials are solved, and efficient lifespan prediction is achieved.

CN123120148APending Publication Date: 2026-11-17CHONGQING MUSHEKE OPTICAL TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611367023.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-04
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies are insufficient to fully characterize the multidimensional damage evolution of frame materials, and traditional models lose accuracy when data is insufficient, making it difficult to accurately predict their aging performance and lifespan.

Method used

A method for predicting the aging performance and lifespan of frame materials based on transfer learning is constructed. By collecting aging experimental data, a source domain aging database is established, features are extracted and aging memory features are constructed, and the differences in feature distribution are corrected using a transfer learning model. A digital damage twin model is established, and the damage evolution state is dynamically updated in combination with real-time service environment parameters to predict the failure time.

Benefits of technology

It improves the accuracy and reliability of predicting the aging performance of eyeglass frame materials, reduces the dependence on experimental data of the complete life cycle of the target eyeglass frame material, and enhances the adaptability to different material types and environmental conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120148A_ABST
    Figure CN123120148A_ABST
Patent Text Reader

Abstract

This invention proposes a method and system for predicting the aging performance and lifespan of eyeglass frame materials based on transfer learning. The method includes: constructing a source domain aging database; extracting damage accumulation features, performance degradation rate features, and aging stage transition features for each material at each aging stage to construct source domain aging memory features; establishing an aging memory encoding model to map the aging detection data of the target eyeglass frame material to the target aging memory features; constructing a transfer learning model to correct the differences in feature distribution between the two to obtain an aging damage evolution model; constructing a digital damage twin model, inputting real-time service environment parameters to obtain the damage evolution state of the target eyeglass frame material at different service times; determining the failure time by combining preset failure criteria, and outputting the aging performance and lifespan prediction results. This method dynamically updates damage accumulation, performance degradation, and aging stage states based on the digital damage twin model to predict the failure time, improving accuracy, reliability, and dynamic adaptability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of material aging performance prediction technology, and in particular to a method and system for predicting the aging performance and lifespan of picture frame materials based on transfer learning. Background Technology

[0002] With the development of smart wearables, vision correction, and high-end eyewear products, eyeglass frame materials are subjected to the combined effects of multiple factors such as temperature, humidity, ultraviolet radiation, chemical media, and mechanical loads over a long period of time. This leads to continuous degradation of their internal structure, surface morphology, chemical composition, and mechanical properties. Eyeglass frame materials must balance lightweight, high strength, corrosion resistance, and wearing comfort. The aging process is characterized by multi-factor driven, damage accumulation, and gradual performance degradation. Current life assessments largely rely on accelerated aging tests, evaluating durability by measuring indicators such as color changes, surface defects, and mechanical properties. However, these methods are often based on single indicators or limited periods, making it difficult to comprehensively characterize the multidimensional damage evolution. Furthermore, the significant differences in materials, service environments, and aging mechanisms make it difficult to transfer experimental results to real-world scenarios, resulting in insufficient predictive accuracy and adaptability. In recent years, the development of artificial intelligence, transfer learning, and digital twin technologies has led to the increasing application of data-driven methods in material degradation analysis. These methods typically involve building machine learning models to learn aging data and predict remaining lifespan. However, data acquisition for frame material aging is time-consuming, and data distribution varies significantly among different materials. Traditional models rely on a large number of target samples, and prediction accuracy decreases when data is insufficient. Furthermore, many models only focus on the mapping between aging state and lifespan results, lacking characterization of the damage accumulation process and stage evolution patterns, making accurate prediction difficult. Therefore, there is an urgent need for a method that utilizes existing aging experience and combines it with the actual service state of the target material for knowledge transfer and dynamic prediction, in order to improve the accuracy, reliability, and environmental adaptability of lifespan prediction. Summary of the Invention

[0003] This invention aims to at least solve the technical problems existing in the prior art, and in particular, it innovatively proposes a method and system for predicting the aging performance and lifespan of picture frame materials based on transfer learning.

[0004] To achieve the above-mentioned objectives of this invention, this invention provides a method for predicting the aging performance and lifespan of picture frame materials based on transfer learning, the method comprising: S1. Collect aging test data of different types of frame materials under different aging environmental conditions, and construct a source domain aging database based on the aging test data; S2. Based on aging experiment data in the source domain aging database, feature extraction is performed on the performance degradation process of different types of frame materials at different aging stages, and source domain aging memory features characterizing the historical damage evolution of frame materials are constructed according to the performance degradation process. S3. Based on the source domain aging memory features, establish an aging memory coding model, and use the aging memory coding model to perform feature mapping on the aging test data corresponding to the target frame material, so as to obtain the target aging memory features corresponding to the target frame material. S4. Based on the source domain aging memory features and the target aging memory features, a transfer learning model is constructed, and the feature distribution difference between the source domain aging memory features and the target aging memory features is corrected through the transfer learning model to obtain an aging damage evolution model suitable for the target frame material. S5. Based on the aging damage evolution model, construct a digital damage twin model corresponding to the target frame material, and input the real-time service environment parameters corresponding to the target frame material into the digital damage twin model to obtain the damage evolution state of the target frame material under different service times. S6. Based on the damage evolution state and combined with the preset failure criteria, determine the failure time of the target frame material and output the aging performance life prediction result of the target frame material.

[0005] In another aspect, the present invention also provides a system for predicting the aging performance and lifespan of eyeglass frame materials based on transfer learning, the system including a processor and a memory for storing processor-executable instructions; wherein the processor is configured to implement the method for predicting the aging performance and lifespan of eyeglass frame materials based on transfer learning when executing the executable instructions.

[0006] The beneficial effects of this invention are as follows: By constructing a source domain aging database containing aging experimental data of different types of frame materials, this invention can fully utilize the historical degradation information of existing frame materials under different aging environmental conditions. Furthermore, by extracting damage accumulation features, performance degradation rate features, and aging stage transition features, source domain aging memory features are constructed, thereby achieving effective characterization of the historical damage evolution law of frame materials. Further, by mapping the aging detection data of the target frame material to the aging memory feature space through an aging memory coding model, and using a transfer learning model to correct the feature distribution differences between the source domain aging memory features and the target aging memory features, the aging damage evolution law already obtained in the source domain material can be effectively represented. This method can be transferred to target frame materials with relatively limited aging data, reducing reliance on experimental data covering the entire life cycle of the target frame material and improving the adaptability of life prediction under different material types and service environments. Simultaneously, a digital damage twin model is constructed based on the aging damage evolution model obtained through the transfer. This model dynamically updates the damage accumulation state, performance degradation state, and aging stage state of the target frame material according to real-time service environment parameters. Combined with preset failure criteria, it predicts the time when the target frame material reaches failure, thereby achieving continuous life prediction from historical damage memory and current damage state to future damage evolution, improving the accuracy, reliability, and dynamic adaptability of life prediction for the aging performance of frame materials.

[0007] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0008] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of a method for predicting the aging performance and lifespan of picture frame materials based on transfer learning, according to the present invention. Detailed Implementation

[0009] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0010] Example 1 like Figure 1 As shown, a method for predicting the aging performance and lifespan of picture frame materials based on transfer learning is described, the method comprising: S1. Collect aging test data of different types of frame materials under different aging environmental conditions, and construct a source domain aging database based on the aging test data. The aging test data includes environmental load parameters, surface morphology change parameters, chemical structure change parameters and mechanical property degradation parameters corresponding to the frame materials. The aging test data collected in this step includes publicly available experimental data of mature frame materials that have completed full-life-cycle aging tests, as well as accelerated aging test data of materials under different conditions accumulated in this field, in order to integrate existing research resources and avoid redundant experiments. Environmental load parameters characterize the intensity and duration of external environmental effects, surface morphology change parameters reflect the evolution of external damage, chemical structure change parameters characterize changes in internal structures such as molecular chains and crystal phases, and mechanical property degradation parameters characterize the decline in service performance such as load-bearing capacity and bending resistance. These four types of data characterize the aging process from four dimensions: external effects, external damage, internal changes, and performance degradation. When constructing the database, the raw experimental data is standardized and preprocessed, outliers are removed, parameter units and labeling formats are unified, and index relationships are established to facilitate subsequent data retrieval for model training.

[0011] Environmental load parameters include ambient temperature, humidity, ultraviolet radiation intensity, and concentration of corrosive chemical media under different operating conditions, and the duration of each operating condition is recorded. Surface morphology change parameters include surface roughness, microcrack size and number, yellowness index, gloss, and surface weight loss per unit area at different stages. Chemical structure change parameters include the proportion of molecular chain breakage, degree of crosslinking, type and content of characteristic functional groups, crystallinity, and relative content of oxidation products. Mechanical property degradation parameters include tensile strength, flexural strength, impact strength retention rate, hardness, and elastic modulus at different stages, reflecting the decay law of mechanical properties with the aging process.

[0012] S2. Based on aging experiment data in the source domain aging database, feature extraction is performed on the performance degradation process of different types of frame materials at different aging stages, and source domain aging memory features characterizing the historical damage evolution of frame materials are constructed according to the performance degradation process. S3. Based on the source domain aging memory characteristics, establish an aging memory coding model, and use the aging memory coding model to perform feature mapping on the aging detection data corresponding to the target frame material to obtain the target aging memory characteristics corresponding to the target frame material. The target aging memory characteristics are used to characterize the cumulative damage state of the target frame material during service. S4. Based on the source domain aging memory features and the target aging memory features, a transfer learning model is constructed, and the feature distribution difference between the source domain aging memory features and the target aging memory features is corrected through the transfer learning model to obtain an aging damage evolution model suitable for the target frame material. S5. Based on the aging damage evolution model, construct a digital damage twin model corresponding to the target frame material, and input the real-time service environment parameters corresponding to the target frame material into the digital damage twin model to obtain the damage evolution state of the target frame material under different service times. S6. Based on the damage evolution state and combined with the preset failure criteria, determine the failure time of the target frame material and output the aging performance life prediction result of the target frame material.

[0013] Constructing source domain aging memory features includes: S201. Divide the aging test data in the source domain aging database into stages according to the preset aging time interval to obtain the staged aging data of the frame material at different aging stages. The staged aging data includes environmental load parameters, surface morphology change parameters, chemical structure change parameters and mechanical property degradation parameters corresponding to different aging stages. This step divides the full-cycle aging test data into stages by setting a preset time interval. The core purpose is to decompose the continuously changing aging process into multiple discrete and observable time units, facilitating the accurate extraction of material aging characteristics at different time points and adapting to the input data format requirements of subsequent time-series models. The preset aging time interval can be flexibly adjusted according to the type of frame material and the acceleration rate of the aging test: for conventional thermoplastic resin frame materials, the preset interval for natural aging tests can be set to 30 days, and the preset interval for accelerated aging tests can be adjusted to 72 hours, balancing data acquisition costs and feature extraction accuracy; for metal frame materials with slower aging rates, the division interval can be appropriately extended to reduce redundant data while ensuring that the performance degradation trend can be captured. After the stage division is completed, each aging stage will integrate aging data from all dimensions within that time period, ensuring that the data for each stage can fully reflect the aging state of the frame material after being subjected to environmental effects within that time period, while clearly marking the cumulative aging time corresponding to each stage.

[0014] S202. Based on the phased aging data, feature extraction is performed on the performance degradation process of the frame material at different aging stages to obtain the environmental response characteristics, structural damage characteristics, chemical degradation characteristics, and mechanical property decay characteristics of the frame material. This step extracts features corresponding to different characterization dimensions from the multi-dimensional data of each aging stage after stage division, achieving dimensionality reduction and information purification of the original aging data. The four features correspond to different characterization levels of the aging process: Among them, the environmental response feature is a feature quantity obtained by integrating various environmental load parameters and corresponding action durations within the aging stage, used to characterize the cumulative effect of external environmental action on the frame material within this stage, clearly depicting the intensity of external excitation driving material performance degradation at different aging stages; the structural damage feature is obtained by normalizing and integrating surface morphology data such as surface roughness, surface microcrack parameters, yellowness index, gloss, and surface weight loss rate within this stage, used to quantitatively characterize the frame material under this aging stage. The degree of structural damage already present at the macroscopic level directly reflects the aging process on the material's surface. Chemical degradation characteristics are derived from microstructural data such as the proportion of molecular chain breakage, degree of cross-linking, content of characteristic functional groups, crystallinity, and relative content of oxidation products at this stage. This characterizes the extent of aging chemical reactions in the inner frame material at the molecular level, reflecting the changes in the material's internal structure caused by aging. Mechanical property degradation characteristics are derived from mechanical parameters such as tensile strength retention rate, flexural strength retention rate, impact strength retention rate, hardness, and elastic modulus at this stage. This directly quantifies the degree of degradation in the core service performance of the frame material at this stage, providing the most direct indicator of the material's current aging state. These four dimensions work together to comprehensively depict the aging state of a single aging stage from four perspectives: external excitation, surface damage, internal changes, and performance output. This approach preserves the core information of the original experimental data while achieving data dimensionality reduction.

[0015] S203. Based on the environmental response characteristics, structural damage characteristics, chemical degradation characteristics and mechanical property decay characteristics corresponding to different aging stages, establish the characteristic correlation relationship between the frame material at different aging stages, and obtain the continuous degradation characteristic sequence characterizing the evolution process of the frame material from the initial state to the failure state according to the characteristic correlation relationship. First, the feature vectors of all aging stages are sorted from smallest to largest based on the cumulative aging time of each stage, ensuring that the time trend of the feature sequence aligns with the actual aging process of the material. The changes in the four dimensions of features between adjacent stages are calculated sequentially to obtain the feature change rates between adjacent stages: changes in environmental response reflect the driving effect of changes in external load intensity on the aging process; changes in structural damage, chemical degradation, and mechanical property decay reflect the development rates of aging damage, internal structural changes, and performance degradation, respectively. The feature change rates are then added to the corresponding time-series nodes, ensuring that each node simultaneously contains both static state information and dynamic changes in the aging process. After integrating the features of a single set of experimental data, the annotation information of the corresponding frame material type and aging environmental conditions is associated with the feature sequence. This process is then performed sequentially on multiple sets of experimental data from different materials and environments in the source domain database, ultimately yielding multiple sets of continuous degradation feature sequences. Each set fully covers the entire aging process from the initial unaged, zero-damage state to the experimental endpoint failure state, preserving the dynamic laws governing the aging damage evolution of the frame material under different conditions.

[0016] This temporal feature construction method fully preserves the intrinsic correlation between aging features at different stages, enabling the full-cycle aging patterns contained in the source domain data to be fully learned and utilized by subsequent models, ensuring that the extracted source domain aging memory features accurately represent the historical damage evolution patterns of the frame material.

[0017] S204. Based on the continuous degradation feature sequence, extract the damage accumulation features, performance degradation rate features, and aging stage transition features of the frame material during the aging process, and encode the damage accumulation features, performance degradation rate features, and aging stage transition features to obtain source domain aging memory features that characterize the historical damage evolution law of the frame material.

[0018] The three types of features extracted in this step correspond to the regular information of three core levels in the aging process of the frame material. The specific extraction and encoding process is as follows: First, the extraction of damage accumulation features, which is used to characterize the overall trend of internal damage of the frame material increasing with aging time under continuous external environmental loads. During the extraction process, the cumulative values ​​of structural damage features and chemical degradation features at each time node in the continuous degradation feature sequence are statistically analyzed. The relationship curve of damage accumulation versus cumulative aging time is fitted, and then the shape parameters, cumulative growth coefficient, and other characterization information are extracted from the curve to obtain the damage accumulation features. Second, the extraction of performance degradation rate features, which is used to characterize the response law of the decay rate of the core service performance of the frame material to different environmental loads. During the extraction process, for the environmental response features and mechanical performance decay features of each adjacent time node in the continuous degradation feature sequence, the change in mechanical performance decay corresponding to a unit change in environmental load is calculated. The correspondence between performance degradation rate and environmental load intensity at different aging stages is obtained, and then the relationship is parameterized to obtain the performance degradation rate features. Thirdly, the aging stage transition characteristics are extracted. The aging process of the frame material is divided into three stages: aging induction period, damage stabilization period, and accelerated aging failure period. This feature is used to characterize the damage threshold and performance critical value corresponding to the transition nodes of different stages. During the extraction process, abrupt changes in the performance degradation rate in the continuous degradation feature sequence are detected to identify two rate abrupt change time nodes, which correspond to the transition nodes from the aging induction period to the stabilization period and from the stabilization period to the accelerated failure period, respectively. The damage accumulation and mechanical property retention rate parameters corresponding to the two nodes are extracted to obtain the aging stage transition characteristics.

[0019] After extracting the three types of features, the three types of features are normalized and encoded according to a preset unified dimension, and integrated into a fixed-dimensional feature vector, which is the source domain aging memory feature corresponding to the current set of experiments. The above operation is performed on all sets of experiments in the source domain aging database in sequence, and finally all source domain aging memory features covering different material types and different aging environmental conditions are obtained, which completely preserve the full-cycle aging evolution law of the frame material, remove redundant information in the original data, and retain only the core transferable law information.

[0020] Obtaining the target aging memory features corresponding to the target frame material includes: S301, based on the source domain aging memory features, organizing the damage accumulation features, performance degradation rate features, and aging stage transition features corresponding to different types of frame materials, and constructing an aging memory feature space to characterize the historical damage evolution law of different types of frame materials based on the feature organization results. This step first standardizes and preprocesses all source domain aging memory features. Because the parameter test range and data statistical caliber of different source domain aging experiments are different, all feature values ​​of the three types of features need to be uniformly projected to the preset [0,1] value range to eliminate the difference in dimensions and avoid the difference in numerical magnitude from interfering with the subsequent spatial construction.

[0021] Subsequently, three interconnected sub-feature spaces are constructed according to feature types, corresponding to damage accumulation features, performance degradation rate features, and aging stage transition features, respectively. The three types of feature components of all source domain samples are projected onto their respective sub-spaces. After projection, each feature vector is bound with labels corresponding to the frame material type and aging environment parameters, establishing a mapping between features and material properties and environmental conditions. This facilitates subsequent matching of source domain patterns with similar material properties to the target frame. This domain-specific storage structure avoids interference between different types of pattern information and supports subsequent distribution correction and transfer learning for different feature components, improving the targeting and accuracy of the model processing. Following this, the projected and labeled feature space undergoes manifold alignment to correct the distribution of all feature vectors and correct systemic biases caused by differences in equipment precision and acquisition processes in different experiments. This allows source domain features of the same material type and similar aging environment to cluster in close neighborhoods within the feature space, while features of different aging patterns are naturally separated according to their differences. This improves the separability and regularity of the feature space while preserving the original core pattern information. The final aging memory feature space is the core knowledge carrier of the aging memory coding model. It integrates all the aging law information of source domain frame materials accumulated in existing research. It can structurally characterize the aging laws of different types of frame materials under different service environments, and provide a unified knowledge framework for the feature mapping of subsequent target frame material aging detection data. This ensures that the target aging memory features extracted later can accurately reflect the actual cumulative damage state of the target frame material.

[0022] S302. Establish an aging memory coding model based on the aging memory feature space, and train the aging memory coding model using source domain aging memory features so that the aging memory coding model establishes a mapping relationship between the performance degradation process of the frame material and the aging memory feature space. Training the aging memory coding model includes: S3021, according to the aging stage sequence corresponding to the frame material, arranging the damage accumulation feature, performance degradation rate feature and aging stage transition feature in the source domain aging memory features in a temporal order, and combining the damage accumulation feature, performance degradation rate feature and aging stage transition feature corresponding to the same aging stage to obtain aging memory feature vectors corresponding to different aging stages. The processing steps are as follows: First, for a single source domain experimental sample, the extracted aging stage transition features are retrieved to obtain two pre-identified stage transition node parameters: the first transition node from the aging process induction phase to the stable damage development phase, and the second transition node from the stable damage development phase to the accelerated aging failure phase. Based on these two nodes, the entire aging process of the sample is divided into three independent aging stage intervals, clarifying the temporal order of each stage. Subsequently, for each divided aging stage, corresponding feature components are extracted from the source domain aging memory features: the change parameters of damage accumulation features within the current stage interval, the response parameters of performance degradation rate features to environmental load within the current stage interval, and the aging stage transition threshold parameters corresponding to the boundaries before and after the current stage. Then, the three types of feature components are concatenated and integrated according to a preset dimension order to form a fixed-dimensional one-dimensional feature vector, which is the aging memory feature vector corresponding to the current aging stage.

[0023] After combining features from all stages of a single sample, the aging memory feature vectors corresponding to the three aging stages are sequentially arranged according to the aging process order: aging induction phase first, damage stabilization development phase in the middle, and accelerated aging failure phase last. This yields a temporal feature structure for the source domain sample that meets the model input requirements. The above operation is then performed sequentially on all training samples in the source domain aging database, ultimately resulting in a standardized set of temporal aging memory feature vectors covering all source domain samples. This provides standardized training input for subsequent aging memory encoding model training, ensuring the model fully learns the feature mapping patterns of different types of frame materials under different aging environments.

[0024] S3022. Based on the aging memory feature vectors corresponding to different aging stages and the temporal relationship between each aging memory feature vector, construct the aging memory feature sequence corresponding to the frame material, and configure the corresponding aging stage position information for each aging memory feature vector in the aging memory feature sequence to obtain the aging memory input sequence containing the temporal relationship of the performance degradation of the frame material. The processing flow is as follows: First, according to the aging process sequence, the aging memory feature vectors corresponding to the three aging stages of a single sample are concatenated to obtain the aging memory feature sequence corresponding to that sample. Since the aging memory feature vectors do not contain temporal position attributes, directly inputting them into the model will not allow the model to distinguish the aging stage sequence of different features or learn the evolutionary dependencies between different stages. Therefore, it is necessary to configure corresponding aging stage position information for each feature vector: assign position identifiers according to the aging process sequence, with identifier 1 corresponding to the aging induction period, identifier 2 corresponding to the damage stabilization development period, and identifier 3 corresponding to the accelerated aging failure period. Then, sine and cosine position encoding is used to convert the discrete position identifiers into real-valued position encoding vectors of the same dimension as the original feature vectors, which are then fused with the original feature vectors, so that the final feature carries both aging pattern information and temporal position information. After processing, the aging memory input sequence corresponding to the sample, containing the temporal relationship of the frame material performance degradation, is obtained. This structure helps the aging memory encoding model perceive the aging stage sequence, learn the inherent law of the cumulative damage of the previous stage affecting the degradation process of the next stage, avoid stage mismatch and temporal reversal, and improve model training efficiency and feature mapping accuracy. By performing the above operations sequentially on all training samples in the source domain database, a complete set of aging memory input sequences with uniform format and standard dimensions can be obtained.

[0025] S3023. Input the aging memory input sequence into the aging memory coding model, extract the local degradation correlation features between adjacent aging stages and the long-term damage evolution correlation features between different aging stages through the aging memory coding model, and fuse the local degradation correlation features and the long-term damage evolution correlation features to obtain the aging memory coding representation corresponding to the frame material. This step employs an aging memory coding model based on the TCN-Transformer dual-branch temporal feature coding architecture, which is divided into three core components: local degradation correlation feature extraction, long-term damage evolution correlation feature extraction, and multi-feature fusion. First, it extracts local degradation correlation features between adjacent stages to capture the direct dependencies in damage evolution between adjacent aging stages. The model performs sliding window convolutions on the input aging memory sequence along the temporal dimension using multi-layer one-dimensional convolutional kernels. Each convolutional window covers the feature vectors of two adjacent aging stages, extracting the coupling relationship between feature changes in adjacent stages to obtain multi-channel local degradation correlation features. This accurately depicts the direct impact of damage accumulation in the previous stage on the performance degradation rate of the next stage, preserving local details of the evolutionary patterns of adjacent stages. Second, it extracts long-term damage evolution correlation features across stages to capture the indirect dependencies between different aging stages throughout the entire lifecycle, compensating for the inadequacy of local features in covering the entire lifecycle evolution logic. The model calculates the similarity of feature vectors at any two temporal positions in the aging memory input sequence through a multi-head self-attention mechanism, generates attention weights, and obtains long-term damage evolution correlation features containing full-stage correlation information through weighted aggregation. This captures the potential correlation between initial aging and final failure processes, realizing the overall modeling of the full-cycle aging evolution logic of the frame material. Thirdly, it integrates local and long-term correlation features. First, the two types of features are transformed in dimension through independent fully connected layers to make their dimensions consistent with the input feature dimensions. Then, the two types of features after dimension alignment are summed and fused element-wise, and connected to the layer normalization module and residual connection structure. This retains the single-stage regularity information of the original input while integrating newly added correlation features, avoiding gradient vanishing or distribution shift. The final output is the aging memory encoding representation of the corresponding source domain sample. The aging memory encoding representation obtained through the above processing integrates single-stage aging states, local evolutionary dependencies between adjacent stages, and long-term evolution correlation information throughout the entire cycle, completely preserving the core regularity of aging damage evolution of the frame material in the source domain sample.

[0026] This embodiment of the TCN-Transformer dual-branch temporal feature encoding architecture consists of an input layer, a TCN local feature encoding branch, a Transformer long-range correlation encoding branch, and a fusion output layer. The configuration of each part is as follows: Input Layer: Receives the preprocessed aging memory input sequence with location information, converts it into a tensor format that the model can process, and outputs a sequence tensor with uniform dimensions as the shared input for the two encoding branches. The length of the input tensor sequence is fixed at 3, corresponding to the three aging stages, and the input feature dimension is preset to 64, consistent with the source domain aging memory feature encoding dimension. TCN Local Feature Encoding Branch: Used to extract local degradation correlation features between adjacent aging stages, it consists of two stacked causal dilated convolutional modules connected in series. Each causal dilated convolutional module contains two layers of one-dimensional dilated causal convolutions with progressively increasing dilation rates. Each convolutional layer is followed by a ReLU activation layer and a Dropout regularization layer. The causal convolution constraint ensures that temporal modeling uses only features from the current and previous stages, conforming to the unidirectional irreversible nature of the aging process and preventing the leakage of future features. The progressively increasing dilation rate expands the receptive field of the convolution without increasing the number of parameters or computational cost, allowing the top-level convolution to cover all three aging stages, balancing the need for detail extraction from adjacent stages and modeling local dependencies across multiple stages. The multi-channel local features output from all convolutional modules are reduced in dimensionality by global average pooling to obtain dimension-matched local degradation correlation features, which are then output to the fusion output layer.

[0027] The Transformer long-range correlation coding branch is used to extract long-term damage evolution correlation features between different aging stages across the entire lifecycle. It consists of two stacked standard Transformer coding blocks concatenated. Each Transformer coding block sequentially includes a multi-head self-attention sub-layer, a layer normalization layer, a residual connection layer, a feedforward neural network sub-layer, and another set of layer normalization layers and residual connection layers. The multi-head self-attention sub-layer can calculate the dependency weights between any two aging stage features in the input sequence, capturing the potential correlation between interval stages and compensating for the TCN branch's limitation of only being good at extracting local correlations. The feedforward neural network sub-layer performs a non-linear transformation on the attention output features to extract the regular information in the features. After all coding blocks are processed, the long-term damage evolution correlation features that meet the dimensionality requirements are output and sent to the fusion output layer. The local and long-term correlation features output from the two branches are integrated: First, the dimensions of the two types of features are transformed by two independent fully connected layers to unify the dimensions of the two types of features to a preset 64; then, the two types of features after dimension alignment are fused element-wise, and the normalization module and residual connection structure of the access layer are connected to complete the complementary integration of the two types of correlation features; finally, the fusion result that meets the dimension requirements is output, which is the aging memory encoding representation of the frame material output by this encoding architecture.

[0028] S3024. Map the aging memory encoding representation to the aging memory feature space, and adjust the model parameters of the aging memory encoding model according to the reconstruction deviation between the aging memory encoding representation and the corresponding source domain aging memory features and the evolution continuity deviation between the aging memory encoding representations corresponding to adjacent aging stages. The processing steps are as follows: First, after generating the aging memory encoding representation, the aging memory encoding representations of all source domain samples are connected to a pre-constructed fully connected decoding layer. Through nonlinear transformation, they are projected back into the original aging memory feature space to obtain the reconstructed aging memory features corresponding to each source domain sample. Then, the reconstruction bias is calculated: the reconstructed aging memory features and the original source domain aging memory features of the corresponding sample are substituted into the mean square error formula to obtain the reconstruction bias. This constrains the model to fully retain the core aging law information of the original source domain features, avoiding the loss of key information. Second, the evolution continuity bias is calculated. The aging damage evolution of the frame material is a continuous and gradual process, with irregular jumps and abrupt changes in damage states between adjacent aging stages. Therefore, constraints are added to the evolution continuity of the encoding representations of adjacent stages: for a single source domain sample, the aging memory encoding representations of three aging stages arranged in chronological order are extracted. The Euclidean distance between the encoding vectors of two adjacent stages is calculated sequentially, and the average of the two distances is taken to obtain the evolution continuity bias corresponding to the sample. This constrains the aging evolution process learned by the model to conform to actual physical laws, avoiding discrepancies between damage features of adjacent stages that do not conform to actual aging logic. Finally, after calculating the two types of biases, the total loss is integrated and the parameters are updated: the reconstruction bias and evolution continuity bias are weighted and summed according to preset weights to obtain the total training loss of the model. The weights of the two types of biases are preset hyperparameters, and the influence ratio of the two types of constraints can be adjusted. Then, a gradient optimization algorithm is used to backpropagate along the gradient direction of the total loss, iteratively updating all trainable parameters of the model; the process of forward calculation of loss and backward parameter update is repeated until the total loss drops to a preset convergence threshold or a preset number of iterations is completed, at which point training stops, resulting in an aging memory encoding model with fixed parameters and complete training. This model has fully learned the aging evolution laws of different frame materials and different environments in the source domain data, and can accurately establish the mapping relationship between the performance degradation process of frame materials and the aging memory feature space, which can be used for feature mapping processing of subsequent target frame material detection data.

[0029] S3025. Repeatedly perform the acquisition of aging memory encoding representation, mapping of aging memory encoding representation to aging memory feature space, and adjustment of model parameters until the reconstruction deviation and evolution continuity deviation meet the preset convergence conditions to obtain the trained aging memory encoding model, and establish the mapping relationship between the performance degradation process of the frame material and the aging memory feature space through the trained aging memory encoding model.

[0030] This step sets up dual convergence criteria, judging the model training status from two dimensions: the numerical value of the deviation and the trend of loss change. This avoids underfitting or overfitting problems caused by a single criterion. The judgment logic is as follows: First, independent convergence thresholds are set for reconstruction deviation and evolutionary continuity deviation. Only when the reconstruction deviation of the current iteration is less than the preset threshold of 0.02, and the evolutionary continuity deviation is less than the preset threshold of 0.03, is the basic convergence requirement at the deviation numerical level met. If either deviation fails to meet the standard, the next iteration of training continues. Second, after meeting the basic convergence requirement, the trend of total loss change is further judged: the average decrease in total loss over the most recent 15 iterations is statistically analyzed. If this decrease is less than the preset loss change threshold of 1×10⁻⁴, the model is considered to have fully met the convergence conditions, and iterative training stops. If both convergence conditions are not fully met, the updated model parameters of the current iteration are used as the initial parameters, and the process is returned to S3021 to re-execute the complete process and enter the next training iteration, repeating until both conditions are met. After stopping the iteration, all trainable parameters of the aging memory encoding model are fixed, resulting in the trained model. The model has fully learned the underlying laws of the full-cycle aging evolution of different frame materials and under different aging environments in the source domain database. The established mapping relationship can convert the aging detection data of any target frame material into target aging memory features that are consistent with the source domain measurement standard, thus eliminating the distribution differences between the target domain and the source domain data.

[0031] S303. Obtain the aging test data corresponding to the target frame material, and extract features from the environmental load parameters, surface morphology change parameters, chemical structure change parameters and mechanical property degradation parameters in the aging test data to obtain the target environmental response characteristics, target structural damage characteristics, target chemical degradation characteristics and target mechanical property decay characteristics corresponding to the target frame material. The aging test data for the target lens frame material comes from the target lens frame product whose lifespan is to be predicted. The data acquisition methods are divided into two categories based on the actual application scenario: Laboratory accelerated life prediction scenario, where artificial accelerated aging tests are conducted on the target lens frame material, and multiple rounds of sampling and testing are completed at preset time intervals; and in-service lens frame life assessment scenario, where data sampling is completed by periodically conducting non-destructive testing on the target lens frame in use. All sampling nodes are pre-divided into three aging stages according to the material aging process. At least one multi-parameter test is completed in each stage, obtaining four types of raw test data: environmental load parameters, surface morphology change parameters, chemical structure change parameters, and mechanical property degradation parameters for the corresponding stage. Subsequently, feature extraction is performed on each of the four types of raw parameters. The target environmental response characteristics are extracted based on the environmental load parameters: the environmental load parameters include four types of raw time-series data of temperature, relative humidity, ultraviolet radiation intensity, and external mechanical load stress during the aging process of the target material. For each parameter in each aging stage, four types of statistics are calculated: mean, fluctuation variance, maximum extreme value, and cumulative duration. All statistics are concatenated in sequence to obtain the target environmental response characteristics corresponding to that aging stage, which characterizes the driving effect of environmental factors on the aging process of the target frame material.

[0032] To extract target structural damage features based on surface morphology change parameters: The surface morphology change parameters are microscopic morphology images of the frame material surface collected at different aging stages, reflecting the development of surface damage. This step uses the same convolutional neural network as the source domain feature extraction to downsample and refine the morphology images. After removing the final classification output layer of the network, the resulting 512-dimensional feature vector is the target structural damage feature corresponding to the aging stage, ensuring that the feature extraction rules of the target domain and the source domain are consistent.

[0033] Target chemical degradation characteristics are extracted based on parameters of chemical structure change: These parameters include three types of raw data: Fourier transform infrared spectral absorbance curves, X-ray photoelectron spectroscopy functional group content data, and thermogravimetric analysis thermogravimetric curves. From the infrared spectral curves, four types of parameters are extracted: peak position, peak height, peak area, and half-peak width. Combined with the functional group ratio, 5% thermal decomposition temperature, and residual mass ratio, all parameters are sequentially spliced ​​to obtain the target chemical degradation characteristics corresponding to the aging stage, thus characterizing the degree of aging at the chemical level.

[0034] To extract target mechanical property degradation features from mechanical property degradation parameters, four types of performance indicators of the frame material measured at different aging stages are used: tensile strength, impact strength, Shore hardness, and flexural modulus. Using the indicators corresponding to the unaged initial state as a benchmark, the performance retention rate and degradation rate relative to the previous stage are calculated for each indicator at the current stage. After concatenating all statistics, the target mechanical property degradation features corresponding to that aging stage are obtained, reflecting the degree to which the load-bearing capacity of the target frame material decreases with the aging process. After the above extraction, four types of target features can be obtained for each aging stage. The entire feature extraction process is completely consistent with the feature extraction rules of the source domain samples, avoiding the impact of inconsistent feature measurement standards on the accuracy of subsequent migration prediction.

[0035] S304. Input the target environmental response characteristics, target structural damage characteristics, target chemical degradation characteristics, and target mechanical property decay characteristics into the aging memory coding model, and map the performance degradation information corresponding to the target frame material to the aging memory feature space based on the mapping relationship to obtain the initial target aging memory characteristics corresponding to the target frame material. First, this step involves constructing the target aging memory feature sequence. For a single test sample of the target frame material, after obtaining the four types of target features for each of the three aging stages according to the aforementioned feature extraction process, the four types of features within a single aging stage are first concatenated: the target environmental response feature, target structural damage feature, target chemical degradation feature, and target mechanical property decay feature are concatenated in a preset order to obtain the target aging memory feature vector corresponding to that aging stage. Subsequently, the target aging memory feature vectors of the three aging stages are concatenated sequentially according to the aging process to form the serialized feature structure of the target sample, resulting in the initial target aging memory feature sequence. Next, positional information is configured for the target feature sequence, and feature fusion is completed. Consistent with the source domain feature processing flow, each feature vector in the sequence is assigned a corresponding position identifier according to the aging process sequence: position identifier 1 for the aging induction period, position identifier 2 for the damage stable development period, and position identifier 3 for the accelerated aging failure period. Then, using the same sine and cosine position encoding method as in the model training phase, the discrete position identifiers are converted into real-valued position encoding vectors with the same dimension as the original feature vectors. These position encoding vectors are then fused element-wise with the original target aging memory feature vectors to obtain a target aging memory input sequence with positional information that meets the model input format requirements. Finally, the initial target aging memory features are obtained through model forward propagation. The processed target aging memory input sequence is then input into the trained aging memory encoding model. Following the pre-trained dual-branch encoding process, the model sequentially extracts local degradation correlation features of adjacent stages, extracts long-term damage evolution correlation features throughout the entire cycle, and fuses the two types of features. The final fused feature output by the model is the initial target aging memory feature corresponding to the target frame material.

[0036] The initial target aging memory features obtained through the above processing have been transformed into an aging memory feature space that is consistent with the source domain samples through the trained mapping relationship. This fully preserves the performance degradation law information of the target frame material itself, and achieves alignment with the feature measurement standard and distribution space of the source domain. This eliminates the negative impact of cross-domain distribution differences for subsequent transfer learning lifetime prediction, and can improve the accuracy of the final lifetime prediction results.

[0037] S305. Based on the feature correlation between the initial target aging memory features and the source domain aging memory features, feature calibration is performed on the initial target aging memory features to obtain target aging memory features that characterize the cumulative damage state of the target frame material during service.

[0038] This processing flow consists of three core steps: First, calculating feature association similarity. The initial target aging memory features obtained from mapping are substituted with all source domain aging memory features in the source domain feature set into the cosine similarity calculation formula to obtain the cosine similarity of each target-source domain feature pair. This value reflects the closeness of the aging states corresponding to the two types of features; the higher the similarity, the stronger the correlation of the aging evolution law. Second, selecting nearest neighbor reference source features. All source domain aging memory features are sorted from high to low similarity, and the top 5 are selected as nearest neighbor reference features for calibration. This filters out irrelevant source domain features with excessive differences, avoids interference from invalid information, and controls the complexity of subsequent calculations, improving processing efficiency. Third, weighted fusion to complete feature calibration. First, the similarity of the selected nearest neighbor reference features is normalized to obtain the calibration weight corresponding to each nearest neighbor feature. The higher the similarity, the greater the weight. Then, based on the initial target aging memory features, the normalized calibration weights are multiplied point by point with the corresponding nearest neighbor features and summed. The weighted reference features are then fused with the initial target aging memory features element by element according to a preset ratio. The fusion result is the calibrated target aging memory features.

[0039] The target aging memory features obtained through this calibration step fully preserve the unique performance degradation information of the target frame material itself. At the same time, the maturity law information of similar aging states in the source domain is introduced as a supplement, which solves the feature representation bias problem caused by insufficient sample size in the target domain and further reduces the distribution difference between target features and source domain features.

[0040] The aging damage evolution model applicable to the target frame material includes: S401, based on the source domain aging memory features and the target aging memory features, constructing the source domain aging memory feature set and the target aging memory feature set respectively, and establishing the transfer relationship between the source domain aging memory features and the target aging memory features according to the feature correlation relationship between the source domain aging memory feature set and the target aging memory feature set; Source domain aging memory feature set construction process: The aging memory features of all source domain samples that have completed encoding and calibration are organized and stored along with the corresponding labeled remaining lifespan information, material type, and aging environment parameter information for each sample. All labeled source domain features together constitute the source domain aging memory feature set. The set adopts a structured storage format for easy subsequent rapid retrieval and matching calculations. Target aging memory feature set construction process: The aging memory features of all target frame materials that have completed feature extraction, encoding mapping, and calibration are organized and stored along with the corresponding material type and actual service or test environment parameter information for each target sample. Since the target domain samples to be predicted usually do not have pre-labeled remaining lifespan information, the set only stores feature vectors and corresponding basic attribute information, thus obtaining a complete target aging memory feature set. After completing the construction of both types of sets, a two-level matching process is used to establish the transfer relationship between source domain and target features: The first level is coarse matching based on basic attributes. Based on the material type and environmental parameter information of each target sample in the target set, source domain features with consistent material types and environmental parameter deviations within a preset range are selected from the source domain set to obtain a subset of candidate matching source domain features. This can directly exclude irrelevant source domain samples with excessive differences, reduce the amount of subsequent matching calculations, and avoid irrelevant patterns interfering with subsequent migration correction from the source.

[0041] The second level is fine-grained matching of feature distribution. For the subset of candidate source domain features obtained by coarse matching, the maximum mean difference between the candidate source feature set and the target feature set is calculated. Candidate source features whose distribution difference exceeds the threshold are successively eliminated. Finally, the remaining candidate source domain features are associated with the corresponding target features according to feature similarity, thus completing the construction of the transfer relationship.

[0042] S402. Construct a transfer learning model based on the transfer relationship, and input the source domain aging memory feature set and the target aging memory feature set into the transfer learning model to obtain the source domain damage evolution representation corresponding to the source domain aging memory feature and the target damage evolution representation corresponding to the target aging memory feature, respectively. This transfer learning model uses domain adversarial transfer learning as its core framework, comprising two core modules: a shared feature extraction backbone and a domain discriminator. The core training logic involves an adversarial game between the backbone and the discriminator to align the feature distributions of the source and target domains while preserving the aging lifetime patterns of the source domain, thus eliminating prediction errors caused by cross-domain differences at the feature level. The shared feature extraction backbone consists of three sequentially connected fully connected layers. The input dimension is a preset 64-dimensional feature, matching the dimension of the calibrated aging memory features. After a nonlinear transformation by two fully connected layers, the final output is a 32-dimensional representation of damage evolution features. This backbone simultaneously receives aging memory feature inputs from both the source and target domains, ensuring consistent transformation rules for both types of features and providing a foundation for distribution alignment. The domain discriminator is a binary fully connected subnetwork. Its input is the damage evolution representation output by the backbone, and its output is the probability that an input feature belongs to either the source or target domain. Its training objective is to accurately distinguish the domain origin of the input feature. The training objective of the shared feature extraction backbone is to generate damage evolution representations that the domain discriminator cannot distinguish the domain origin. Through adversarial training, the distribution of the source and target damage evolution representations generated by the backbone is gradually forced to converge. Based on the basic adversarial loss, the model introduces an association constraint loss based on transfer matching relationships: for source-target feature pairs that establish an association correspondence in S401, the Euclidean distance between their damage evolution representations output by the backbone is calculated. The average of the Euclidean distances of all paired feature pairs is taken as the association constraint loss. This constraint ensures that source and target features with similar aging states are adjacent in the feature space after transformation, avoiding the destruction of the feature's ability to distinguish aging damage states by adversarial training alone, thus balancing distribution alignment and state representation requirements. After model construction and training are completed, the aging memory features of the source and target domains are sequentially input into the trained shared features to extract the backbone. After the same transformation process, 32-dimensional damage evolution representations of the corresponding samples are output. At this point, the two types of damage evolution representations have achieved distribution alignment, which not only fully preserves their respective aging damage state information, but also eliminates the distribution differences between the source and target domains.

[0043] S403. Based on the source domain damage evolution representation and the target damage evolution representation, calculate the feature distribution difference between the source domain aging memory features and the target aging memory features, and obtain the migration bias between the source domain aging memory features and the target aging memory features according to the feature distribution difference. This step first calculates the global feature distribution difference. All source domain damage evolution representations output by S402 are integrated into a source domain feature distribution point set, and all target damage evolution representations are integrated into a target domain feature distribution point set. Following the maximum mean difference calculation framework from the feature matching stage, the existing Gaussian kernel function is used to map the two point sets to a high-dimensional regenerative kernel Hilbert space. The average center vectors of the source domain point set and the target domain point set in the kernel space are calculated respectively. The Euclidean distance between the two center vectors is the global feature distribution difference between the source and target domains, quantifying the degree of deviation between the two types of feature distributions at an overall level.

[0044] The final transfer bias is then generated. Based on the source-target feature pairing association established by S401, for each paired feature pair, the difference between the corresponding source domain damage evolution representation and the target damage evolution representation in each feature dimension is calculated. The differences in all dimensions are then concatenated sequentially to obtain the local transfer bias vector for that paired feature pair. Next, using the similarity of the corresponding feature pairs as weights, all local transfer bias vectors are weighted and summed. The weighted overall local bias is then fused with the global feature distribution difference according to preset weights, ultimately yielding a transfer bias result that comprehensively represents the cross-domain distribution bias. After obtaining the transfer bias, it is substituted into the correction layer of the transfer learning model to compensate for the bias in the target damage evolution representation, ultimately completing the feature distribution difference correction and obtaining a model that accurately represents the aging damage evolution law of the target frame material.

[0045] S404. Adjust the feature mapping parameters in the transfer learning model based on the transfer bias, and use the transfer learning model to correct the feature distribution of the source domain aging memory features and the target aging memory features, so that the damage evolution law of the source domain aging memory features and the target aging memory features in the unified aging memory space is consistent. The specific implementation process of parameter adjustment and correction is as follows: The fusion transfer bias output by S403 is split into independent correction biases of corresponding dimensions according to the output feature dimension, resulting in a correction bias vector of length 32. The bias value of each dimension represents the overall offset direction and offset amount of the target domain feature relative to the source domain feature of the same state. Subsequently, inverse correction is performed according to the bias direction. Since the transfer bias represents the overall distribution offset of the target feature output layer, only the bias parameters of the last fully connected layer of the shared feature extraction backbone need to be adjusted, without modifying other network parameters. Specifically, the obtained correction bias vector is added element-wise to the original output layer bias parameter vector to complete the parameter update. This lightweight adjustment method can accurately compensate for the distribution offset without changing the general aging feature transformation rules learned by the network, avoiding overfitting caused by full network fine-tuning when the target sample size is too small, and also significantly reducing the computational load. After the parameter adjustment is completed, the source domain and target aging memory features are sequentially input into the updated transfer learning model to obtain the corrected source domain and target damage evolution representations. For source domain features, a minor adjustment to the output layer bias does not alter the relative positional relationships of the source domain features, and the correspondence between the labeled lifetime information and the features remains unchanged, preserving the original aging patterns. For target domain features, after bias correction, residual cross-domain distribution offsets are further eliminated, and the distance between target features and source domain features under the same aging cumulative damage state in the unified feature space is significantly reduced, ultimately ensuring that the damage evolution patterns of source and target domain features remain consistent in the unified aging memory space. After this correction step, the transfer learning model completes the cross-domain distribution difference correction, and the final model accurately reflects the aging damage evolution patterns of the target frame material. After obtaining the aging damage evolution model adapted to the target frame material, the currently collected target aging memory features can be input into the model, and the corresponding remaining lifetime prediction results for the target frame can be output.

[0046] S405. Based on the source domain aging memory characteristics and target aging memory characteristics after feature distribution correction, establish the correlation between the damage accumulation process, performance degradation process and aging stage transition process of the target frame material, and obtain an aging damage evolution model applicable to the target frame material based on the correlation.

[0047] The specific modeling and association establishment process in this step is as follows: All source domain damage evolution representations corrected by the S404 feature distribution are used as model input samples. The labeled remaining lifetime corresponding to each source domain sample is used as the supervised output label for that sample. A unified regression training dataset is obtained, while retaining the material properties and environmental parameters corresponding to each source domain sample for validation and parameter tuning during training. This invention uses a lightweight regression output head composed of two sequentially connected fully connected layers. Its input dimension is consistent with the damage evolution representation output by S402, which is 32 dimensions. After a 16-dimensional hidden layer nonlinear transformation, the final output is a one-dimensional continuous value, which is the remaining lifetime result predicted by the model. The training process uses mean squared error as the loss function, with the optimization objective being to minimize the error between the predicted remaining lifetime and the labeled remaining lifetime. The parameters of the regression output head are optimized using a mini-batch gradient descent algorithm. Training is stopped early when the validation set error no longer decreases after 20 consecutive iterations to avoid model overfitting. Because this invention encodes the development information of damage accumulation into aging memory features according to the chronological order of the aging process during the feature extraction and encoding stages, the features themselves already contain information about the complete aging chain of "environmental factors driving - surface structure damage accumulation - molecular chemical structure degradation - macroscopic mechanical property decay - gradual transition of aging stages". During the training process, the regression output head automatically learns the mapping rules of the remaining life corresponding to different damage features, and thus simultaneously establishes the corresponding relationship between the damage accumulation process, the performance degradation process and the aging stage transition process: the feature vectors at different positions in the feature space correspond to different degrees of damage accumulation, the degree of damage accumulation corresponds to different levels of performance degradation, and the damage accumulation crossing the preset threshold corresponds to the transition of the aging stage. This relationship is entirely based on the existing mature data in the source domain for training. Since the feature distribution alignment between the source domain and the target domain has been completed beforehand, this relationship can be directly transferred to the target frame material.

[0048] Obtaining the damage evolution state of the target frame material under different service times includes: S501. Based on the damage accumulation process, performance degradation process and aging stage transition process in the aging damage evolution model, construct a digital damage state model corresponding to the target frame material, and establish a digital damage twin model corresponding to the target frame material based on the digital damage state model. The core function of the digital twin model is to construct a virtual mirror image in digital space that perfectly maps to the actual physical target mirror frame. Based on existing aging damage evolution patterns, it can dynamically simulate and predict the damage development process of the mirror frame throughout its entire lifecycle. The model's construction revolves around three core stages: initial state mapping, evolution rule embedding, and the establishment of an interaction mechanism. For the target mirror frame to be modeled, basic information such as its original material grade, processing parameters, initial mechanical properties, and initial microstructure parameters are collected. If the target mirror frame is already in service, existing damage data is also collected. These parameters and information are input into the model, and an initial digital feature vector is generated according to the feature encoding rules of the aging damage evolution model. This initial vector serves as the starting state for the digital twin model's calculation, ensuring that the initial conditions of the digital model are consistent with the physical entity. The "damage accumulation - performance degradation - stage transition" correlation rules established in the final aging damage evolution model obtained from S405 are fully embedded into the digital twin model's calculation kernel, while damage rate correction coefficients corresponding to different environmental parameters are pre-embedded. The correction coefficient is obtained through synchronous learning from multi-environment aging data in the source domain. It can characterize the influence of different environmental parameters on the damage accumulation rate and is used to dynamically adjust the damage increment calculation results according to the actual service environment. The digital damage twin model has an open entity monitoring data interface, which can be connected to the environmental monitoring sensors and appearance damage online detection module of the target frame service scenario to obtain the actual environmental parameters and actual damage detection results of the physical entity in real time. Whenever a new actual damage detection result is obtained, the actual damage degree is compared with the current predicted damage degree of the digital model. If the deviation exceeds a preset threshold, the current damage state of the digital model is updated based on the actual detection result to eliminate the accumulated calculation error and ensure that the evolution process of the digital model remains synchronized with the physical entity.

[0049] After completing the above three steps, a digital damage twin model corresponding one-to-one with the target frame is obtained. This model can rely on the transfer learning results of the aging damage evolution model, and can achieve high-precision modeling with only a small amount of initial detection information. It does not require large-scale long-term aging tests on the target frame material to obtain modeling data, which greatly reduces the construction cost and cycle of the digital twin model.

[0050] S502. Obtain the real-time service environment parameters corresponding to the target lens frame material, and divide the real-time service environment parameters into time series according to the preset time interval to obtain real-time environmental status data corresponding to different service time stages. The real-time service environment parameters include temperature parameters, humidity parameters, ultraviolet irradiation parameters, and chemical environment parameters. For different application scenarios, this step can adopt two data acquisition methods: For ordinary eyeglass frames without a real-time monitoring module, based on publicly available long-term meteorological and environmental statistics of the target eyeglass frame's usage area, the average temperature, humidity, and ultraviolet radiation intensity data for each time period within the corresponding usage cycle are extracted. Combined with the statistical results of the target user group's daily usage habits, the average frequency and concentration levels of chemical media that the eyeglass frame comes into contact with are supplemented. Then, according to a preset 7-day time interval, the entire service cycle is divided into several continuous stages, and a set of average environmental state data is obtained for each stage, thus completing the time series division. For smart eyeglass frames with a built-in environmental monitoring module, the real-time raw environmental data of the actual service process is collected directly through the embedded micro-sensor at a preset sampling frequency. The raw data in each time interval is processed by a moving average to obtain the average environmental state data for the corresponding stage. This method can filter random noise and accurately capture the dynamic changes in the environment.

[0051] The above four types of parameters correspond to four core environmental factors driving the aging and damage of the frame material: the temperature parameter corresponds to the actual temperature of the service environment. The aging reaction rate of polymer materials is sensitive to temperature, and the higher the temperature, the faster the aging rate; the humidity parameter corresponds to the relative humidity of the environment. High humidity will accelerate the hydrolysis of polar polymer materials, promote the precipitation of small molecule additives, and accelerate performance degradation; the ultraviolet irradiation parameter corresponds to the intensity of ultraviolet irradiation in the environment. It is the core driving factor for photoaging degradation and surface microcrack formation of frame materials in outdoor scenarios, and determines the overall rate of damage accumulation; the chemical environment parameter corresponds to the type and concentration of corrosive chemical media that the frame material comes into contact with, characterizing the promoting effect of external chemical corrosion on the aging process.

[0052] After all processing is completed, the environmental state data corresponding to different service time stages are organized and stored in chronological order as a continuous environmental load time series, which can be used as the basic data for subsequent input into the digital damage twin model, providing driving conditions for the phased dynamic calculation of damage accumulation increment.

[0053] S503. Input the real-time environmental status data into the digital damage twin model, and simulate the environmental load process corresponding to the real-time environmental status data through the aging damage evolution model in the digital damage twin model, so as to obtain the staged damage evolution data corresponding to the target frame material. This step involves iteratively advancing the calculation process according to the service time stages divided in S502. The starting point of the calculation is the initial damage state constructed in S501. The specific single-stage calculation logic is as follows: First, extract the real-time environmental state data corresponding to the current iteration stage, retrieve the pre-embedded damage rate correction coefficient module, and calculate the damage rate correction result adapted to the current environment based on the combination of temperature, humidity, ultraviolet irradiation, and chemical environment parameters of the current stage. Then, input the damage state feature vector stored in the digital twin model after the end of the previous stage, combined with the damage rate correction result of the current stage, into the aging damage evolution model to calculate the newly added damage increment within the current service time stage. Next, fuse the damage increment with the original damage state feature vector according to the aging memory encoding rule to obtain the updated damage state after the end of the current stage. Simultaneously record the core information such as the service time, damage accumulation degree, macroscopic mechanical properties, and current aging stage corresponding to the current stage to complete the calculation of this stage.

[0054] During the calculation process, the digital twin model continuously connects to the entity monitoring interface to obtain the latest damage detection data. After each stage of calculation is completed, it will determine whether new measured damage data has been received. If new measured data exists and the deviation between the measured damage level and the current damage level calculated by the model exceeds a preset threshold, the current damage state stored in the model will be updated directly with the measured damage data to eliminate the previously accumulated calculation error, and then the next stage of iterative calculation will begin, ensuring that the simulation process is always synchronized with the actual damage development of the physical entity.

[0055] The entire iterative calculation process continues until the cumulative service time reaches the preset maximum predicted time, or the current calculated damage accumulation exceeds the preset damage threshold corresponding to frame failure, at which point the iteration stops. Then, the core information recorded in all iteration stages is organized according to the service time to obtain complete staged damage evolution data of the target frame material, covering the entire cycle from the initial state at the factory to final failure.

[0056] S504. Based on the staged damage evolution data, the digital damage state in the digital damage twin model is dynamically updated, and the damage accumulation state, performance degradation state and aging stage state of the target frame material at different service times are obtained according to the updated digital damage state, so as to obtain the damage evolution state of the target frame material at different service times.

[0057] Dynamic updates follow a phased incremental update rule. After S503 completes a single-stage iteration and obtains the latest stage damage data, this step directly replaces the previous stage damage state data stored in the model with the latest damage state data. This eliminates the need to backtrack and recalculate all historical stages, ensuring the timeliness of the state while reducing the model's real-time computational burden. This approach is suitable for scenarios with low-power processors embedded in smart frames. If S503 triggers the measured data correction mechanism—that is, if the actual damage result deviates from the predicted result beyond a preset threshold—this step first corrects the current core damage parameters stored in the model based on the measured results. Then, the corrected state is used as the initial benchmark for subsequent iterations, preventing the accumulation of existing deviations and ensuring the accuracy of the digital model's evolution.

[0058] After completing the status update, this step extracts and decodes three types of target status information from the updated digital damage status: First, the damage accumulation status. This invention uses a standardized damage accumulation index to characterize the degree of damage accumulation. The index ranges from 0 to 1, where 0 corresponds to the initial state of the frame without damage at the factory and 1 corresponds to the critical state of complete material failure. The damage accumulation index corresponding to the service time node can be obtained by decoding the encoded current damage feature vector, quantifying the overall damage degree. Second, the performance degradation status. Based on the damage accumulation-macroscopic performance mapping relationship obtained by training the aging damage evolution model, the current tensile strength, cantilever beam impact strength, visible light transmittance, and other key mechanical and optical performance parameters of the target frame material are converted from the current damage accumulation index, presenting the performance degradation range. Third, the aging stage status. By comparing the current damage accumulation index with the preset aging stage division threshold, the current aging stage of the target frame can be determined. This invention defaults to dividing the aging process into four stages: initial integrity, mild degradation, moderate damage, and severe failure. The threshold for each stage can be pre-adjusted according to the frame material type and usage requirements to adapt to different scenarios and quickly determine the needs.

[0059] The output of the aging performance life prediction results corresponding to the target lens frame material includes: S601. Based on the initial performance state of the target lens frame material and the preset service reliability requirements, establish the preset failure criteria for the target lens frame material. The preset failure criteria include mechanical property degradation threshold, structural damage threshold, chemical degradation threshold and comprehensive damage threshold. Different types of eyeglass frames and different application scenarios have different reliability requirements for material aging performance. A uniform and fixed failure judgment standard cannot adapt to the diverse product needs. It is necessary to establish specific failure criteria based on the specific attributes and usage requirements of the target eyeglass frame. The four types of criteria are as follows: The first category is the mechanical performance degradation threshold, which uses the retention rate of the core mechanical properties of the frame material as the criterion for failure assessment based on the degradation of its load-bearing capacity. Using the initial core mechanical parameters measured at the frame's factory as a benchmark, a minimum performance retention threshold is set based on product design requirements. When the mechanical properties drop below this threshold, the failure condition is met. For example, conventional myopia frames typically require a flexural yield strength retention rate of no less than 70% of the initial value; below this threshold, the frame is prone to bending and breakage, thus meeting the failure requirement.

[0060] The second category is the structural damage threshold, which is based on the degree of damage to the macroscopic structure and appearance of the material. It assesses damage that affects the user experience and the integrity of the appearance. It includes multiple indicators such as the proportion of visible surface damage, the maximum crack length, and the degree of optical performance degradation. When any one of the following exceeds the corresponding preset threshold: the proportion of visible surface damage, the maximum crack length, or the decrease in the visible light transmittance of the transparent frame, it is determined that the structural damage threshold has been reached.

[0061] The third category is the chemical degradation threshold, which is based on the degree of degradation of the internal chemical structure of the material and is used to determine the cumulative degree of aging degradation within the polymer material. Frame materials are mostly polymers, and during aging, chemical reactions such as molecular chain breakage and functional group changes occur. The relative change in characteristic functional groups is usually used as the criterion. For example, the increase in the carbonyl index relative to its initial value can be used as a core parameter. When the increase in the carbonyl index exceeds a preset threshold of 100%, it indicates that the molecular degradation within the material has reached a critical level, and the chemical degradation threshold can be determined.

[0062] The fourth category is the comprehensive damage threshold, which integrates the judgment results of the above three categories of criteria to avoid the limitations of a single criterion. For situations where a single indicator fails to reach the threshold but multiple indicators are close to the critical point, and the material as a whole meets the failure requirements, the three types of damage are first standardized and scored separately. Then, corresponding weights are assigned according to the usage scenario, and a weighted comprehensive damage score is obtained. If the score exceeds the preset comprehensive threshold, the failure condition is determined. This method allows for flexible adjustment of weights; for example, outdoor sunglasses can have increased weights for chemical degradation and structural damage thresholds, while indoor children's myopia frames can have increased weights for mechanical performance degradation thresholds. The specific values ​​of all thresholds can be flexibly adjusted according to industry standards, enterprise quality requirements, and user needs, ensuring the scientific nature of the judgment while adapting to diverse product positioning.

[0063] S602. Obtain the damage evolution state of the target frame material at different service times as output by the digital damage twin model, and perform fusion analysis based on the damage accumulation state, performance degradation state and aging stage state in the damage evolution state to obtain the comprehensive damage state of the target frame material. This step of the fusion analysis process proceeds in three steps: First, for each updated service time point, three types of basic state information are extracted: the damage accumulation index, the current values ​​of each key macroscopic performance, and the aging stage label, ensuring that all input information is consistent with the latest output of the digital damage twin model. Second, the extracted state information of each dimension is matched with the corresponding preset failure criteria established in S601, and the degree of damage deviation from the initial state and approach to the failure threshold of each dimension is quantified and normalized to obtain a separate damage proximity score for each dimension. The score range is uniformly 0~1, and the higher the score, the closer the damage of that dimension is to the failure threshold. Finally, all normalized damage proximity scores are weighted and summed according to the preset weights of each dimension in S601 to obtain the comprehensive damage state value corresponding to the current time point, thus completing the fusion analysis of that time point.

[0064] This step adapts to the phased incremental update rule proposed earlier. Each time S504 completes a digital damage status update, it only needs to perform a fusion calculation on the status information of the latest time node, without needing to recalculate the comprehensive damage status for all historical time nodes. This ensures the timeliness of the comprehensive damage status while maintaining low real-time computational overhead, making it suitable for embedded application scenarios with low computing power. The final comprehensive damage status sequence for the entire service life clearly shows the trend of the overall damage degree of the target frame material gradually increasing with service time.

[0065] S603. Based on the changing trend of the comprehensive damage state, predict the damage evolution process of the target frame material during its future service life, and match the predicted comprehensive damage state with the preset failure criteria to determine the predicted time when the target frame material reaches the failure state. This step matches different calculation paths based on the actual service status of the target frame: The first is for brand-new target frames that have not yet been put into use. The comprehensive damage state change sequence for the entire service life has already been fully generated by the stage-by-stage iterative calculation in S503. It is only necessary to traverse the organized full-cycle time series and retrieve the service time node where the first comprehensive damage state value exceeds the corresponding failure threshold set in S601. The cumulative service time corresponding to this node is the predicted total service life of the target frame material, which can be directly used for material performance calibration and service life labeling of the frame product. The second is for target frames that have been in service for a certain period of time. This method only needs to take the current updated actual damage state as the starting point of the calculation, substitute the subsequent environmental load time series of the target frame's corresponding usage scenario, and continue to calculate the subsequent comprehensive damage growth process stage-by-stage through the digital twin model until the comprehensive damage state meets the preset failure criteria. Finally, the interval from the current time point to the failure time point is calculated, which is the predicted remaining service life of the target frame, which can provide a direct reference for end users to plan the frame replacement time.

[0066] Throughout the entire prediction iteration process, this step will always maintain open access to the entity monitoring interface. If new measured damage data is obtained during the process, and the deviation between the measured damage and the model-predicted damage exceeds a preset threshold, a correction mechanism will be immediately triggered to update the current damage state of the model. Then, based on the updated state, subsequent damage evolution prediction will be carried out again. This process avoids the interference of accumulated calculation deviations on the final lifetime prediction results, ensuring that the prediction results always match the actual aging process of the target frame.

[0067] S604. Determine the failure time corresponding to the target frame material based on the predicted time, and obtain the remaining service life corresponding to the target frame material based on the time difference between the failure time and the current service time of the target frame material, so as to output the aging performance life prediction result corresponding to the target frame material.

[0068] For brand-new eyeglass frames not yet in use, this step directly outputs the total service life prediction result determined by S603. This result can serve as a core evaluation indicator for the aging performance of frame materials, used by manufacturers to conduct material reliability verification, formula iteration optimization, and product quality grading. It can also be directly labeled as the official nominal service life in the product description, providing consumers with a clear reference for purchasing. For eyeglass frames already in actual use, this step first reads the accumulated service time recorded in the model, then subtracts the accumulated service time from the predicted failure time to calculate the current remaining service life of the frame. Finally, the remaining service life is integrated with information such as the current damage accumulation status, performance degradation level, and aging stage of the frame, providing end users with a clear and definite reference for planning frame replacement time.

[0069] Example 2 A system for predicting the aging performance and lifetime of picture frame materials based on transfer learning includes: a processor and a memory for storing processor-executable instructions; wherein the processor is configured to perform the transfer learning-based prediction of the aging performance and lifetime of picture frame materials when executing the executable instructions. It should be noted that the computer device includes a processor and a memory, and may also include one or more of a multimedia component, an input / output (I / O) interface, and a communication component. The processor controls the overall operation of the computer device, completing all or part of the steps of the transfer learning-based prediction of the aging performance and lifetime of picture frame materials. The memory stores various types of data to support device operation and can be implemented by volatile or non-volatile storage devices or combinations thereof, such as SRAM, EEPROM, etc. The multimedia component includes a screen (such as a touch screen) and an audio component. The audio component has a microphone to receive external audio signals and at least one speaker to output audio signals. The I / O interface provides an interface for the processor and other interface modules (such as a keyboard, mouse, buttons, etc.). The communication component is used for wired or wireless communication between devices. Wireless communication includes Wi-Fi, Bluetooth, etc., and the communication component includes a Wi-Fi module, etc. As a preferred embodiment, the computer device can be implemented using electronic components such as ASICs and DSPs to execute a method for predicting the aging performance and lifespan of frame materials based on transfer learning.

[0070] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for predicting the aging performance and lifespan of picture frame materials based on transfer learning, characterized in that, The method includes: Aging test data of different types of frame materials under different aging environmental conditions were collected, and a source domain aging database was constructed based on the aging test data; Based on aging experimental data in the source domain aging database, feature extraction was performed on the performance degradation process of different types of frame materials at different aging stages, and source domain aging memory features characterizing the historical damage evolution of frame materials were constructed according to the performance degradation process. An aging memory coding model is established based on the source domain aging memory features, and the aging memory coding model is used to perform feature mapping on the aging test data corresponding to the target frame material in order to obtain the target aging memory features corresponding to the target frame material. Based on the source domain aging memory features and the target aging memory features, a transfer learning model is constructed, and the feature distribution difference between the source domain aging memory features and the target aging memory features is corrected through the transfer learning model to obtain an aging damage evolution model suitable for the target frame material. A digital damage twin model corresponding to the target lens frame material is constructed based on the aging damage evolution model, and the real-time service environment parameters corresponding to the target lens frame material are input into the digital damage twin model to obtain the damage evolution state of the target lens frame material under different service times. Based on the damage evolution state and combined with preset failure criteria, the failure time of the target frame material is determined, and the aging performance life prediction result of the target frame material is output.

2. The method for predicting the aging performance and lifespan of picture frame materials based on transfer learning as described in claim 1, characterized in that, Constructing source domain aging memory features includes: The aging experiment data in the source domain aging database is divided into stages according to the preset aging time interval to obtain the stage aging data of the frame material at different aging stages. Based on the phased aging data, the performance degradation process of the frame material at different aging stages is characterized by feature extraction to obtain the environmental response characteristics, structural damage characteristics, chemical degradation characteristics, and mechanical property decay characteristics of the frame material. Based on the environmental response characteristics, structural damage characteristics, chemical degradation characteristics, and mechanical property decay characteristics corresponding to different aging stages, the characteristic correlation relationship between the frame material at different aging stages is established, and a continuous degradation characteristic sequence characterizing the evolution process of the frame material from the initial state to the failure state is obtained according to the characteristic correlation relationship. Based on the continuous degradation feature sequence, damage accumulation features, performance degradation rate features, and aging stage transition features of the frame material during the aging process are extracted. The damage accumulation features, performance degradation rate features, and aging stage transition features are then encoded to obtain source domain aging memory features that characterize the historical damage evolution of the frame material.

3. The method for predicting the aging performance and lifespan of picture frame materials based on transfer learning as described in claim 1, characterized in that, The target aging memory characteristics corresponding to the target frame material include: Based on the source domain aging memory characteristics, the damage accumulation characteristics, performance degradation rate characteristics, and aging stage transition characteristics of different types of frame materials are organized into features, and an aging memory feature space is constructed based on the feature organization results to characterize the historical damage evolution law of different types of frame materials. An aging memory coding model is established based on the aging memory feature space, and the aging memory coding model is trained using source domain aging memory features so that the aging memory coding model establishes a mapping relationship between the performance degradation process of the frame material and the aging memory feature space. Acquire aging test data corresponding to the target lens frame material, and extract features from the environmental load parameters, surface morphology change parameters, chemical structure change parameters and mechanical property degradation parameters in the aging test data to obtain the target environmental response characteristics, target structural damage characteristics, target chemical degradation characteristics and target mechanical property decay characteristics corresponding to the target lens frame material. The target environmental response characteristics, target structural damage characteristics, target chemical degradation characteristics, and target mechanical property decay characteristics are input into the aging memory coding model. Based on the mapping relationship, the performance degradation information corresponding to the target frame material is mapped to the aging memory feature space to obtain the initial target aging memory characteristics corresponding to the target frame material. Based on the feature correlation between the initial target aging memory features and the source domain aging memory features, feature calibration is performed on the initial target aging memory features to obtain target aging memory features that characterize the cumulative damage state of the target frame material during service.

4. The method for predicting the aging performance and lifespan of picture frame materials based on transfer learning as described in claim 3, characterized in that, Training the aging memory encoding model includes: According to the aging stage sequence corresponding to the frame material, the damage accumulation feature, performance degradation rate feature and aging stage transition feature in the source domain aging memory feature are arranged in time sequence, and the damage accumulation feature, performance degradation rate feature and aging stage transition feature corresponding to the same aging stage are combined to obtain the aging memory feature vector corresponding to different aging stages. Based on the aging memory feature vectors corresponding to different aging stages and the temporal relationship between each aging memory feature vector, an aging memory feature sequence corresponding to the frame material is constructed, and the aging stage position information is configured for each aging memory feature vector in the aging memory feature sequence to obtain an aging memory input sequence containing the temporal relationship of the performance degradation of the frame material. The aging memory input sequence is input into the aging memory coding model. The aging memory coding model extracts the local degradation correlation features between adjacent aging stages and the long-term damage evolution correlation features between different aging stages. The local degradation correlation features and the long-term damage evolution correlation features are then fused to obtain the aging memory coding representation corresponding to the frame material. The aging memory encoding representation is mapped to the aging memory feature space, and the model parameters of the aging memory encoding model are adjusted according to the reconstruction deviation between the aging memory encoding representation and the corresponding source domain aging memory features, as well as the evolutionary continuity deviation between the aging memory encoding representations corresponding to adjacent aging stages. Repeatedly perform the acquisition of aging memory encoding representation, mapping of aging memory encoding representation to aging memory feature space, and adjustment of model parameters until the reconstruction bias and evolution continuity bias meet the preset convergence conditions to obtain the trained aging memory encoding model.

5. The method for predicting the aging performance and lifespan of picture frame materials based on transfer learning as described in claim 1, characterized in that, Obtaining an aging damage evolution model applicable to the target frame material includes: Based on source domain aging memory features and target aging memory features, source domain aging memory feature sets and target aging memory feature sets are constructed respectively, and the transfer relationship between source domain aging memory features and target aging memory features is established according to the feature association relationship between the source domain aging memory feature sets and target aging memory feature sets. A transfer learning model is constructed based on the transfer relationship. The source domain aging memory feature set and the target aging memory feature set are input into the transfer learning model to obtain the source domain damage evolution representation corresponding to the source domain aging memory feature and the target damage evolution representation corresponding to the target aging memory feature, respectively. Based on the source domain damage evolution representation and the target damage evolution representation, the feature distribution difference between the source domain aging memory features and the target aging memory features is calculated, and the transfer bias between the source domain aging memory features and the target aging memory features is obtained based on the feature distribution difference. Based on the transfer bias, the feature mapping parameters in the transfer learning model are adjusted, and the feature distribution of the source domain aging memory features and the target aging memory features is corrected using the transfer learning model, so that the damage evolution law of the source domain aging memory features and the target aging memory features is consistent in the unified aging memory space. Based on the source domain aging memory features and target aging memory features after feature distribution correction, the correlation between the damage accumulation process, performance degradation process and aging stage transition process of the target frame material is established, and an aging damage evolution model applicable to the target frame material is obtained based on the correlation.

6. The method for predicting the aging performance and lifespan of picture frame materials based on transfer learning as described in claim 1, characterized in that, Obtaining the damage evolution state of the target frame material under different service times includes: Based on the damage accumulation process, performance degradation process and aging stage transition process in the aging damage evolution model, a digital damage state model corresponding to the target frame material is constructed, and a digital damage twin model corresponding to the target frame material is established based on the digital damage state model. The real-time service environment parameters corresponding to the target lens frame material are obtained, and the real-time service environment parameters are divided into time series according to a preset time interval to obtain real-time environmental status data corresponding to different service time stages. The real-time service environment parameters include temperature parameters, humidity parameters, ultraviolet irradiation parameters, and chemical environment parameters. Real-time environmental status data is input into the digital damage twin model, and the aging damage evolution model in the digital damage twin model is used to simulate the environmental load process corresponding to the real-time environmental status data in order to obtain the staged damage evolution data of the target frame material. The digital damage state in the digital damage twin model is dynamically updated based on the phased damage evolution data. The damage accumulation state, performance degradation state and aging stage state of the target frame material at different service times are obtained according to the updated digital damage state, so as to obtain the damage evolution state of the target frame material at different service times.

7. The method for predicting the aging performance and lifespan of picture frame materials based on transfer learning as described in claim 1, characterized in that, The output aging performance and lifetime prediction results for the target lens frame material include: Based on the initial performance state of the target lens frame material and the preset service reliability requirements, preset failure criteria are established for the target lens frame material. The preset failure criteria include mechanical property degradation threshold, structural damage threshold, chemical degradation threshold and comprehensive damage threshold. The damage evolution state of the target frame material at different service times is obtained from the output of the digital damage twin model. The damage accumulation state, performance degradation state and aging stage state in the damage evolution state are fused and analyzed to obtain the comprehensive damage state of the target frame material. Based on the changing trend of the comprehensive damage state, the damage evolution process of the target frame material during its future service life is predicted, and the predicted comprehensive damage state is matched with the preset failure criteria to determine the predicted time when the target frame material reaches the failure state. The failure time of the target frame material is determined based on the predicted time, and the remaining service life of the target frame material is obtained based on the time difference between the failure time and the current service time of the target frame material, so as to output the aging performance life prediction result of the target frame material.

8. A system for predicting the aging performance and lifespan of picture frame materials based on transfer learning, characterized in that, The system includes: processor; Memory used to store processor-executable instructions; The processor is configured to implement the transfer learning-based method for predicting the aging performance of frame materials according to any one of claims 1 to 7 when executing executable instructions.