Natural rubber aging state evaluation method and system based on machine learning

Abstract

The invention discloses a natural rubber aging state evaluation method and system based on machine learning, and belongs to the technical field of machine learning, and the method comprises the following steps: 1, constructing a first parameterization function family corresponding to synthetic rubber and a second parameterization function family corresponding to natural rubber; step 2, performing parameter optimization on the first parameterization function family by using the obtained aging data sequence of the natural rubber, and generating a rule deviation vector; 3, according to the rule deviation vector, analyzing the dimension sensitivity of the original signal space to determine a specific sensitive channel; according to the method, a differential parameterization function family of synthetic rubber and natural rubber is constructed, the core difference of aging laws of two types of rubber is mined in combination with a law deviation vector, and a specific sensitive channel is extracted in a targeted manner, so that the problem of negative migration caused by the difference of aging mechanisms of synthesis and natural rubber in traditional migration learning is effectively avoided; and the accuracy of natural rubber aging state evaluation is improved.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and more specifically, to a method and system for assessing the aging state of natural rubber based on machine learning. Background Technology

[0002] Due to its unique molecular structure, the aging process of natural rubber involves the coupling of multiple mechanisms such as complex oxidation, cross-linking, and microstructure rearrangement. This makes it costly and time-consuming to obtain sufficient, accurate, and fully labeled aging state data through experiments. To address the problem of data scarcity, a common technical approach is to use a transfer learning strategy. This involves first using a large amount of data from synthetic rubbers with relatively regular molecular structures, such as styrene-butadiene rubber (SBR), which are easier to obtain for accelerated aging experiments, to pre-train machine learning models such as deep neural networks. This allows the models to learn generalized aging characteristics. Then, a small amount of natural rubber data is used to fine-tune the models in order to assess the aging state of natural rubber. This method often relies on the transfer and updating of model parameters under the control of computer programs.

[0003] Synthetic rubber and natural rubber differ fundamentally in molecular chain structure, unsaturated bond distribution, and antioxidant reaction characteristics, leading to different dominant aging mechanisms and kinetic pathways. For example, the aging of synthetic rubber may be more inclined towards a uniform oxidation chain reaction, while the aging of natural rubber may be affected by multiple competing processes, such as oxidation induction and the ordering of natural amorphous regions. When a pre-trained model learns on synthetic rubber data, the characteristic aging state mapping relationship constructed internally is essentially a mathematical representation of the specific aging laws of synthetic rubber. When the features of this pre-trained model are directly extracted based on natural rubber data, due to the systematic bias of the underlying physicochemical driving factors, a profound mismatch will occur between the feature space on which the model depends and the signal change patterns induced by the actual aging process of natural rubber. This mismatch not only makes it difficult for the model to converge or achieve poor results during the fine-tuning stage, but more commonly, it triggers negative transfer, that is, the prior knowledge obtained from pre-training interferes with the model's learning of the specific aging patterns of natural rubber, causing the final evaluation results to deviate from reality with systematic errors, making it difficult to meet the reliability and accuracy requirements of industrial applications. Summary of the Invention

[0004] To address the problems existing in the prior art, the present invention aims to provide a machine learning-based method and system for assessing the aging state of natural rubber. This method can construct a family of differentiated parameterized functions for synthetic and natural rubber, combine them with regular deviation vectors to mine the core differences in the aging patterns of the two types of rubber, and specifically extract sensitive channels. This effectively avoids the negative transfer problem caused by the difference in aging mechanisms between synthetic and natural rubber in traditional transfer learning, thereby improving the accuracy of natural rubber aging state assessment.

[0005] This invention is based on a machine learning framework. By constructing a family of differentiated parameterized functions, generating regular deviation vectors, extracting specific sensitive channels, and designing a collaborative training and conditional evaluation mechanism, it achieves accurate assessment of the aging state of natural rubber and effectively avoids the negative transfer problem in traditional transfer learning.

[0006] To solve the above problems, the present invention adopts the following technical solution: Firstly, a machine learning-based method for assessing the aging state of natural rubber includes the following steps: Step 1: Construct the first family of parameterized functions for synthetic rubber and the second family of parameterized functions for natural rubber, respectively; Step 2: Use the obtained natural rubber aging data sequence to perform parameter optimization on the first parameterized function family to generate a regular deviation vector; Step 3: Based on the deviation vector, analyze the dimensionality sensitivity of the original signal space to determine the specific sensitive channels; Step 4: Use specific sensitive channels to extract the divergence enhancement signal set and the full signal set from the labeled natural rubber aging data, and use the two to train the first evaluation unit and the second discrimination unit together; wherein, the first evaluation unit fits the divergence enhancement signal set, the second discrimination unit makes regularity judgments based on the full signal set, and the two establish a correlation through a coupled loss function; Step 5: After the collaborative training reaches the predetermined convergence condition, solidify the second discrimination unit as the pattern identification unit, and configure the conditional evaluation unit based on the output of the pattern identification unit. Step 6: For the sample to be tested, obtain its divergence enhancement signal and full signal through a specific sensitive channel; input the full signal into the regularity discrimination unit to obtain the discrimination score, adjust the parameter state of the conditional evaluation unit accordingly, and input the divergence enhancement signal into the adjusted conditional evaluation unit.

[0007] Furthermore, the construction of the parameterized function family includes: Step 11: Analyze the aging chain reaction pathways of synthetic rubber and natural rubber respectively, and extract the first set of control variables for synthetic rubber and the second set of control variables for natural rubber. The first set of control variables includes the free radical generation rate and the chain termination efficiency, and the second set of control variables also includes the degradation rate of the activity of amorphous region chain segments. Step 12: Based on the first set of control variables and the second set of control variables, construct a unidirectional saturated coupling network for synthetic rubber and a triangular asymmetric coupling network for natural rubber. Step 13: Instantiate the unidirectional saturated coupled network and the triangular asymmetric coupled network into a dynamic system, assign parameterized nonlinear functions to the coupling edges, and define macroscopic response values ​​based on the state variables of the dynamic system to generate the first family of parameterized functions and the second family of parameterized functions.

[0008] Furthermore, the generation of the regular deviation vector includes: Step 21: Perform phase space reconstruction on the natural rubber aging data sequence to generate empirical state trajectories; Step 22: Calculate the Hausdorff distance spectrum between the empirical state trajectory and the simulated state trajectory generated by the first family of parameterized functions. Step 23: Perform geodesic-based parameter perturbation on the coupled network of synthetic rubber based on the Hausdorff distance spectrum; Step 24: Extract the edge weight changes of the coupled network before and after the disturbance that exceed a predetermined threshold, and arrange them in topological order to generate a regular deviation vector.

[0009] Furthermore, the identification of specific sensitive channels includes: Step 31: Decode the regular deviation vector into a virtual physical field excitation mode; Step 32: Based on the virtual physical field excitation mode, simulate the disturbance response of each dimension of the original signal space to generate a full-dimensional disturbance response spectrum; Step 33: Perform mode decomposition on the full-dimensional perturbation response spectrum and extract coherent sensitive clusters; Step 34: Calculate the topological density of the coherent sensitive cluster in the original signal space, and determine the specific sensitive channel based on the topological density.

[0010] Furthermore, the steps for collaboratively training the first evaluation unit and the second discriminant unit include: Step 41: Modulate the entire signal set based on the specific sensitive channel to generate a modulated signal, and project the modulated signal onto the difference subspace to obtain the divergence-enhanced signal set; Step 42: Based on the statistical summary of the modulated signal, dynamically generate the convolutional kernel weight set for the first evaluation unit and the second discrimination unit.

[0011] Furthermore, the collaborative training first evaluation unit and second discrimination unit also include: Step 43: Based on the fitting loss gradient of the first evaluation unit and the discrimination loss gradient of the second discrimination unit, dynamically adjust the learning rate of the parameters of both units, and exchange feature summary information during the adjustment process. Step 44: Perform bidirectional attention calibration on the high-level feature maps of the first evaluation unit and the second discrimination unit.

[0012] Furthermore, the configuration of the regularity identification unit and the conditional evaluation unit includes: Step 51: Perform sparsity pruning on the second discrimination unit; Step 52: Construct a nonlinear mapping rule from the decision space of the second discriminant unit after sparse pruning to the parameter subspace of the first evaluation unit; Step 53: Based on the nonlinear mapping rule, the pre-stored basic parameter blocks are weighted and combined to dynamically generate a subset of parameters for the conditional evaluation unit. Step 54 introduces a softening function based on decision space distance during the weighted combination process to establish a smooth transition mechanism between parameter subsets.

[0013] Furthermore, the evaluation of the sample to be tested includes: Step 61: Perform phase synchronization processing on the full signal and the divergence enhancement signal of the sample to be tested to generate a phase-synchronized divergence enhancement signal; Step 62: Utilize the feature extraction kernel of the dynamic configuration conditional evaluation unit for the entire signal after phase synchronization.

[0014] Furthermore, the evaluation of the sample to be tested also includes: Step 63: Feed the phase-synchronized divergence enhancement signal into the configured conditional evaluation unit, and calculate the sub-band attention mask based on the hybrid weights corresponding to the parameter subsets to modulate the intermediate feature map. Step 64: Perform consistency verification based on adversarial perturbation on the preliminary evaluation results of the conditional evaluation unit, and output the final evaluation results based on the verification results.

[0015] Secondly, a machine learning-based system for assessing the aging state of natural rubber includes: The function family generation module is used to construct the first parameterized function family corresponding to synthetic rubber and the second parameterized function family corresponding to natural rubber, respectively. The parameter optimization module is used to perform parameter optimization on the first parameterized function family using the aging data sequence of natural rubber, and generate a regular deviation vector; The channel determination module analyzes the dimensionality sensitivity of the original signal space based on the regular deviation vector to determine the specific sensitive channels; The collaborative training module is used to extract the divergence enhancement signal set and the full signal set from the labeled natural rubber aging data using a specific sensitive channel, and to use the two to collaboratively train the first evaluation unit and the second discrimination unit. The first evaluation unit fits the divergence enhancement signal set, and the second discrimination unit makes regularity judgments based on the full signal set. The two are associated through a coupled loss function. The unit configuration module is used to solidify the second discriminant unit as a pattern identification unit after the collaborative training reaches the predetermined convergence condition, and to configure the conditional evaluation unit based on the output of the pattern identification unit. The state evaluation module is used to acquire the divergence enhancement signal and the full signal of the sample under test through a specific sensitive channel; the full signal is input into the regularity discrimination unit to obtain the discrimination score, and the parameter state of the conditional evaluation unit is adjusted accordingly, and the divergence enhancement signal is input into the adjusted conditional evaluation unit.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This scheme constructs a family of differentiated parameterized functions for synthetic rubber and natural rubber, combines the regular deviation vector to mine the core differences in the aging patterns of the two types of rubber, and extracts specific sensitive channels in a targeted manner. This effectively avoids the negative transfer problem caused by the difference in aging mechanisms between synthetic and natural rubber in traditional transfer learning, improves the accuracy of natural rubber aging status assessment, and meets the industrial demand for assessment accuracy.

[0017] (2) Based on nonlinear mapping rules, this scheme realizes the accurate correlation between the decision output of the rule identification unit and the parameters of the conditional evaluation unit. By dynamically generating parameter subsets through weighted combination of pre-stored basic parameter blocks, it can adapt to different aging scenarios without retraining, greatly reducing the model adaptation cost and improving evaluation efficiency and scenario applicability.

[0018] (3) This scheme ensures the alignment of timing features by synchronizing the phase of the full signal and the divergence enhancement signal; it also strengthens key aging features by combining sub-band attention masking to suppress noise interference, enabling the conditional evaluation unit to accurately capture the specific aging signal of natural rubber, thereby further improving the accuracy and stability of the evaluation results.

[0019] (4) This scheme introduces an adversarial perturbation consistency verification mechanism. By generating small reasonable perturbations, the robustness of the preliminary evaluation results is verified, effectively filtering random noise and detecting evaluation deviations caused by anomalies, ensuring the reliability of the final evaluation results, and providing strong support for the industrial application of natural rubber aging status assessment. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0021] Figure 1 This is a flowchart of a machine learning-based method for assessing the aging state of natural rubber according to the present invention. Figure 2 This is a flowchart illustrating the relationships between various modules in a machine learning-based natural rubber aging status assessment system according to the present invention. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0023] Example 1: The machine learning technologies involved in this solution mainly include deep learning, such as attention mechanisms, as well as improved applications of transfer learning, such as avoiding negative transfer through parameterized function families, and branches of meta-learning such as dynamic parameter generation and adversarial learning for consistency verification. The fusion of multiple branch technologies enables accurate assessment of the aging state of natural rubber.

[0024] Please refer to details. Figure 1 A machine learning-based method for assessing the aging state of natural rubber, comprising: Step 1: Construct the first family of parameterized functions for synthetic rubber and the second family of parameterized functions for natural rubber, respectively; The construction of the parameterized function family also includes the following steps: Step 11: Analyze the aging chain reaction pathways of synthetic rubber and natural rubber respectively, and extract the first set of control variables for synthetic rubber and the second set of control variables for natural rubber. The first set of control variables includes the free radical generation rate and chain termination efficiency, while the second set of control variables also includes the decay rate of amorphous region chain segment activity. The specific operations are as follows: To construct a family of parameterized functions suitable for the aging characteristics of synthetic and natural rubber, it is necessary to first clarify the driving mechanisms of their aging processes. Therefore, the analysis of the aging chain reaction pathways of synthetic and natural rubber is conducted first. The analysis begins with the differences in their molecular structures, combining the chemical kinetics of aging reactions such as oxidation and cross-linking, to trace the material transformation and energy transfer processes in key stages such as reaction initiation, chain growth, and termination. For synthetic rubber, its molecular chain structure is relatively regular, and the aging process is dominated by a uniform oxidation chain reaction. Based on this analysis, a first set of control variables capable of characterizing the regulatory nodes of this aging pathway is extracted. The set mainly includes the free radical generation rate and chain termination efficiency. These two variables directly determine the initiation speed and decay characteristics of the oxidation chain reaction of synthetic rubber. However, due to the more complex distribution of unsaturated bonds in the molecular chain structure and the presence of natural amorphous regions, the aging process of natural rubber not only includes oxidation chain reactions but also multiple competing processes such as the ordering of amorphous regions. The aging mechanism is more complex. Therefore, when analyzing its aging chain reaction path, in addition to considering the free radical generation rate and chain termination efficiency shared with synthetic rubber, it is also necessary to extract the decay rate of amorphous region chain segments that characterize the aging characteristics of amorphous regions. This forms a second set of control variables that covers three types of variables.

[0025] Step 12: Based on the first set of control variables and the second set of control variables, construct the unidirectional saturated coupling network for synthetic rubber and the triangular asymmetric coupling network for natural rubber. The specific operations are as follows: After obtaining the first set of control variables for synthetic rubber and the second set of control variables for natural rubber, a coupling network structure adapted to both types of rubber is constructed based on the differences in aging driving characteristics represented by the control variable sets. For synthetic rubber, its first set of control variables only includes the free radical generation rate and chain termination efficiency, and the corresponding aging chain reaction path exhibits a unidirectional progressive characteristic. The interaction between the control variables during the reaction process has a clear sequence and saturation characteristics. Therefore, a unidirectional saturated coupling network is specifically constructed. This network characterizes the unidirectional regulatory effect of the free radical generation rate on the chain termination efficiency through directed edges, while introducing a saturation constraint mechanism to match the aging reaction of synthetic rubber. The decay law of the influence of control variables in the process; for natural rubber, the second set of control variables adds the decay rate of amorphous region chain segment activity. This variable has multiple competitive relationships with the original two types of variables, and is not a simple one-way action. Therefore, a triangular asymmetric coupling network is constructed, and the three types of control variables are respectively used as nodes of the network. The difference in the intensity of interaction between variables is represented by asymmetric directed edges. Among them, the edge weight configuration between the amorphous region chain segment activity decay rate node and the free radical generation rate and chain termination efficiency nodes focuses on matching their interference effect on the oxidation chain reaction, so as to realize the networked characterization of the complex aging mechanism of natural rubber.

[0026] Step 13: Instantiate the unidirectional saturated coupled network and the triangular asymmetric coupled network as dynamic systems, assign parameterized nonlinear functions to the coupling edges, and define macroscopic response values ​​based on the state variables of the dynamic system to generate the first and second parameterized function families. The specific operations are as follows: To transform the constructed coupled network into a quantifiable and computable function model, the unidirectional saturated coupled network and the triangular asymmetric coupled network need to be further instantiated into dynamic systems. During the instantiation process, a parameterized nonlinear function is assigned to each coupled edge in both types of networks. The parameter dimension and value range of this function are determined according to the physicochemical properties of the corresponding control variables. For example, the parameterized nonlinear function of the coupled edge characterizing the relationship between the free radical generation rate and the chain termination efficiency needs to incorporate the influence coefficients of environmental factors such as temperature and oxygen concentration to characterize the changes in aging reaction rate under different conditions. At the same time, a macroscopic response value is defined based on the state variables of the dynamic system. Here, the state variables are the real-time quantized values ​​of each node in the coupled network, and the macroscopic response value is selected from physical quantities that can reflect the macroscopic state of rubber aging, such as the tensile strength decay rate and the change in crosslinking density. By establishing a mapping relationship between the state variables and the macroscopic response value, the interaction of variables within the network is transformed into observable and verifiable macroscopic indicators. Through the above process, a first family of parameterized functions that can accurately characterize the aging law of synthetic rubber and a second family of parameterized functions that adapt to the complex aging mechanism of natural rubber are finally generated.

[0027] In a preferred embodiment of the present invention, step 2 is further included: using the acquired natural rubber aging data sequence to perform parameter optimization on the first parameterized function family to generate a regular deviation vector. The generation of the regular deviation vector also includes the following steps: Step 21: Perform phase space reconstruction on the natural rubber aging data sequence to generate empirical state trajectories. The specific operations are as follows: First, an accelerated aging method using an oven, closely mimicking the results of natural aging, was employed. Natural rubber vulcanizate samples were prepared according to a standard formula and placed in an aging test chamber with controlled environmental parameters, such as temperature, humidity, and oxygen partial pressure, for aging treatment. Subsequently, at preset time intervals, key aging characterization indicators of the samples were tested using professional instruments. Tensile strength was measured using an electronic tensile testing machine, crosslinking density was measured using a nuclear magnetic resonance (NMR) crosslinking density meter, and infrared characteristic peak intensity was detected using an infrared spectrometer. Finally, the test data from different time points were organized chronologically to form a time-series signal set reflecting the dynamic changes in the natural rubber aging process, i.e., the natural rubber aging data sequence.

[0028] Natural rubber aging data sequences are time-series signals reflecting the aging process, containing dynamic changes in multiple dimensions such as tensile strength, crosslinking density, and infrared characteristic peak intensity. These information are scattered in one-dimensional time-series data, making it difficult to directly reflect the dynamic evolution characteristics and intrinsic relationships of the aging process. To uncover the true aging state evolution laws of natural rubber contained in the data sequences, a phase space reconstruction operation needs to be performed on the natural rubber aging data sequences. The reconstruction process uses the time delay τ and the embedding dimension m of the time-series data as core parameters. The time delay τ is determined by calculating the moment when the autocorrelation function of the data sequence first approaches zero, and the embedding dimension m is selected based on the spurious nearest neighbor method to ensure that the reconstructed phase space can completely restore the dynamic characteristics of the aging process. By translating and stacking the one-dimensional aging data sequences according to the time delay τ, a set of discrete points in a high-dimensional phase space is constructed. The trajectory formed by connecting these discrete points in time order is the empirical state trajectory, which completely describes the co-evolution path of various state variables in the natural rubber aging process.

[0029] Step 22: Based on the empirical state trajectory and the simulated state trajectory generated by the first family of parameterized functions, calculate the Hausdorff distance spectrum between the two. The specific operation is as follows: The first family of parameterized functions is a mathematical model built based on the aging law of synthetic rubber. By inputting the same environmental parameters as those in the natural rubber aging experiment, such as temperature, humidity, and oxygen partial pressure, this family of functions can generate simulated state trajectories characterizing the evolution of the aging state of synthetic rubber under corresponding conditions. To quantify the differences between the aging laws of natural and synthetic rubber, it is necessary to calculate the Hausdorff distance spectrum between the empirical state trajectory and the simulated state trajectory. The Hausdorff distance reflects the maximum and minimum distance between two sets of points. The logic is to first calculate the minimum distance from each point in the empirical state trajectory to the simulated state trajectory, and then take the maximum value among these minimum distances. At the same time, the minimum distance from each point in the simulated state trajectory to the empirical state trajectory is calculated and the maximum value is taken. Finally, the larger of the two maximum values ​​is taken as the Hausdorff distance at a single scale. By repeating the above calculation process at different time scales and different state variable dimensions, a Hausdorff distance spectrum containing multiple distance values ​​is obtained. The magnitude of this distance spectrum directly characterizes the degree of difference in the laws of the two types of rubber at the corresponding aging stages and corresponding state dimensions. The larger the distance value, the worse the adaptability of the synthetic rubber aging law to the natural rubber aging process.

[0030] Step 23: Based on the Hausdorff distance spectrum, perform geodesic-based parameter perturbation on the coupled network of synthetic rubber. The specific operation is as follows: Hausdorff distance spectrum clearly reveals the differential distribution of aging patterns between synthetic and natural rubber. This difference essentially stems from the different coupling network structures and parameter configurations of the two types of rubber. Therefore, targeted parameter perturbations need to be applied to the coupling network of synthetic rubber based on the distance spectrum to locate the core parameter nodes causing the differences in patterns. Considering that the unidirectional saturated coupling network of synthetic rubber has a clear topological structure and parameter correlations, a geodesic-based parameter perturbation method is adopted to avoid disrupting the inherent aging response logic of the network during perturbation. A geodesic characterizes the shortest path between two points in the topological space of the coupling network. Geodesic-based perturbation can accurately apply perturbations along the effective influence path of network parameters. Specifically, the parameter nodes of the coupling network are first prioritized according to the magnitude of the Hausdorff distance spectrum values. Parameter nodes corresponding to dimensions with larger distance spectrum values ​​are set as high-priority perturbation objects. Then, parameter perturbations are applied along the geodesic direction corresponding to each parameter node, with a perturbation amplitude positively correlated with the distance spectrum value. The larger the distance spectrum value, the larger the perturbation amplitude, thereby achieving differentiated and precise adjustment of the parameters of the synthetic rubber coupling network.

[0031] Step 24: Extract the edge weight changes exceeding a predetermined threshold in the coupled network before and after the perturbation, and arrange them in topological order to generate a regular deviation vector. The specific operation is as follows: After perturbating the parameters of the synthetic rubber coupled network, the weights of each coupled edge in the network will change due to the perturbation. These changes in edge weights contain both effective information related to the differences in the aging patterns of the two types of rubber and random noise generated during the perturbation process. To filter out the effective information, a predetermined threshold needs to be set. This threshold is determined by statistically analyzing the distribution of edge weight changes in multiple blank perturbation experiments, where blank perturbation experiments are perturbations without actual differences in aging patterns. Typically, the value corresponding to three times the standard deviation of the edge weight change distribution in the blank experiments is taken as the predetermined threshold. To ensure effective removal of random noise, the absolute value of the change in edge weights before and after the perturbation in the coupled network is extracted if it exceeds a predetermined threshold. The changes in edge weights directly reflect the parameter information that the synthetic rubber coupling network needs to adjust to adapt to the aging law of natural rubber. To facilitate the subsequent analysis of the sensitivity of the original signal spatial dimension, the extracted edge weight changes need to be arranged according to the topological order of the synthetic rubber coupling network, that is, along the action path of the control variables in the network, from the initial node to the terminal node, specifically from the free radical generation rate node to the chain termination efficiency node. The final ordered set of edge weight changes is the law deviation vector, which completely describes the difference distribution of the aging laws of synthetic rubber and natural rubber.

[0032] In a preferred embodiment of the present invention, step 3 is further included: based on the regular deviation vector, the dimensional sensitivity of the original signal space is analyzed to determine the specific sensitive channel; The determination of specific sensitive channels also includes the following steps: Step 31: Decode the regular deviation vector into a virtual physics field excitation mode. The specific operation is as follows: The deviation vector is a set of ordered edge weight changes that characterize the core differences in the aging patterns of synthetic and natural rubber. Essentially, it represents abstract network parameter changes and cannot be directly used for dimensionality sensitivity analysis of the original signal space. Therefore, it needs to be decoded into a virtual physical field excitation mode with clear physical meaning. The decoding process uses the topology and physical meaning of the synthetic rubber coupled network as a bridge to establish a mapping relationship between the edge weight changes in the deviation vector and aging-related physical field parameters, such as the local stress field, equivalent oxidation concentration field, and molecular chain activity energy field. Specifically, the edge weight changes related to the free radical generation rate are decoded as changes in the excitation intensity of the equivalent oxidation concentration field; the edge weight changes related to the decay rate of chain activity in the amorphous region are decoded as adjustments in the excitation distribution of the molecular chain activity energy field. Through this mapping transformation, the abstract parameter difference information is converted into a virtual physical field distribution mode that can act on the original signal generation process. This mode not only includes the spatial distribution information of the excitation intensity but also encompasses the evolution characteristics of the excitation over time.

[0033] Step 32: Based on the virtual physical field excitation mode, simulate the perturbation response in each dimension of the original signal space to generate a full-dimensional perturbation response spectrum. The specific operations are as follows: The original signal space consists of signals from various detection dimensions during the aging process of natural rubber, covering multiple dimensions such as tensile strength, crosslinking density, infrared characteristic peak intensity, and dynamic mechanical properties. The sensitivity of each dimension to the differences in aging patterns varies significantly. Based on the virtual physical field excitation mode obtained in step 31, the perturbation response process of each signal dimension under different physical field excitations is simulated through the established original signal response simulation model. During the simulation, the virtual physical field excitation mode is used as input and substituted into the physical generation model corresponding to each signal dimension, such as the correlation model between tensile strength and molecular chain crosslinking degree, and the quantitative model between infrared characteristic peak intensity and oxidation product concentration. The numerical change of each signal dimension under excitation is calculated. By repeating the above simulation under different excitation intensities and different excitation time points, a full-dimensional perturbation response spectrum containing response data of each signal dimension in the full excitation range is obtained. In this response spectrum, the response amplitude of each dimension signal directly reflects its sensitivity to the differences in aging patterns between synthetic rubber and natural rubber. The larger the response amplitude, the easier it is for that dimension to capture the differences in aging patterns between the two types of rubber.

[0034] Step 33: Perform mode decomposition on the full-dimensional perturbation response spectrum and extract coherent sensitive clusters. The specific operations are as follows: The full-dimensional perturbation response spectrum contains complex response information from multiple signal dimensions, with some dimensions exhibiting strong correlations. These correlations stem from the common aging physicochemical mechanisms underlying each dimension. Therefore, it is necessary to separate sensitive components with synergistic response characteristics through mode decomposition. The mode decomposition process, based on the frequency characteristics and phase consistency of the signal, decomposes the full-dimensional perturbation response spectrum into multiple independent modal components. Each modal component corresponds to a class of response modes with similar frequency characteristics and phase changes. Subsequently, by calculating the correlation coefficient between each modal component and the deviation vector of the law, modal components with correlation coefficients higher than a preset threshold are selected. These components are the effective response components directly related to the differences in aging laws. Further cluster analysis is performed on the signal dimensions corresponding to the effective response components, grouping signal dimensions with highly similar response characteristics and strong phase synchronization into one class to form coherent sensitive clusters. Each coherent sensitive cluster represents a set of signal dimensions driven by the same aging law difference mechanism, thus realizing the structured extraction of the original complex response information.

[0035] Step 34: Calculate the topological density of the coherent sensitive clusters in the original signal space, and determine the specific sensitive channels based on the topological density. The specific operations are as follows: The topological compactness of a coherent sensitive cluster is an indicator that measures the degree of correlation between signal dimensions within the cluster and the stability of the response. It directly determines whether the cluster can accurately capture the specific aging signals of natural rubber. Therefore, it is necessary to screen high-quality coherent sensitive clusters as specific sensitive channels by calculating the topological compactness. The calculation of topological compactness is based on the distribution position and correlation strength of each signal dimension within the cluster in the original signal space. The specific logic is as follows: First, calculate the correlation coefficient of each signal dimension within the cluster with all other dimensions to obtain the correlation matrix; then, calculate the average correlation strength of the dimensions within the cluster based on the correlation matrix, and combine it with the spatial distribution entropy of each dimension in the signal space. The topological compactness value is obtained by weighted summation, where the weight coefficient of the average correlation strength is set to 0.6 and the weight coefficient of the spatial distribution entropy is set to 0.4 to balance the influence of correlation degree and distribution concentration. A topological compactness threshold is also set. This threshold is determined by statistically analyzing the density distribution of all coherent sensitive clusters, typically taking the value corresponding to the upper quartile of the distribution. Clusters with topological density exceeding the threshold are considered. The coherent sensitive clusters were identified as specific sensitive channels, and the signal dimensions corresponding to these channels can collaboratively and stably capture the specific patterns of natural rubber aging.

[0036] In a preferred embodiment of the present invention, step 4 is further included: extracting the divergence enhancement signal set and the full signal set from the labeled natural rubber aging data using a specific sensitive channel, and using the two to train the first evaluation unit and the second discrimination unit in tandem; wherein, the first evaluation unit fits the divergence enhancement signal set, the second discrimination unit makes a pattern discrimination based on the full signal set, and the two establish a correlation through a coupled loss function. The steps of collaborative training for the first evaluation unit and the second discrimination unit include: Step 41: Modulate the entire signal set based on the specific sensitive channel to generate a modulated signal, and project the modulated signal onto the difference subspace to obtain the bifurcation enhancement signal set. The specific operation is as follows: The labeled natural rubber aging data contains aging signal information across all dimensions, forming a complete signal set. This set includes both effective signals reflecting the specific aging patterns of natural rubber and redundant noise signals unrelated to the aging patterns. The specific sensitive channel, as a set of high-quality signal dimensions filtered by topological density, has the ability to accurately capture the differences in aging patterns between the two types of rubber. Therefore, the complete signal set is modulated based on this channel to enhance the effective signal components and suppress redundant noise. The modulation process assigns weights to each signal dimension based on the topological density value corresponding to the specific sensitive channel. The higher the topological density value, the larger the weight coefficient of the corresponding signal dimension. For example, the weight coefficient of the channel with a topological density value of 0.8 is set to 0.9, while the weight coefficient of the signal dimensions not included in the sensitive channel is set to 0.2. This weighted processing generates a modulated signal that highlights specific characteristics. The difference subspace is a feature space constructed based on the deviation vector of the aforementioned patterns. Its dimensional direction is consistent with the divergence direction of the two types of rubber aging patterns. Projecting the modulated signal onto this subspace can further amplify the signal components related to the divergence of aging patterns. The final projected signal set is the divergence-enhanced signal set, which can accurately focus on the specific aging characteristics of natural rubber.

[0037] Step 42: Based on the statistical summary of the modulated signal, dynamically generate the convolutional kernel weight set for the first evaluation unit and the second discrimination unit. The specific operation is as follows: The first evaluation unit and the second discrimination unit are responsible for fitting the divergent enhanced signal and discriminating the pattern of the entire signal, respectively. The convolutional kernel weights of the two units directly determine the accuracy of feature extraction. The statistical summary of the modulated signal can intuitively reflect the distribution pattern of specific aging features. Therefore, dynamically generating the convolutional kernel weight set based on this statistical summary can achieve precise matching of weights to task requirements. First, the statistical summary of the modulated signal is calculated, covering key statistical quantities such as the mean, variance, peak density, and correlation coefficient of adjacent time-series signals for each signal dimension. For example, the mean μ=0.5 and variance σ=0.15 for a certain modulated signal dimension are calculated. The peak density ρ=0.3. Subsequently, a mapping rule between statistical summary and convolutional kernel weights is established. For the convolutional kernel of the first evaluation unit, the focus is on the correlation between variance and peak density. The larger the variance and the higher the peak density of the signal features, the greater the weight of the convolutional kernel, so as to strengthen the extraction of divergent features. For the convolutional kernel of the second discrimination unit, the focus is on the correlation between mean and correlation coefficient to ensure that the overall regularity of the whole signal can be captured. Through this mapping rule, a set of convolutional kernel weights that are adapted to the different task requirements of the two units is dynamically generated to avoid feature extraction bias caused by using fixed weights and improve the effectiveness of initial feature extraction of the unit.

[0038] Step 43: Based on the fitting loss gradient of the first evaluation unit and the discrimination loss gradient of the second discrimination unit, dynamically adjust the learning rates of both parameters, and exchange feature summary information during the adjustment process. The specific operations are as follows: The fitting loss gradient of the first evaluation unit reflects the degree of deviation between its fitting result for the bifurcation enhancement signal and the true labeled value, while the discrimination loss gradient of the second discrimination unit reflects the accuracy of its discrimination of the aging law of the entire signal. Dynamically adjusting the learning rates of both can achieve synergistic optimization of the training process, avoiding excessively fast or slow convergence of a single unit. The loss value of the first evaluation unit is the fitting loss value, and its calculation logic revolves around the fitting deviation of the bifurcation enhancement signal: First, the true labeled value of the bifurcation enhancement signal set is identified. This labeled value is a quantitative indicator obtained through precise detection in the natural rubber aging experiment, such as the measured value of crosslinking density and the measured value of tensile strength decay rate at the corresponding aging time point; then, the true labeled value of the bifurcation enhancement signal set is obtained. The first evaluation unit outputs fitted values ​​for the bifurcation enhancement signal set. It calculates the difference between the fitted value and the corresponding true label value for each sample, and squares this difference to obtain the squared fitting deviation for a single sample. Finally, it takes the arithmetic mean of the squared fitting deviations for all training samples, which is the final fitting loss value. This mean square error calculation logic effectively amplifies the impact of large fitting deviations, driving the evaluation unit to accurately fit the specific aging characteristics in the bifurcation enhancement signal. The fitting loss value is positively correlated with the fitting deviation; the larger the difference, the higher the fitting loss value. The second discrimination unit's loss value is the discrimination loss value. The calculation logic focuses on the matching accuracy of the aging pattern discrimination across the entire signal. First, the entire signal set is labeled with true aging pattern labels. Label types include aging stages (early, middle, late), dominant aging mechanisms (oxidation-dominated, amorphous region ordering-dominated), and discrete category labels. The second discrimination unit performs pattern discrimination on the entire signal set and outputs the probability distribution of each sample corresponding to different aging pattern labels. Based on this probability distribution and the one-hot encoding of the true aging pattern labels (i.e., the category corresponding to the true label is encoded as 1, and the other categories as 0), the cross-entropy is calculated. Specifically, the logarithm of the predicted probability for the category corresponding to the true label is taken as the negative value, and then the arithmetic mean of this negative value is taken for all training samples to obtain the discrimination loss value. This cross-entropy form... The computational logic can accurately quantify the deviation between the predicted probability distribution and the true label. The lower the matching degree between the discrimination result and the true label, the smaller the predicted probability of the corresponding category, the larger the negative logarithm, and the higher the discrimination loss value. During the adjustment process, the absolute values ​​of the loss gradients of the two units are calculated separately. When the absolute value of the fitting loss gradient of the first evaluation unit is greater than the absolute value of the discrimination loss gradient of the second discrimination unit, the learning rate of the first evaluation unit is increased and the learning rate of the second discrimination unit is decreased. For example, when the absolute value of the fitting loss gradient is 0.6 and the absolute value of the discrimination loss gradient is 0.3, the learning rate of the first evaluation unit is adjusted from 0.001 to 0.0015, and the learning rate of the second discrimination unit is adjusted to 0.0008; Conversely, the adjustment is reversed. Simultaneously, during the learning rate adjustment process, the feature summary information of the two units is exchanged. The first evaluation unit transmits the extracted divergent feature summary to the second discriminant unit to assist it in accurately identifying the full signal features related to specific patterns. The second discriminant unit then feeds back the overall pattern feature summary of the full signal to the first evaluation unit, guiding it to optimize the fitting direction. The coupling loss function specifically correlates the loss values ​​of the two units through a weighted summation. The weight coefficient of the fitting loss value is set as α, and the weight coefficient of the discriminant loss value is set as β, with α + β = 1. The weight coefficients can be dynamically adjusted according to the training stage. For example, in the early stage of training, α = 0.6 and β = 0.4, focusing on the fitting accuracy of the first evaluation unit; in the later stage of training, α = 0.4 and β = 0.6, focusing on the discrimination accuracy of the second discriminant unit. This weighted summation yields the overall coupling loss value, ensuring that the learning rate adjustment and feature exchange always revolve around a common training objective, achieving collaborative convergence.

[0039] Step 44: Perform bidirectional attention calibration on the high-level feature maps of the first evaluation unit and the second discrimination unit. The specific operation is as follows: The high-level feature maps of the first evaluation unit and the second discriminant unit focus on divergent features and overall regularity features, respectively, and each has its own redundant feature regions. Performing bidirectional attention calibration enables complementary optimization of the two types of feature maps, improving the accuracy of feature representation. First, the high-level feature maps of the two units are extracted, and the cross-correlation coefficients of each region in both feature maps are calculated to locate feature regions with collaborative significance. For example, if the cross-correlation coefficient between a region in the high-level feature map of the first evaluation unit and a region in the feature map of the second discriminant unit is 0.75, it is determined to be a collaborative feature region. Then, a bidirectional attention weight matrix is ​​constructed, assigning attention weights to each feature region based on the cross-correlation coefficients. Collaborative feature regions with higher cross-correlation coefficients have greater weights, while redundant feature regions (those with cross-correlation coefficients below 0.2) are assigned lower weights. Based on this attention weight matrix, the high-level feature maps of the two units are calibrated and adjusted, enhancing the signal strength of high-weight collaborative feature regions and suppressing signal interference from low-weight redundant regions. Through this bidirectional calibration, the high-level feature map of the first evaluation unit incorporates the overall regularity information of the entire signal, while the high-level feature map of the second discriminant unit strengthens specific divergent feature information, achieving complementary optimization of feature representation and further improving the effectiveness of collaborative training.

[0040] In a preferred embodiment of the present invention, step 5 is further included: after the collaborative training reaches the predetermined convergence condition, the second discrimination unit is solidified as a pattern identification unit, and a conditional evaluation unit is configured based on the output of the pattern identification unit. The configuration of the regularity identification unit and the conditional evaluation unit also includes the following steps: Step 51: Perform sparsity pruning on the second discrimination unit. The specific operation is as follows: Although the second discriminant unit, after collaborative training, possesses accurate full-signal pattern discrimination capabilities, it contains some redundant parameters. These parameters not only increase the computational cost of subsequent deployments but may also introduce irrelevant noise that interferes with the accuracy of pattern identification. Therefore, sparsity pruning is necessary. The pruning logic is to retain the core parameters that play a crucial role in pattern discrimination and remove redundant parameters with extremely low contribution. Specifically, the absolute value of each parameter's weight is used as the basis for contribution evaluation. First, the absolute value of the weights of all parameters in the second discriminant unit is calculated. Then, a pruning threshold is set. This threshold is determined by statistically analyzing the distribution of the absolute values ​​of parameter weights, typically taking the value corresponding to the lower 10th quantile of the distribution. For example, if the lower 10th quantile of the absolute value of the parameter weights is 0.02, then the pruning threshold is set to 0.02. Parameters with absolute weights below this threshold are directly set to zero to achieve pruning. At the same time, the performance of the pruned unit is verified to ensure that the decrease in discrimination accuracy on the validation set does not exceed 3%, thus ensuring that the pruned unit still retains its core discrimination capabilities.

[0041] Step 52: Construct a nonlinear mapping rule from the decision space of the second discriminant unit after sparse pruning to the parameter subspace of the first evaluation unit. The specific operations are as follows: The pre-stored basic parameter block is a set of parameters pre-trained based on different aging scenarios, such as different temperatures and different aging stages. It covers parameter templates adapted to various natural rubber aging patterns. Based on a nonlinear mapping rule, these templates are weighted and combined to dynamically generate the core parameters of the conditional evaluation unit. This nonlinear mapping rule is not an abstract construction, but a concrete construction based on historical data accumulated during collaborative training. The specific logic is as follows: First, all decision outputs of the second discriminator unit in the collaborative training stage are extracted, i.e., the probability distribution of aging patterns of different samples, and the core parameter configuration of the first evaluation unit at the corresponding time, forming a set of decision output and parameter configuration sample pairs. Then, the carrier of the mapping rule is determined to be a Gaussian kernel function. This kernel function is used to measure the similarity between the current decision output and the historical decision outputs in the sample pair. The bandwidth parameter of the Gaussian kernel is... The standard deviation of historical decision outputs in a sample pair is used to determine the standard deviation. For example, if the calculated standard deviation of historical decision outputs is 0.12, then the standard deviation is set to... =0.15, to ensure that the sensitivity of similarity calculation is adapted to the distribution characteristics of the decision space; based on this Gaussian kernel function, the similarity between the current decision output and each historical decision output is calculated, and then the similarity is weighted and summed with the corresponding historical parameter configuration to obtain the initial mapping result. Finally, the mapping result is calibrated by the least squares method to ensure that the mapping error, that is, the deviation between the mapped parameter configuration and the historical true parameter configuration, is less than a preset threshold, such as setting the deviation threshold to 0.05, thus completing the construction of the nonlinear mapping rule; based on the completed mapping rule, the fit between each pre-stored basic parameter block and the current decision output is calculated: the standard parameter configuration corresponding to the basic parameter block is input into the mapping rule to obtain the mapped standard decision output, and then the fit is calculated. The Gaussian kernel similarity between the standard decision output and the current decision output is called the fit. The higher the fit, the larger the weight coefficient of the basic parameter block. For example, if the fit between a basic parameter block and the decision output is 0.85, the corresponding weight coefficient is set to 0.7, while the weight coefficient of a basic parameter block with a fit of 0.3 is set to 0.1. Then, all basic parameter blocks are weighted and summed according to their corresponding weight coefficients to obtain the parameter set that fits the current decision output, which is the parameter subset of the conditional evaluation unit. This process avoids the tedious process of retraining parameters through the flexible combination of basic parameter blocks, while ensuring that the generated parameter subset can accurately match the aging pattern identified by the pattern recognition unit, thereby improving the targeting and efficiency of the evaluation unit.

[0042] Step 53: Based on the nonlinear mapping rule, the pre-stored basic parameter blocks are weighted and combined to dynamically generate a subset of parameters for the conditional evaluation unit. The specific operation is as follows: Different aging patterns correspond to different parameter subsets. Directly switching parameter subsets may cause abrupt changes in the output of the conditional evaluation unit, affecting the stability of the evaluation results. Therefore, a softening function based on decision space distance is introduced in the weighted combination process to establish a smooth transition mechanism between parameter subsets. Specifically, the S-shaped softening function adopts the Logistic function, a commonly used continuous smoothing function in this field, which can map the input distance value to a continuous interval of 0 to 1. Its calculation logic is as follows: first, the distance value between the current decision output and the standard decision output corresponding to each pre-stored basic parameter block in the decision space is normalized, compressing the distance value to the range of 0 to 1. The normalization method is to divide the original distance value by the maximum distance between all standard decision outputs in the decision space. For example, if the maximum distance is 0.5 and a certain original distance value is 0.2, then the normalized distance value... The initial distance is set to 0.4. The normalized distance value is then input into the Logistic function, and the function output becomes the initial weight adjustment coefficient. To ensure the adjustment coefficient range matches the weight correction requirements, the function output value is further scaled using the logic of 0.1 + 0.9 × the function output value, ultimately obtaining a weight adjustment coefficient ranging from 0.1 to 1. The specific characteristics of the Logistic function are controlled by the slope coefficient k and the offset b. The slope coefficient k determines the steepness of the function curve; the larger the k value, the steeper the curve, and the more sensitive the adjustment coefficient is to changes in the distance value. In this scheme, k=2 is set, and experimental results show... This value balances sensitivity and smoothness; the offset b determines the midpoint of the function curve. In this scheme, b=0.5 is set to ensure that when the normalized distance value is 0.5, the adjustment coefficient is 0.55, achieving a symmetrical distribution in the transition interval. In specific implementation, the original distance values ​​between the current decision output and each standard decision output are first calculated using Euclidean distance calculation, i.e., the square root of the sum of the squares of the differences in the corresponding dimensions of the two, after normalization, is input into the Logistic function to obtain the initial adjustment coefficient. Then, the final adjustment coefficient is obtained through scaling. The smaller the distance value, the more pronounced the aging trend. The closer the samples are, the closer the adjustment coefficient is to 1; the greater the distance, the closer the adjustment coefficient is to 0.1. This adjustment coefficient is then used to correct the initial weights of the basic parameter blocks. The corrected weights = initial weights × adjustment coefficients. The corrected weight coefficients enable a gradual change in the weights of the parameter subsets corresponding to adjacent aging patterns, rather than an abrupt change. This allows the parameter subsets generated by the weighted combination to transition smoothly with the changes in the decision output. This smooth transition mechanism ensures that the conditional evaluation unit has continuous and stable output results when facing samples with similar aging patterns, avoiding evaluation bias caused by parameter abrupt changes.

[0043] Step 54: In the weighted combination process, a softening function based on the distance in the decision space is introduced to establish a smooth transition mechanism between parameter subsets. The specific operation is as follows: Different aging patterns correspond to different parameter subsets. Directly switching parameter subsets may cause abrupt changes in the output of the conditional evaluation unit, affecting the stability of the evaluation results. Therefore, a softening function based on decision space distance is introduced in the weighted combination process to establish a smooth transition mechanism between parameter subsets. First, the distance between the current decision output and the standard decision output corresponding to each pre-stored basic parameter block in the decision space is calculated. The smaller the distance value, the more similar the aging patterns are. Then, this distance value is input into the softening function. For example, an S-shaped softening function is used. The smaller the distance value, the closer the weight adjustment coefficient of the softening function output is to 1. The larger the distance value, the closer the weight adjustment coefficient gradually approaches 0.1. The initial weights of the basic parameter blocks are corrected through this adjustment coefficient. The corrected weight coefficient can achieve a gradual change in the weights of parameter subsets corresponding to adjacent aging patterns, rather than an abrupt change. This allows the parameter subsets generated by the weighted combination to transition smoothly with the change of decision output. This smooth transition mechanism ensures that the output results of the conditional evaluation unit have continuity and stability when facing samples with similar aging patterns, avoiding evaluation bias caused by parameter abrupt changes.

[0044] In a preferred embodiment of the present invention, step 6 is further included: for the sample to be tested, its divergence enhancement signal and full signal are obtained through a specific sensitive channel; the full signal is input into the regularity discrimination unit to obtain a discrimination score, and the parameter state of the conditional evaluation unit is adjusted accordingly; and the divergence enhancement signal is input into the adjusted conditional evaluation unit. The evaluation of the sample to be tested also includes the following steps: Step 61: Perform phase synchronization processing on the full signal and the divergence enhancement signal of the sample to be tested to generate a phase-synchronized divergence enhancement signal. The specific operation is as follows: Although both the full signal and the divergent enhanced signal of the test sample originate from the same aging process, they may have phase deviations due to factors such as the response delay of the detection instrument and differences in signal transmission paths. This deviation can lead to mismatches in key aging information during subsequent feature extraction, affecting the evaluation accuracy. Therefore, phase synchronization processing is required. The logic of phase synchronization is to use the full signal as a reference to calibrate the phase of the divergent enhanced signal, ensuring that the key feature points of the two are accurately aligned in the time dimension. First, the cross-correlation function of the full signal and the divergent enhanced signal is calculated. The phase offset of the divergent enhanced signal relative to the full signal is determined by finding the peak position of the cross-correlation function. For example, if the time difference corresponding to the peak of the cross-correlation function is calculated to be 0.2 seconds, the phase angle corresponding to the phase offset of 0.2 seconds is determined. Then, the phase of the divergent enhanced signal is adjusted according to this phase offset by shifting all the time-series data of the divergent enhanced signal along the time axis by the corresponding offset, so that the key feature points such as the peak and valley values ​​of the two coincide on the same time scale. Finally, a phase-synchronized divergent enhanced signal is generated. This synchronization processing ensures the consistency of the aging information carried by the two types of signals in the time dimension.

[0045] Step 62: Utilize the feature extraction kernel of the conditional evaluation unit for dynamic configuration of the entire signal after phase synchronization. The specific operation is as follows: The feature extraction kernel of the conditional evaluation unit directly determines the ability to extract specific aging features from the divergent enhanced signal. The phase-synchronized full signal eliminates phase deviation and can more accurately reflect the true aging pattern of the sample under test. Therefore, by dynamically configuring the feature extraction kernel using this full signal, precise adaptation between the extraction kernel and the aging pattern of the sample under test can be achieved. First, key statistical features of the phase-synchronized full signal are extracted, including the frequency distribution, temporal variation slope, and time interval of characteristic peaks in each signal dimension. For example, the main frequency f=2Hz and the temporal variation slope of a certain full signal dimension are extracted. =0.3. Combining the discrimination score output by the pattern recognition unit, which characterizes the degree of matching between the test sample and each standard aging pattern, the adjustment amount of each parameter of the feature extraction kernel is calculated using the nonlinear mapping rule constructed earlier. Specific parameters of the feature extraction kernel include the convolution kernel size, stride, and initial weight values. For the feature extraction kernel size corresponding to the dominant frequency, it is adjusted according to the inverse relationship between frequency and kernel size; the higher the dominant frequency, the smaller the kernel size. The initial weight values ​​are determined based on the synergistic effect of the discrimination score and statistical features, ensuring that the configured feature extraction kernel can accurately capture specific features matching the aging pattern of the test sample, thus improving the targeting and effectiveness of feature extraction.

[0046] Step 63: Feed the phase-synchronized divergence enhancement signal into the configured conditional evaluation unit, and calculate the sub-band attention mask based on the hybrid weights corresponding to the parameter subsets to modulate the intermediate feature map. The specific operation is as follows: The phase-synchronized divergence enhancement signal focuses on the specific aging characteristics of natural rubber. Feeding this signal into a dynamically configured conditional evaluation unit enables in-depth extraction and fitting evaluation of these specific characteristics. The hybrid weights corresponding to the parameter subsets are key weight information adapted to the aging patterns of the current test sample. Based on these weights, a sub-band attention mask is calculated, which further enhances the sub-band features that significantly contribute to the evaluation and suppresses irrelevant noise interference. First, the phase-synchronized divergence enhancement signal is decomposed into multiple continuous sub-bands, each corresponding to aging characteristic information within a different frequency range, such as 0 to 1 Hz, 1 to 3 Hz, and 3 to 5 Hz. The system calculates the attention weight for each sub-band based on the mixed weight of the parameter subset. The higher the mixed weight of the core parameter, the larger the attention weight coefficient of the corresponding sub-band. For example, if the mixed weight of a core parameter is 0.8, the attention weight coefficient of its corresponding sub-band is set to 0.9, while the weight coefficient of the sub-band corresponding to the core parameter with a mixed weight of 0.2 is set to 0.3. Based on this attention weight, a sub-band attention mask is generated and applied to the intermediate feature map extracted by the conditional evaluation unit. By weighting and strengthening the feature regions corresponding to high-weight sub-bands and weakening low-weight regions, the intermediate feature map can more accurately focus on key aging features.

[0047] Step 64: Perform consistency verification based on adversarial perturbation on the preliminary evaluation results of the conditional evaluation unit, and output the final evaluation results based on the verification results. The specific operations are as follows: The preliminary evaluation results of the conditional evaluation unit may be affected by factors such as random noise and outliers in the test sample, resulting in potential bias. Therefore, consistency verification based on adversarial perturbation is required to ensure the reliability and stability of the evaluation results. The adversarial perturbation is a small perturbation generated based on the loss gradient of the preliminary evaluation results. Its intensity is controlled within the range that does not change the true aging state of the sample. For example, the perturbation intensity threshold is set to 0.05 to ensure that the signal after perturbation still conforms to the physicochemical laws of natural rubber aging. First, the loss gradient corresponding to the preliminary evaluation results is calculated, and an adversarial perturbation is generated along the gradient ascending direction. This perturbation is then applied to the input phase-synchronized divergence enhancement signal to obtain the perturbated signal. The perturbed signal is then input into the configured conditional evaluation unit to obtain the perturbed evaluation result. The deviation between the preliminary evaluation result and the perturbed evaluation result is calculated, and a consistency verification threshold is set. This threshold is determined by statistically analyzing the deviation distribution of a large number of validation set samples. For example, the value corresponding to the upper 20th percentile of the validation set deviation distribution is taken as 0.08. If the deviation value is less than this threshold, it indicates that the preliminary evaluation result is not sensitive to small perturbations and has good consistency. The preliminary evaluation result is then output as the final evaluation result. If the deviation value is greater than the threshold, the parameter configuration of the conditional evaluation unit is readjusted, and the evaluation and verification process is executed again until the deviation value meets the requirements and the final evaluation result is output.

[0048] Example 2: Please see Figure 2 Based on Example 1, this embodiment provides a machine learning-based natural rubber aging state assessment system, comprising: The function family generation module is used to construct the first parameterized function family corresponding to synthetic rubber and the second parameterized function family corresponding to natural rubber, respectively. The parameter optimization module is used to perform parameter optimization on the first parameterized function family using the aging data sequence of natural rubber, and generate a regular deviation vector; The channel determination module analyzes the dimensionality sensitivity of the original signal space based on the regular deviation vector to determine the specific sensitive channels; The collaborative training module is used to extract the divergence enhancement signal set and the full signal set from the labeled natural rubber aging data using a specific sensitive channel, and to use the two to collaboratively train the first evaluation unit and the second discrimination unit. The first evaluation unit fits the divergence enhancement signal set, and the second discrimination unit makes regularity judgments based on the full signal set. The two are associated through a coupled loss function. The unit configuration module is used to solidify the second discriminant unit as a pattern identification unit after the collaborative training reaches the predetermined convergence condition, and to configure the conditional evaluation unit based on the output of the pattern identification unit. The state evaluation module is used to acquire the divergence enhancement signal and the full signal of the sample under test through a specific sensitive channel; the full signal is input into the regularity discrimination unit to obtain the discrimination score, and the parameter state of the conditional evaluation unit is adjusted accordingly, and the divergence enhancement signal is input into the adjusted conditional evaluation unit.

[0049] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.

Claims

1. A machine learning-based method for assessing the aging state of natural rubber, characterized in that, The steps include: Step 1, constructing the first family of parameterized functions for synthetic rubber and the second family of parameterized functions for natural rubber respectively; Step 2: Use the obtained natural rubber aging data sequence to perform parameter optimization on the first parameterized function family to generate a regular deviation vector; Step 3: Based on the deviation vector, analyze the dimensionality sensitivity of the original signal space to determine the specific sensitive channels; Step 4: Use specific sensitive channels to extract the divergence enhancement signal set and the full signal set from the labeled natural rubber aging data, and use the two to train the first evaluation unit and the second discrimination unit together; wherein, the first evaluation unit fits the divergence enhancement signal set, the second discrimination unit makes regularity judgments based on the full signal set, and the two establish a correlation through a coupled loss function; Step 5: After the collaborative training reaches the predetermined convergence condition, solidify the second discrimination unit as the pattern identification unit, and configure the conditional evaluation unit based on the output of the pattern identification unit. Step 6: For the sample to be tested, obtain its divergence enhancement signal and full signal through a specific sensitive channel; input the full signal into the regularity discrimination unit to obtain the discrimination score, adjust the parameter state of the conditional evaluation unit accordingly, and input the divergence enhancement signal into the adjusted conditional evaluation unit.

2. The method for assessing the aging state of natural rubber based on machine learning according to claim 1, characterized in that, The construction of a family of parameterized functions includes: Step 11: Analyze the aging chain reaction pathways of synthetic rubber and natural rubber respectively, and extract the first set of control variables for synthetic rubber and the second set of control variables for natural rubber. The first set of control variables includes the free radical generation rate and the chain termination efficiency, and the second set of control variables also includes the degradation rate of the activity of amorphous region chain segments. Step 12: Based on the first set of control variables and the second set of control variables, construct a unidirectional saturated coupling network for synthetic rubber and a triangular asymmetric coupling network for natural rubber. Step 13: Instantiate the unidirectional saturated coupled network and the triangular asymmetric coupled network into a dynamic system, assign parameterized nonlinear functions to the coupling edges, and define macroscopic response values ​​based on the state variables of the dynamic system to generate the first family of parameterized functions and the second family of parameterized functions.

3. The method for assessing the aging state of natural rubber based on machine learning according to claim 2, characterized in that, The generation of the regular deviation vector includes: Step 21: Perform phase space reconstruction on the natural rubber aging data sequence to generate empirical state trajectories; Step 22: Calculate the Hausdorff distance spectrum between the empirical state trajectory and the simulated state trajectory generated by the first family of parameterized functions. Step 23: Perform geodesic-based parameter perturbation on the coupled network of synthetic rubber based on the Hausdorff distance spectrum; Step 24: Extract the edge weight changes of the coupled network that exceed a predetermined threshold before and after the disturbance, and arrange them in topological order to generate a regular deviation vector.

4. The method for assessing the aging state of natural rubber based on machine learning according to claim 3, characterized in that, The identification of specific sensitive channels includes: Step 31: Decode the regular deviation vector into a virtual physical field excitation mode; Step 32: Based on the virtual physical field excitation mode, simulate the disturbance response of each dimension of the original signal space to generate a full-dimensional disturbance response spectrum; Step 33: Perform mode decomposition on the full-dimensional perturbation response spectrum and extract coherent sensitive clusters; Step 34: Calculate the topological density of the coherent sensitive cluster in the original signal space, and determine the specific sensitive channel based on the topological density.

5. The method for assessing the aging state of natural rubber based on machine learning according to claim 4, characterized in that, The steps for collaboratively training the first evaluation unit and the second discriminant unit include: Step 41: Modulate the entire signal set based on the specific sensitive channel to generate a modulated signal, and project the modulated signal onto the difference subspace to obtain the divergence-enhanced signal set; Step 42: Based on the statistical summary of the modulated signal, dynamically generate the convolutional kernel weight set for the first evaluation unit and the second discrimination unit.

6. The method for assessing the aging state of natural rubber based on machine learning according to claim 5, characterized in that, The collaborative training first evaluation unit and second discrimination unit also include: Step 43: Based on the fitting loss gradient of the first evaluation unit and the discrimination loss gradient of the second discrimination unit, dynamically adjust the learning rate of the parameters of both units, and exchange feature summary information during the adjustment process. Step 44: Perform bidirectional attention calibration on the high-level feature maps of the first evaluation unit and the second discrimination unit.

7. The method for assessing the aging state of natural rubber based on machine learning according to claim 6, characterized in that, The configuration of the regularity identification unit and the conditional evaluation unit includes: Step 51: Perform sparsity pruning on the second discrimination unit; Step 52: Construct a nonlinear mapping rule from the decision space of the second discrimination unit to the parameter subspace of the first evaluation unit after sparse pruning. Step 53: Based on the nonlinear mapping rule, the pre-stored basic parameter blocks are weighted and combined to dynamically generate a parameter subset of the conditional evaluation unit; Step 54: In the weighted combination process, a softening function based on the decision space distance is introduced to establish a smooth transition mechanism between parameter subsets.

8. The method for assessing the aging state of natural rubber based on machine learning according to claim 7, characterized in that, The evaluation of the sample to be tested includes: Step 61: Perform phase synchronization processing on the full signal and the divergence enhancement signal of the sample to be tested to generate a phase-synchronized divergence enhancement signal; Step 62: Dynamically configure the feature extraction kernel of the conditional evaluation unit using the full signal after phase synchronization.

9. The method for assessing the aging state of natural rubber based on machine learning according to claim 8, characterized in that, The evaluation of the sample to be tested also includes: Step 63: Feed the phase-synchronized divergence enhancement signal into the configured conditional evaluation unit, and calculate the sub-band attention mask based on the hybrid weights corresponding to the parameter subsets to modulate the intermediate feature map. Step 64: Perform consistency verification based on adversarial perturbation on the preliminary evaluation results of the conditional evaluation unit, and output the final evaluation results based on the verification results.

10. A machine learning-based natural rubber aging state assessment system, applied to the machine learning-based natural rubber aging state assessment method according to any one of claims 1-9, characterized in that, include: The function family generation module is used to construct the first parameterized function family corresponding to synthetic rubber and the second parameterized function family corresponding to natural rubber, respectively. The parameter optimization module is used to perform parameter optimization on the first parameterized function family using the aging data sequence of natural rubber, and generate a regular deviation vector; The channel determination module analyzes the dimensionality sensitivity of the original signal space based on the regular deviation vector to determine the specific sensitive channels; The collaborative training module is used to extract the divergence enhancement signal set and the full signal set from the labeled natural rubber aging data using a specific sensitive channel, and to use the two to collaboratively train the first evaluation unit and the second discrimination unit. The first evaluation unit fits the divergence enhancement signal set, and the second discrimination unit makes regularity judgments based on the full signal set. The two are associated through a coupled loss function. The unit configuration module is used to solidify the second discriminant unit as a pattern identification unit after the collaborative training reaches the predetermined convergence condition, and to configure the conditional evaluation unit based on the output of the pattern identification unit. The state evaluation module is used to acquire the divergence enhancement signal and the full signal of the sample under test through a specific sensitive channel; the full signal is input into the regularity discrimination unit to obtain the discrimination score, and the parameter state of the conditional evaluation unit is adjusted accordingly, and the divergence enhancement signal is input into the adjusted conditional evaluation unit.

Images