A method for predicting the life of a petg bottle against cracking

Abstract

The present application belongs to the technical field of polymer material life evaluation, and relates to a PETG bottle anti-cracking life prediction method. The method comprises the following steps: acquiring multi-source environment time sequence data of the PETG bottle in the service process, performing semantic segmentation processing based on the environmental response characteristics of the PETG material to generate a stress fragment set with semantic labels; constructing a stress correlation graph structure, mapping the stress fragments into graph nodes and establishing a weighted directed edge; fusing and encoding the semantic labels and node attributes to generate a feature matrix; dynamically adjusting the adjacent information injection intensity according to the material response priority in the graph topology propagation process; applying time sequence constraints to the propagation feature matrix and extracting cross-fragment dependent features to output a health attenuation trend vector; combining the life prediction model parameters to calculate the remaining useful life and generate an evaluation result containing a risk level and a confidence degree. The present application improves the accuracy of the PETG bottle anti-cracking life prediction under complex dynamic environments through semantic segmentation and differentiated feature fusion.

Description

Technical Field

[0001] This invention relates to the field of polymer material life assessment technology, and in particular to a method for predicting the crack resistance life of PETG bottles. Background Technology

[0002] Existing methods for predicting the service life of PETG bottles typically employ cumulative damage models, degradation trajectory fitting, or general deep learning models to process environmental time-series stress data. These methods establish a mapping relationship between material health status and service environment by extracting statistical characteristics or time-series trends of physical quantities such as temperature, humidity, and vibration, in order to support the service life assessment needs in logistics, warehousing, and transportation scenarios.

[0003] However, PETG materials are subjected to diverse environmental stresses during service, including continuous high-temperature exposure, moisture absorption and plasticization in humid environments, condensation phase transformation, mechanical impact, and ultraviolet radiation. Different types of stresses have significantly different mechanisms for promoting crack initiation and propagation, and the damage accumulation path under multi-physics coupling is even more complex. Existing methods often uniformly convert various environmental stresses into numerical feature input models, failing to distinguish the differentiated contributions of different stress types in the material failure process. This weakens key damaging stress information and results in insufficient prediction accuracy.

[0004] In addition, some solutions attempt to capture more complex stress-damage mapping relationships by increasing model depth, but this leads to problems such as increased model size and inference latency, making it difficult to achieve real-time lifetime assessment on resource-constrained edge devices. Summary of the Invention

[0005] This application provides a method for predicting the crack resistance life of PETG bottles, aiming to solve the problem of insufficient prediction accuracy caused by the difficulty in distinguishing the differentiated contributions of different types of environmental stress to the damage of PETG materials in existing methods.

[0006] This application provides a method for predicting the crack resistance life of PETG bottles, specifically including:

[0007] Multi-source environmental time-series data during the service life of PETG bottles are acquired. Based on the environmental response characteristics of PETG material, the multi-source environmental time-series data is semantically segmented to generate a set of stress segments with semantic tags.

[0008] Based on the set of stress segments, a stress correlation graph structure is constructed. Each stress segment is mapped as a graph node and given node attributes that include stress statistical features and inter-segment correlation features. Directed edges are established between segments that satisfy temporal proximity and physical coupling, and the stress mutation degree between segments is quantified to generate a stress correlation graph object with topological connection relationship.

[0009] The semantic labels and node attributes in the stress correlation graph object are fused and encoded to generate a feature matrix representing the initial state of each stress segment.

[0010] Based on the feature matrix, a graph topology propagation operation is performed. During the neighborhood information aggregation process, the intensity of adjacency information injection is dynamically adjusted according to the material response priority corresponding to the semantic type of each fragment, thereby generating a propagation feature matrix that has been adaptively weighted and optimized.

[0011] A temporal order constraint is applied to the propagation feature matrix, and the segments are processed in an orderly manner according to their occurrence order. Cross-segment dependency features are extracted, and a temporal feature vector characterizing the material health degradation trend is output.

[0012] Based on the time-series feature vectors and the parameters of the life prediction model obtained through training, the remaining life of PETG bottles under the current environmental stress is calculated, and an assessment result including risk level and confidence level is generated.

[0013] The method for predicting the crack resistance life of PETG bottles provided in this application has the following beneficial effects:

[0014] By analyzing multi-source environmental time-series data on temperature, humidity, and vibration acquired during the service life of PETG bottles, semantic segmentation is performed based on the environmental response characteristics of PETG material itself. This results in a set of stress segments with physical semantic labels, effectively preserving the temporal integrity and contextual information of key damaging stress events. A stress correlation graph structure is then constructed, mapping each stress segment to a graph node and assigning stress statistical features and inter-segment correlation features. Weighted directed edges are established between segments that satisfy temporal proximity and physical coupling, forming a graph object with topological connections. This achieves a structured expression of the physical causal relationships between different stress events. Furthermore, the semantic labels and node attributes in the graph object are fused and encoded to generate a feature matrix. During graph topology propagation, the intensity of adjacency information injection is dynamically adjusted according to the material response priority corresponding to the semantic type of each segment. This allows the model to adaptively filter key information and suppress irrelevant interference based on the differentiated contribution of stress type to material damage. A temporal order constraint is then applied to the propagated feature matrix, and cross-segment dependency features are extracted according to the segment sequence. Finally, the remaining service life is calculated by combining the parameters of the life prediction model, and an assessment result including risk level and confidence level is output.

[0015] This solution forms a complete technical loop from data semantic reconstruction, physical correlation modeling, differential feature fusion to temporal dependency mining. It effectively solves the problem of insufficient prediction accuracy caused by the difficulty in distinguishing the differential contribution of different types of environmental stress to the damage of PETG materials in existing methods, and improves the accuracy and reliability of crack resistance life assessment under complex dynamic service environments. Attached Figure Description

[0016] Figure 1 This is the main flowchart of a method for predicting the crack resistance life of PETG bottles.

[0017] Figure 2 This is a flowchart of the decision path for ultraviolet superposition and composite disturbance determination in a multiphysics coupling analysis of a PETG bottle crack resistance life prediction method.

[0018] Figure 3 This is a sub-flowchart of a method for predicting the crack resistance life of PETG bottles. Detailed Implementation

[0019] 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.

[0020] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0021] It should be noted that PETG (Polyethylene Terephthalate Glycol-modified) is a transparent, non-crystalline copolyester material. PETG is widely used in the manufacture of various bottles such as cosmetic bottles, laboratory culture medium bottles, and disposable sterile containers.

[0022] like Figure 1 As shown, this application provides a method for predicting the crack resistance life of PETG bottles, specifically including:

[0023] S1: Acquire multi-source environmental time-series data during the service of PETG bottles. The data includes temperature, humidity and vibration acceleration sequences. Based on the glass transition temperature range, moisture absorption-induced plasticity threshold and vibration spectrum energy change characteristics of PETG material, perform non-overlapping semantic segmentation on the multi-source environmental time-series data to generate a set of stress segments with six basic semantic tags.

[0024] S2: Construct a stress semantic graph structure based on the stress fragment set, map each stress semantic fragment to a graph node and assign node attribute vectors containing the mean stress amplitude, slope of change, duration and adjacent segment transition entropy. Only when the time interval between two stress semantic fragments is less than a preset threshold and there is a physical coupling relationship, a directed edge is established and the Jensen-Shannon divergence is calculated as the edge weight to generate a stress semantic graph object with topological connection relationship.

[0025] S3: The learnable discrete label encoder maps the six basic semantic labels in the stress semantic graph object into low-dimensional dense vectors, and then performs a fusion operation with the node attribute vectors to generate a semantic embedding node representation matrix that represents the initial physical state of each stress semantic segment.

[0026] S4: Perform multi-hop topology-aware propagation based on semantic embedded node representation matrix. In the process of second-order neighborhood information aggregation, introduce stress-sensitive gating mechanism. Dynamically adjust the information injection intensity of adjacent nodes according to the material response priority corresponding to the semantic type of the current node to suppress irrelevant semantic interference and generate a topology propagation node representation matrix that has been optimized by local stress sensitivity adaptive weighting.

[0027] S5: Apply time order constraints to the topology propagation node representation matrix, perform ordered splicing of node representations according to the order of occurrence of stress semantic segments, input them into lightweight long short-term memory network units for processing, extract cross-segment dependency features, and output a single temporal health decay trend vector.

[0028] S6: Based on the parameters of the life prediction model obtained by combining a single time-series health decay trend vector with the course learning strategy, calculate the remaining service life prediction value of PETG bottles under the current dynamic environmental stress, and generate a crack resistance life assessment report including failure risk level and prediction confidence.

[0029] Step S1: Acquire multi-source environmental time-series data during the service life of the PETG bottle. The data includes temperature, humidity, and vibration acceleration sequences. Based on the glass transition temperature range, moisture-induced plasticity threshold, and vibration spectrum energy abrupt change characteristics of PETG material, the multi-source environmental time-series data is processed into non-overlapping semantic segments to generate a set of stress segments with six basic semantic tags. Specifically, this includes:

[0030] S1.1: Acquire multi-source environmental time-series data during the service life of PETG bottles. The data includes temperature, humidity and vibration acceleration sequences. Perform timestamp alignment and outlier removal on the multi-source environmental time-series data to generate a standardized multi-source environmental time-series dataset.

[0031] By deploying IoT sensor nodes on logistics pallets or in warehousing environments, raw analog signals of the environment around PETG bottles are collected in real time, including voltage signals output by thermocouples, impedance change signals from capacitive humidity sensors, and triaxial acceleration signals from MEMS accelerometers, forming a raw multi-source data stream with non-uniform sampling intervals.

[0032] To address the time asynchrony issue caused by clock drift from multiple sensors, a linear interpolation resampling algorithm is employed. Using the least common multiple sampling frequency as a benchmark, temperature, humidity, and vibration data are uniformly mapped to the same time axis, ensuring that the data points of each physical quantity at any time t have a strict time correspondence.

[0033] A sliding window statistical test is performed on the resampled data. The window length is set to N. The mean μ and standard deviation σ of the data within the window are calculated. The 3σ criterion is used to identify and remove outliers that deviate from the mean by more than three times the standard deviation, thus eliminating spike noise caused by electromagnetic interference or sensor transient failures.

[0034] To address the data gaps resulting from outlier removal, cubic spline interpolation is used for smooth filling, maintaining the continuity of the second derivative of the data curve, restoring the physical continuity of environmental stress changes, and avoiding high-frequency artifacts introduced by linear interpolation.

[0035] The cleaned multi-source data is Z-score standardized, which converts the original measurements into standardized values ​​based on the historical statistical mean and standard deviation of each physical quantity, thereby eliminating the order-of-magnitude differences between different physical units (degrees Celsius, percentage RH, g).

[0036] The environmental response characteristics of PETG materials refer to the physical and mechanical property response laws exhibited by PETG materials under environmental stresses such as temperature, humidity, and mechanical vibration. These include: the thermal response characteristics of intensified molecular chain segment motion when the temperature reaches the glass transition temperature range; the wet response characteristics of plasticizing effect when humidity exceeds the moisture-induced plasticity threshold; and the mechanical response characteristics of microcrack initiation triggered when vibration energy exceeds the material fatigue damage threshold. In this embodiment, the above environmental response characteristics are quantitatively characterized by the glass transition temperature range determination logic in S1.2, the moisture-induced plasticity threshold rule, and the vibration spectrum energy mutation feature extraction operation in S1.3, serving as the basis for semantic segmentation processing.

[0037] Semantic segmentation refers to the process of segmenting and labeling continuous multi-source environmental time-series data according to the type of environmental stress and its physical significance on the material failure mechanism, based on the environmental response characteristics of PETG materials. Each labeled time-series interval is called a stress semantic segment, and its "semantics" reflects the triggering or promoting effect of environmental stress on specific failure modes of PETG materials within that time period, such as thermal aging, moisture absorption plasticization, condensation cracking, mechanical impact damage, ultraviolet light aging, or multi-field coupling damage.

[0038] Through the above time alignment, anomaly removal and normalization processes, the heterogeneous and noisy raw sensor data is transformed into a standardized multi-source environmental time-series dataset with consistent dimensions and stable distribution, providing a high-quality data foundation for subsequent semantic segmentation based on material response thresholds.

[0039] For example, the sampling frequency is set to 10Hz and the window length N is 50. For temperature data, the mean is 25℃ and the standard deviation is 2℃. Based on the 3σ criterion, the anomaly threshold is calculated to be 25 + 3 × 2 = 31℃. If the reading at a certain moment is 33℃, exceeding this threshold, it is judged as an anomaly and removed. After cubic spline interpolation, the temperature, humidity, and vibration data are Z-score standardized to ensure that the mean of each sequence is 0 and the variance is 1. This processing effectively eliminates sensor noise interference, ensuring the subsequent semantic slicing algorithm's sensitivity to minute stress changes and significantly improving the signal-to-noise ratio of feature extraction.

[0040] S1.2: Based on the standardized multi-source environmental time series dataset, the glass transition temperature range determination logic of PETG material is used to initially screen the temperature sequence for high-temperature slow change segment and condensation cycle segment. At the same time, the humidity sequence is identified for high humidity retention segment according to the moisture absorption induced plasticity threshold rule, so as to generate a set of thermal and humid stress candidate segments with preliminary material response labels.

[0041] Temperature and humidity sequences from a standardized multi-source environmental time-series dataset are read as input data sources for thermal and humidity stress semantic segmentation.

[0042] The logic for determining the glass transition temperature range of PETG material is invoked, and 70℃ to 85℃ is set as the critical phase transition threshold range. A sliding window scan is performed on the temperature sequence.

[0043] When the average temperature within the window falls into this range and the rate of change is lower than the preset gradient threshold, it is marked as a candidate segment for slow temperature change, which represents the state of intensified molecular chain segment motion.

[0044] The system detects the process of rapidly decreasing from above the dew point temperature to below the dew point temperature in a temperature sequence, identifies candidate segments of the condensation cycle, and captures the risk of microcracks caused by surface condensation.

[0045] The humidity sequence was processed synchronously, and continuous segments with relative humidity greater than 65% and lasting for more than 4 hours were selected based on the hygroscopic induced plasticity threshold rule.

[0046] Segments that meet the above humidity conditions are marked as high humidity retention candidate segments, reflecting the material plasticizing effect caused by water molecule penetration.

[0047] The preliminary screening of high temperature slow change, condensation cycle and high humidity retention candidate segments were checked for time sequence alignment, and invalid segments with overlapping time or blurred boundaries were removed.

[0048] By using a thermal-humidity coupling determination mechanism, the results of the previous step are transformed into a set of thermal-humidity stress candidate fragments with preliminary material response labels, thereby achieving structured extraction of the physical meaning of environmental stress.

[0049] S1.3: Perform spectral energy mutation feature extraction on the vibration acceleration sequence in the standardized multi-source environmental time series dataset, calculate the root mean square value jump amplitude and duration of acceleration in a specific frequency band, and generate a candidate fragment set of impact clusters to identify mechanical impact events.

[0050] Vibration acceleration sequences were extracted from the standardized multi-source environmental time-series dataset. The sampling frequency was set to 1000Hz. The original time-domain signal was processed by Hanning window with a window length of 0.5 seconds and an overlap rate of 50% to preserve the local time-frequency characteristics of the signal.

[0051] A fast Fourier transform is performed on the framed acceleration signal to convert the time-domain signal into a frequency-domain spectrum, focusing on the 10–50 Hz frequency band, which corresponds to the main resonance range of PETG bottles in transportation and stacking scenarios.

[0052] Calculate the root mean square value of acceleration within a specified frequency band to quantify the vibration energy intensity within that band. The root mean square value is calculated by taking the square root of the sum of the squares of the acceleration amplitudes at each sampling point within the frequency band divided by the number of sampling points.

[0053] Real-time monitoring of the trajectory of the root mean square acceleration value within the frequency band over time; when the current window RMS is detected... f A potential shock initiation point is marked when the jump from the previous stable baseline exceeds 200%.

[0054] The duration of the high-energy state is continuously tracked. If the high-energy state is maintained for more than 0.3 seconds, it is determined to be a valid mechanical impact event, excluding instantaneous noise interference.

[0055] Time segments that satisfy both amplitude and duration constraints are extracted and encapsulated to generate a candidate set of impact cluster segments that identify mechanical impact events.

[0056] By extracting the spectral energy mutation features and using the dual threshold determination mechanism, the continuous vibration flow is transformed into discrete high-confidence impact event fragments, thus achieving accurate capture of mechanical damage sources in dynamic service environments.

[0057] For example, vibration data of a batch of PETG bottles during logistics transportation was analyzed, with a sampling frequency of 1000Hz. In the interval t=120s to 125s, the RMS acceleration value in the 10–50Hz frequency band surged from a baseline of 0.5g to 1.8g, a jump of 260%, and this high-energy state lasted for 0.45 seconds. Based on its judgment logic, the system marked this 0.45-second period as a candidate segment of "impact cluster" and recorded its start and end timestamps and peak energy, effectively eliminating instantaneous road bump noise lasting only 0.1 seconds, ensuring the physical authenticity of the input data for subsequent life prediction models.

[0058] S1.4: Based on the candidate fragment set of thermal and damp stress and the candidate fragment set of impact cluster, the ultraviolet radiation superposition determination rule and the composite disturbance correlation logic are introduced to carry out multi-physics field coupling analysis. Overlapping or adjacent candidate fragments are merged and labeled as ultraviolet superposition segments or composite disturbance segments to generate a complete set of stress fragments containing six basic semantic tags.

[0059] like Figure 2 As shown, Figure 2 This is the flowchart for S1.4, which shows the complete decision-making path from the input candidate fragment set, through ultraviolet superposition judgment, composite perturbation judgment, to the final output of a set of six semantic label fragments. Specifically:

[0060] Obtain candidate fragment sets for thermal and humid stress and impact clusters, extract the start and end timestamps and spatial location identifiers of each fragment, and construct a multiphysics spatiotemporal index table.

[0061] Retrieve the preset ultraviolet radiation sensor data stream, calculate the ultraviolet irradiance integral value within the time window corresponding to each candidate segment, and mark it as a potential ultraviolet superposition segment if the integral value exceeds the material photoaging threshold and overlaps with the heat and humidity or impact segment in time.

[0062] The multi-physics coupling judgment logic is applied to detect the adjacency relationship between the thermal and damp segment and the impact segment on the time axis. If the interval between the two segments is less than the preset coupling time threshold and they belong to different physical field types, they are judged as a composite disturbance event.

[0063] For regions marked as potential UV overlay segments, the original thermal or shock segments are merged with UV radiation data to generate a four-dimensional attribute vector containing temperature, humidity, vibration, and UV intensity, and then assigned the semantic label "UV overlay".

[0064] For regions identified as complex disturbance events, adjacent thermal and humid segments and impact segments are spliced ​​together in time sequence. The stress change rate mutation index of the spliced ​​sequence is calculated. If the mutation index exceeds the dynamic coupling threshold, a semantic label of "complex disturbance" is assigned.

[0065] For the remaining candidate segments that do not trigger the determination of ultraviolet superposition or composite perturbation, their original semantic tags of "high temperature slow change", "high humidity retention", "condensation cycle" or "impact clustering" are retained to ensure that all time periods are covered.

[0066] Integrate all labeled segments, arrange them in chronological order, eliminate gaps and overlaps on the timeline, and generate a complete set of stress segments containing six basic semantic tags.

[0067] Through multiphysics coupling analysis and semantic merging, scattered single-physics candidate fragments are transformed into a complete set of semantic fragments covering complex environmental interactions, realizing a panoramic semantic description of the service environment of PETG bottles and providing complete node data for subsequent graph structure construction.

[0068] For example, a coupling time threshold of 5 minutes and an ultraviolet aging threshold of 200 J / m² are set. Within a certain time period, 10:00-10:30 is a high-humidity retention period, and 10:32-10:35 is an impact clustering period. The interval between these two periods is 2 minutes, less than the threshold, and the integrated ultraviolet irradiance value during 10:00-10:35 is 250 J / m², exceeding the threshold. The system merges this time period into a composite segment, calculates its four-dimensional attribute vector, and determines it as a "composite disturbance" tag according to priority rules, because the synergistic effect of mechanical impact and hydrothermal coupling on PETG crack propagation is significantly higher than that of single ultraviolet aging. In the final output segment set, this time period is uniquely identified as the composite disturbance segment, avoiding tag conflicts and ensuring the mutual exclusivity and completeness of semantic description.

[0069] Step S2: Construct a stress semantic graph structure based on the stress fragment set. Map each stress semantic fragment to a graph node and assign it a node attribute vector containing the mean stress amplitude, slope of change, duration, and transition entropy of adjacent segments. Only when the time interval between two stress semantic fragments is less than a preset threshold and a physical coupling relationship exists, a directed edge is established, and the Jensen-Shannon divergence is calculated as the edge weight, generating a stress semantic graph object with topological connections. Specifically, this includes:

[0070] S2.1: Obtain each stress semantic segment in the stress segment set as the processing object. Based on the statistical distribution characteristics of the temperature sequence, humidity sequence and vibration acceleration sequence inside the segment, calculate four key indicators: mean stress amplitude, slope of change, duration and adjacent segment transition entropy. Combine and encapsulate the above four key indicators to generate a node attribute vector containing complete physical state information.

[0071] Substep S2.1 aims to quantify the physical features of each segment in the stress segment set, generating a node attribute vector containing complete physical state information. This step follows the stress segment set with six basic semantic labels generated in S1, providing node attribute data for subsequent construction of the stress semantic graph.

[0072] It should be noted that stress statistical features refer to statistical quantities extracted from the temperature sequence, humidity sequence, and vibration acceleration sequence within a stress semantic segment to characterize the environmental stress level and its variation law of that segment. In this embodiment, stress statistical features are specifically manifested in the following four key indicators: mean stress amplitude, slope of change, duration, and transition entropy of adjacent segments.

[0073] The system iterates through the set of stress segments, extracting the raw data of the temperature, humidity, and vibration acceleration sequences corresponding to the current processing object. For the temperature and humidity sequences, the arithmetic mean over the segment duration is calculated as the mean stress amplitude, characterizing the average environmental load level during that period. The temperature and humidity sequences are linearly fitted using the least squares method, and the slope of the fitted line is extracted as the change slope, quantifying the evolution trend and rate of environmental stress within the segment. The difference between the start and end timestamps of the segment is statistically analyzed to determine the duration, reflecting the cumulative time effect of a specific stress state on the PETG material.

[0074] For a vibration acceleration sequence, the difference in probability distribution at the boundary between two adjacent stress semantic segments is calculated. Based on Shannon entropy, the information entropy values ​​of the end window of the preceding segment and the beginning window of the following segment are calculated, i.e., the sum of the negative values ​​of the product of the probability distribution and its logarithm for each amplitude interval. The absolute difference between the entropy values ​​of the preceding and following segments is then calculated as the transition entropy of the adjacent segment, measuring the severity of the abrupt change in stress state.

[0075] The four scalar indices—mean stress amplitude, slope of change, duration, and adjacent segment transition entropy—are calculated and concatenated into a vector. Z-score normalization is then performed to eliminate numerical magnitude differences between different physical dimensions. A four-dimensional node attribute vector is generated, serving as the unique identifier and state description of this stress semantic segment within the graph structure.

[0076] Through the above-mentioned quantitative calculation and standardized encapsulation processing, the discrete semantic fragments of the previous step are transformed into node attribute vectors with continuous physical meaning, realizing a refined mathematical representation of the local stress state of the PETG bottle in service environment, and providing high-fidelity input data for the subsequent topology propagation of graph neural networks.

[0077] For example, a stress semantic segment labeled "high temperature gradual change" is selected, with a duration of 3600 seconds. The temperature sequence within this segment is extracted, and the average temperature is calculated to be 75℃, with a linear fitting slope of 0.02℃ / s. The humidity sequence is extracted, with an average relative humidity of 40% and a slope of -0.01% / s. The amplitude distribution probability of the vibration acceleration sequence 10 seconds before the end of the segment and 10 seconds after the start of the next segment is calculated, with an entropy value of 2.3 bits for the first segment, 2.8 bits for the second segment, and 0.5 bits for the adjacent segment transition entropy. The duration of 3600 seconds, the average temperature of 75℃, the temperature slope of 0.02, and the adjacent segment transition entropy of 0.5 are combined into an initial vector [3600, 75, 0.02, 0.5]. After Z-score normalization, a node attribute vector [0.85, 1.2, -0.3, 0.6] is generated. This vector accurately reflects the long-term high-temperature, low-speed heating characteristics and moderate state abrupt changes of this segment, significantly improving the accuracy of subsequent models in capturing the thermal aging effect of materials.

[0078] S2.2: Obtain the timestamp information of all generated node attribute vectors and their corresponding stress semantic segments, filter out candidate segment pairs with time intervals less than the preset threshold according to the preset time interval threshold, and further determine whether there is a specific physical coupling relationship between the candidate segment pairs that is high humidity retention pointing to condensation cycle or impact cluster pointing to high temperature slow change according to the physical coupling relationship rule base, thereby generating a set of potential connection edges that meet the spatiotemporal constraints.

[0079] Obtain the set of node attribute vectors generated by S2.1 and the timestamp sequence of the corresponding stress semantic segments as the input data source for spatiotemporal correlation analysis. Traverse all stress semantic segment pairs, calculate the time difference between the end time and the start time of any two segments, filter out candidate segment pairs with a time interval of less than a preset threshold of 30 minutes, and construct a preliminary temporal proximity set.

[0080] The physical coupling relationship rule base is invoked to retrieve whether the semantic label of the preceding segment in the candidate segment pair belongs to the physical cause type of the following segment. Specifically, the coupling logic for determining the high humidity retention segment pointing to the condensation cycle segment is as follows: if the preceding node label is high humidity retention and the succeeding node label is condensation cycle, then it is determined that there is a moisture absorption-phase change coupling path.

[0081] The specific logic for determining the coupling between impact clusters and high-temperature gradual-change segments is as follows: if the predecessor node is labeled as an impact cluster and the successor node is labeled as a high-temperature gradual-change segment, then a mechanical damage-thermal relaxation coupling path is determined to exist. Candidate pairs that satisfy the above specific physical coupling relationship are marked as valid potential connection edges. Random neighboring segment pairs without physical mechanism support are removed to ensure the causal rationality of the graph structure.

[0082] Through a dual filtering mechanism of spatiotemporal constraint screening and physical mechanism verification, discrete fragments are transformed into a set of potential connecting edges with clear failure evolution logic, realizing a structured mapping from independent stress events to associated failure chains.

[0083] For example, a time interval threshold is set to 30 minutes. Segment A is a high-humidity retention segment, ending at 10:00; segment B is a condensation cycle segment, starting at 10:15. The time difference of 15 minutes is less than the threshold. A rule base query confirms a physical coupling between high-humidity retention and condensation cycle. A potential directed edge from A to B is generated. Segment C is an impact cluster, and segment D is ultraviolet superposition; although they are temporally adjacent, no coupling relationship is defined, so no connection is established. The final output is a set of potential connection edges containing valid physical causal relationships.

[0084] S2.3: Obtain the node attribute vector of each pair of connected stress semantic segments in the potential connection edge set, use the Jensen-Shannon divergence algorithm to calculate the degree of difference between the two node attribute vectors in the probability distribution, quantify the calculated degree of difference into edge weight values ​​that characterize the severity of the sudden change in the stress parameters of the two segments, and generate weighted directed edge data with quantitative weight information.

[0085] Obtain the node attribute vector for each pair of connected stress semantic segments in the potential connection edge set. The vector contains physical feature data in four dimensions: mean stress amplitude, slope of change, duration, and adjacent segment transition entropy.

[0086] The node attribute vectors are processed by probability distribution mapping. The Softmax normalization function is used to convert the physical feature values ​​of each dimension into non-negative probability distribution vectors with a sum of 1, so as to eliminate the numerical scale difference between different physical units and construct a comparable statistical distribution model.

[0087] Based on the transformed probability distribution vector, the Jensen-Shannon divergence algorithm is used to calculate the distribution differences between the source node and the target node in the multidimensional stress feature space, and to quantify the severity of the abrupt change in the two stress parameters.

[0088] The specific calculation formula is as follows:

[0089]

[0090] Where P and Q represent the normalized probability distribution vectors of the source node and the target node, respectively, M is the arithmetic mean distribution vector of P and Q, and KLD represents the Kullback-Leibler divergence.

[0091] The calculated Jensen-Shannon divergence values ​​are directly assigned to the corresponding directed edges as edge weights characterizing the intensity of environmental stress state transitions.

[0092] Through the above processing method, the potential connection relationship generated in the previous step is transformed into weighted directed edge data with quantitative weight information, realizing the accurate quantification of stress mutation risk in dynamic environment, and providing topological constraints that conform to physical failure mechanism for subsequent information aggregation of graph neural network.

[0093] For example, for a candidate edge pointing from "high humidity retention" to "condensation cycle", the source node attribute vector is [0.8, 0.1, 5.0, 0.2], and the target node attribute vector is [0.3, 0.6, 2.0, 0.5]. After Softmax normalization, the probability distributions are obtained as P=[0.45, 0.12, 0.38, 0.05] and Q=[0.15, 0.40, 0.25, 0.20]. The average distribution M=[0.30, 0.26, 0.315, 0.125]. Substituting into the Jensen-Shannon divergence formula, KLD(P||M) is calculated to be 0.18, KLD(Q||M) is 0.22, and the final edge weight JSD value is 0.20. This weighting accurately reflects the high-risk state of stress concentration and microcrack initiation on the surface of PETG materials when the environment changes abruptly from a stable high-humidity environment to a condensing environment, significantly improving the model's accuracy in identifying the coupling effects of complex environments.

[0094] S2.4: Obtain the mapped stress semantic fragments as graph nodes and the weighted directed edge data with quantitative weight information as graph edge connection relationships. Perform graph structure assembly operation to integrate the graph nodes and weighted directed edge data with the topological connection logic in chronological order, and finally generate a stress semantic graph object with complete topological connection relationships and containing node attributes and edge weight information.

[0095] Obtain the set of stress semantic fragment nodes after mapping and the weighted directed edge data with quantitative weight information, as the input object for graph structure assembly. Traverse all generated graph nodes and assign a unique topological index identifier to each node. This identifier is generated based on the start time of the stress semantic fragment on the time axis, ensuring that the storage order of nodes in the graph data structure strictly follows the monotonically increasing characteristic of physical time.

[0096] Each node's attribute vector, including the mean stress amplitude, slope of change, duration, and adjacent segment transition entropy, is bound to its corresponding topological index to form a node feature dictionary, achieving a precise mapping between physical state information and graph topological location. Weighted directed edge data is parsed, and source and target node indices are extracted to construct row and column index arrays of a sparse adjacency matrix, representing the directed connections between nodes.

[0097] The Jensen-Shannon divergence values ​​calculated in the previous step are filled into the corresponding positions in the adjacency matrix as edge weights. For node pairs that do not have a physical coupling relationship, their adjacency matrix elements remain zero, thus reflecting the sparsity characteristics of the graph structure.

[0098] A temporal causality constraint mechanism is introduced to verify the directionality of all directional edges, ensuring that only semantic fragments from earlier times are allowed to point to semantic fragments from later times or at the same moment. Any reverse edge connections that violate the temporal causality law are eliminated to prevent future information from being leaked into the current state evaluation. A node feature dictionary and a sparse adjacency matrix are integrated to instantiate a graph data object. This object contains three core components: a node feature matrix, an edge index matrix, and an edge weight vector, completing the digital encapsulation of the stress semantic graph.

[0099] Through the above graph structure assembly operation, discrete stress semantic fragments and their physical coupling relationships are transformed into a standardized graph topology structure that can be processed by a computer, realizing the structured representation of multi-source environmental stress data, and providing a data foundation with spatiotemporal consistency for the subsequent topology-aware propagation of time-series graph neural networks.

[0100] For example, based on service data of a batch of PETG bottles in a tropical marine transportation scenario, a node set containing 120 stress semantic segments is generated. Each node is assigned a unique topological index from 0 to 119, where index 0 corresponds to the high-temperature gradual change segment at the start time, and index 119 corresponds to the composite perturbation segment at the end time. The node attribute vector has a dimension of 4, storing the normalized mean temperature (0.75), humidity change slope (-0.12), duration (3600 seconds), and neighboring segment transition entropy (0.45). A sparse adjacency matrix of size 120x120 is constructed, with 350 non-zero elements, corresponding to segments with physical coupling relationships. For example, there is a directed edge between the high-humidity retention segment at index 15 and the condensation cycle segment at index 16, with a Jensen-Shannon divergence weight of 0.08, representing a small distributional difference and strong physical continuity between the two. Verification revealed that the connection from index 50 to index 49 violated temporal causality and was therefore removed. The final generated graph data object contains a 120x4 node feature matrix, a 2x350 edge index matrix, and an edge weight vector of length 350. It successfully constructs a stress semantic graph that reflects the dynamic environment evolution law, and significantly improves the model's ability to perform structured modeling of complex service conditions.

[0101] Step S3: Utilize a learnable discrete label encoder to map the six basic semantic labels in the stress semantic graph object into low-dimensional dense vectors. Then, fuse these low-dimensional dense vectors with node attribute vectors to generate a semantic embedding node representation matrix that characterizes the initial physical state of each stress semantic segment. Specifically, this includes:

[0102] S3.1: Based on a pre-built learnable discrete label encoder, vector mapping processing is performed on six basic semantic labels contained in the stress semantic map object, namely high temperature slow change, high humidity retention, condensation cycle, impact cluster, ultraviolet superposition and composite perturbation, to convert discrete category labels into a low-dimensional dense vector set with differentiable properties.

[0103] It should be noted that fusion encoding refers to the process of jointly encoding the discrete semantic label information carried by each graph node in the stress semantic graph object with the continuous physical attribute information to generate a feature vector of a unified dimension. This process aims to eliminate the distributional differences between discrete category labels and continuous physical dimensions, enabling deep fusion of semantic information and physical state in the feature space.

[0104] The system receives six basic semantic labels carried by each node in the stress semantic graph object, including high temperature gradual change, high humidity retention, condensation cycle, impact clustering, ultraviolet superposition, and composite perturbation. A learnable discrete label encoder is constructed, and the embedding matrix is ​​initialized with an embedding dimension of 16 to ensure that the low-dimensional space can fully represent semantic differences and avoid the curse of dimensionality.

[0105] One-hot encoding is performed on the discrete category identifier of each node to generate a sparse binary vector representation, which serves as the input data for the encoder. Matrix multiplication is then performed between the one-hot vector and the embedding matrix to obtain the corresponding low-dimensional dense semantic vector.

[0106] A random dropout mechanism is introduced to mask some semantic features with a probability of 0.1 during the training phase, enhancing the model's robustness to environments with missing labels. The embedding matrix parameters are updated using the backpropagation algorithm, bringing semantically similar labels closer together in the vector space and achieving a continuous representation of semantic features.

[0107] Through the above mapping process, discrete category identifiers are converted into a low-dimensional dense vector set with differentiable properties, realizing the quantitative transformation of semantic information into numerical features and providing standardized input for subsequent multimodal feature fusion.

[0108] For example, the embedding dimension is set to 16, and the initial embedding matrix E is 6x16. For a node labeled "impact cluster", its one-hot vector is [0,0,0,1,0,0]. After performing matrix multiplication, the vector in the 4th row is obtained as the initial semantic embedding of the node, such as [0.12,-0.35,...,0.08]. During training, if the discarding mechanism is triggered, some elements of this vector are set to zero. After 100 rounds of training iterations, the embedding matrix converges, and the cosine similarity between the "impact cluster" and "composite perturbation" vectors improves to 0.85, significantly better than the 0.2 of random initialization, indicating that the model successfully captures the coupling relationship between physical semantics and effectively improves the discriminative power of feature representation.

[0109] S3.2: Using the feature alignment algorithm, normalization and dimension matching are performed on the node attribute vectors of each graph node in the stress semantic graph object, which consist of the mean stress amplitude, slope of change, duration and adjacent segment transition entropy, in order to generate a standard numerical attribute feature matrix consistent with the dimension of the low-dimensional dense vector space.

[0110] Obtain the original node attribute vector carried by each graph node in the stress semantic graph object. The vector contains four feature dimensions with different physical dimensions: mean stress amplitude, slope of change, duration, and adjacent segment transition entropy.

[0111] The original node attribute vectors are subjected to Min-Max normalization to eliminate the order-of-magnitude differences between different physical dimensions, and each feature value is mapped to the [0,1] interval to construct a dimensionless standard feature subset.

[0112] The Z-Score normalization algorithm is used to align the distribution of the normalized feature subsets. The mean and standard deviation of each feature dimension are calculated. By subtracting the mean and dividing by the standard deviation, the data distribution conforms to the normal distribution characteristics of zero mean and unit variance.

[0113] A linear projection matrix is ​​constructed to map the standardized four-dimensional physical feature vector to the same low-dimensional dense vector space dimension as the discrete semantic label embedding vector, ensuring dimensional consistency during subsequent multimodal feature fusion.

[0114] A fully connected layer is used to perform a nonlinear transformation on the projected feature vector, and the ReLU activation function is introduced to enhance the sparsity and nonlinear resolution of the feature representation, thereby generating a preliminary standard numerical attribute feature representation.

[0115] The standard numerical attribute feature representation is normalized using the L2 norm, and the feature vector magnitude is constrained to 1 to prevent gradient explosion caused by excessively large feature magnitudes, thereby optimizing the convergence stability of subsequent neural network training.

[0116] By using a feature alignment algorithm, the heterogeneous physical dimension data obtained in the previous step is transformed into a numerical attribute feature matrix that is consistent with the semantic embedding space dimension and has a standardized distribution, thus achieving seamless fusion preparation of multi-source heterogeneous features.

[0117] For example, for a service segment of a PETG bottle, its original node attribute vector is [average temperature 80℃, humidity slope 0.5% / h, duration 4h, transition entropy 0.8bit]. First, Min-Max normalization is performed, assuming a temperature range of [20,100], a humidity slope range of [0,2], a duration range of [0,10], and an entropy range of [0,1], resulting in a normalized vector [0.75,0.25,0.4,0.8]. Next, the mean vector [0.5,0.5,0.5,0.5] and standard deviation vector [0.2,0.2,0.2,0.2] of this batch of data are calculated, and Z-Score normalization is performed to obtain [-1.25,-1.25,-0.5,1.5]. The target embedding dimension is set to 64-dimensional, and a 64x4 learnable projection matrix W is constructed. The 4-dimensional vector is mapped to a 64-dimensional vector H=W·X+b through matrix multiplication. After ReLU activation, negative values ​​are set to zero, while positive feature responses are preserved. Finally, L2 normalization is performed, the vector magnitude is calculated and divided by this magnitude, and the final standard numerical attribute feature vector is output. This process eliminates the dimensional conflict between temperature and entropy, allowing the numerical features and the 64-dimensional semantic embedding vector generated by S3.1 to be concatenated on the same spatial scale, significantly improving the efficiency of subsequent graph neural networks in aggregating multimodal information.

[0118] S3.3: Based on the multimodal feature fusion mechanism, the generated low-dimensional dense vector set is concatenated or weighted summed with the standard numerical attribute feature matrix element by element to eliminate the distribution difference between discrete semantic information and continuous physical dimensions and generate the fused hybrid feature representation tensor.

[0119] The set of low-dimensional dense semantic vectors generated by S3.1 and the standard numerical attribute feature matrix processed by S3.2 are obtained as the input data source for multimodal feature fusion.

[0120] A fusion weighted gating network is constructed, and learnable scalar fusion coefficients are initialized for six basic semantic labels to balance the contributions of discrete semantic information and continuous physical dimensions in the feature space.

[0121] Perform a linear projection transformation on the standard numerical attribute feature matrix to map its dimensions to the same feature space as the low-dimensional dense semantic vector, thus eliminating the distribution shift caused by the difference in dimensions.

[0122] An element-wise concatenation strategy is adopted to concatenate the mapped numerical feature vector with the corresponding semantic label vector along the feature dimension to form a preliminary hybrid feature representation tensor.

[0123] An adaptive weighted summation mechanism is introduced, which uses a learnable weight matrix to weight the semantic and numerical vectors element-wise, and then outputs the fused node representations through a nonlinear activation function. This mechanism ensures that the model can dynamically adjust the fusion ratio of semantic labels and physical values ​​according to the sensitivity of PETG materials to different stress types, avoiding the dominance of a single mode in feature expression.

[0124] The fused hybrid feature representation tensor is normalized by the layer to stabilize the feature distribution and accelerate the convergence speed of subsequent neural network layers.

[0125] Through the aforementioned multimodal feature fusion mechanism, the discrete semantic vectors and continuous physical attributes generated in the previous step are transformed into a hybrid feature representation tensor of a unified dimension, realizing deep coupling between semantic information and physical state, and providing node initial states rich in physical meaning for subsequent topology propagation.

[0126] S3.4: Perform feature reconstruction processing on the fused hybrid feature representation tensor through a nonlinear activation function to suppress noise interference and enhance the nonlinear resolution capability of feature expression, thereby outputting a preliminary optimized node state feature vector sequence.

[0127] Receive the fused hybrid feature representation tensor generated by S3.3, which contains concatenated data of discrete semantic vectors and continuous physical properties.

[0128] A linear transformation operation is performed on the hybrid feature representation tensor, and the high-dimensional fusion features are mapped to the hidden layer space through a learnable weight matrix to eliminate the dimensional differences between different modal features and to initially extract nonlinear interaction information.

[0129] The LeakyReLU nonlinear activation function is introduced to perform element-wise activation processing on the linearly transformed feature vector, preserving small gradients in the negative region to prevent neuron death, while enhancing the model's sensitivity to stress-induced abrupt signals.

[0130] By applying the Dropout regularization mechanism, some neuron outputs are randomly masked at a preset dropout rate to block feature co-adaptation, thereby suppressing the influence of sensor noise and environmental interference on node state representation.

[0131] Batch normalization is performed on the regularized data to readjust the feature distribution so that its mean is zero and its variance is one, thereby accelerating network convergence and improving the model's generalization stability in dynamic environments.

[0132] By employing the aforementioned nonlinear activation and regularization methods, the hybrid features from the previous step are transformed into a robust preliminary optimized node state feature vector sequence, achieving the expected technical effects of enhanced noise suppression and nonlinear resolution.

[0133] For example, the hidden layer dimension is set to 64, the LeakyReLU slope is 0.01, and the Dropout rate is 0.2. A 128-dimensional hybrid feature tensor is mapped to a 64x128 weight matrix to obtain a 64-dimensional vector. This is then activated by LeakyReLU, with negative values ​​multiplied by 0.01. The outputs of 13 neurons are randomly masked. After batch normalization, a 64-dimensional pre-optimized node state feature vector is output. This processing significantly improves the discriminative power of features for key semantics such as impact clusters and effectively suppresses background noise interference in the steady-state aging phase.

[0134] S3.5: Based on the topological node index order of the stress semantic graph object, perform structured recombination processing on the initially optimized node state feature vector sequence to generate the semantic embedding node representation matrix that finally represents the initial physical state of each stress semantic segment, which serves as the standardized input data for subsequent topological propagation operations.

[0135] Obtain the preliminary optimized node state feature vector sequence output by S3.4, and parse the topological index identifier and corresponding timestamp metadata of each node in the stress semantic graph object. Sort the nodes in ascending order according to the timestamp metadata, construct a strictly monotonically increasing time index mapping table, and establish the causal logical order of feature vectors in the temporal dimension.

[0136] Based on the time index mapping table, a reordering operation is performed on the discretely distributed node state feature vectors to reorganize the non-contiguously stored feature data into an ordered tensor sequence that conforms to the time evolution law.

[0137] Perform a dimensionality standardization check on the recombined ordered tensor sequence to ensure that the dimension of the feature vector at each time step is consistent with the dimension of the input layer of the subsequent lightweight long short-term memory network unit.

[0138] If dimensionality mismatch exists, the feature vector length is corrected using zero-padding or linear interpolation to ensure the integrity and continuity of the time series data. The corrected ordered feature vector sequence is encapsulated into a three-dimensional matrix structure, where the first dimension represents the batch size, the second dimension represents the time step, and the third dimension represents the feature dimension, forming a standardized semantic embedding node representation matrix.

[0139] Through the above-mentioned structured recombination process, the discrete node features obtained in the previous step are transformed into standardized input data with clear temporal dependencies, realizing the transformation from static graph node representation to dynamic temporal feature flow, and providing a high-quality data foundation with temporal causal constraints for multi-hop topology sensing propagation in step S4.

[0140] For example, for PETG bottle service life data containing 12 stress semantic segments, the timestamps of each segment are parsed to t1 to t12. After ascending order, a time index sequence [1,2,...,12] is generated. Assume the feature vector output by S3.4 has a dimension of 64 and is initially stored in random order. After rearranging according to the index sequence, an ordered tensor of shape (1,12,64) is obtained. If a segment has missing features due to sensor packet loss, linear interpolation is performed to complete the feature vector using the average of the features of adjacent segments. The final generated semantic embedding node representation matrix is ​​directly input to the S4 module to ensure that the topology propagation process can accurately capture the stress coupling effect that evolves over time, significantly improving the model's response sensitivity to dynamic environmental changes.

[0141] like Figure 3 As shown, step S4: Perform multi-hop topology-aware propagation based on the semantic embedding node representation matrix. A stress-sensitive gating mechanism is introduced during the second-order neighborhood information aggregation process. The injection intensity of adjacent node information is dynamically adjusted according to the material response priority corresponding to the semantic type of the current node to suppress irrelevant semantic interference, generating a topology propagation node representation matrix that has undergone adaptive weighted optimization based on local stress sensitivity. Specifically, this includes:

[0142] S4.1: Based on the discrete semantic labels in the semantic embedding node representation matrix, the pre-set PETG material failure mechanism knowledge base is called for matching query, and the crack initiation weight coefficient and crystal region relaxation weight coefficient corresponding to each semantic type are extracted to generate a material response priority vector set containing the material response priority quantification index.

[0143] It should be noted that the graph topology propagation operation refers to the process of transferring and aggregating the feature information of each graph node along the directed edges of the graph to neighboring nodes under the topological constraints of the stress correlation graph structure. This operation aims to enable each node to absorb the environmental stress information of its topologically adjacent nodes in the graph structure when updating its own representation, thereby realizing the modeling of causal relationships and synergistic effects between stress events that are difficult to reflect by a single independent feature.

[0144] Material response priority refers to the quantitative characterization of the promoting effect of stress fragments of different semantic types on different failure mechanisms (crack initiation and grain relaxation) of PETG materials. In this embodiment, the material response priority is implemented in the form of a two-dimensional vector, which includes two dimensions: crack initiation weight coefficient and grain relaxation weight coefficient. The two dimensions satisfy the normalization constraint condition, that is, the sum of the two weights is 1.

[0145] The discrete semantic labels corresponding to each node in the semantic embedding node representation matrix are obtained and used as index keys to query the pre-set PETG material failure mechanism knowledge base. This knowledge base stores the mapping relationship between six types of basic semantic labels and PETG micro-damage mechanisms, including two response modes: crack initiation-dominated and grain relaxation-dominated.

[0146] For nodes identified as 'impact clusters', 'condensation cycles', and 'composite disturbances', it is determined that their stress states are prone to inducing microcrack nucleation and propagation, and high-weight crack initiation weight coefficients are extracted from the knowledge base; for nodes identified as 'high temperature gradual change', 'high humidity retention', and 'ultraviolet superposition', it is determined that their stress states mainly cause molecular chain segment movement and free volume changes, and high-weight crystal region relaxation weight coefficients are extracted.

[0147] Construct a material response priority vector, which includes two dimensions: crack initiation weight and crystal relaxation weight. The two dimensions satisfy the normalization constraint condition, that is, the sum of the two weights is 1.

[0148] For the 'impact cluster' node, the crack initiation weight coefficient is set to 0.85 and the granular region relaxation weight coefficient is set to 0.15 to highlight the dominant role of mechanical impact on brittle fracture; for the 'high temperature gradual change' node, the crack initiation weight coefficient is set to 0.20 and the granular region relaxation weight coefficient is set to 0.80 to emphasize the influence of thermal history on the material creep characteristics; for the 'composite disturbance' node, the weight ratio is dynamically adjusted according to the stress amplitude ratio to ensure accurate characterization of multi-physics coupling effects.

[0149] By traversing all nodes in the stress semantic graph, a set of material response priority vectors, consistent with the number of nodes, is generated. This set quantifies the differences in the contribution of different environmental stress segments to the failure mechanism of PETG bottles, providing a physically interpretable control basis for subsequent gating mechanisms.

[0150] Through the above matching query and weight assignment processing, discrete semantic tags are transformed into a continuous set of material response priority vectors, realizing differentiated modeling of PETG damage evolution paths under different stress types, and significantly improving the model's ability to capture the sensitivity of dynamic environmental stress.

[0151] For example, when processing the stress semantic graph containing 'impact cluster' nodes N1 and 'high-temperature gradual change' nodes N2, a knowledge base query reveals that N1 corresponds to the crack initiation-dominant mechanism, and N2 corresponds to the granular region relaxation-dominant mechanism. The crack initiation weight for N1 is set to 0.9, and the granular region relaxation weight to 0.1; the crack initiation weight for N2 is set to 0.15, and the granular region relaxation weight to 0.85. In the generated material response priority vector set, the N1 vector is [0.9, 0.1], and the N2 vector is [0.15, 0.85]. This quantization result makes N1 more inclined to aggregate nearby high-stress impact information, while N2 focuses more on nearby thermal and moisture history information, effectively suppressing irrelevant semantic interference and significantly improving the robustness of the prediction model under complex working conditions.

[0152] S4.2: Using the set of material response priority vectors as control factors, a nonlinear gating function operation is performed on the feature vectors of adjacent nodes within the second-order neighborhood to calculate an adaptive gating coefficient matrix that reflects the degree of acceptance of neighbor information by the current node, so as to realize the dynamic filtering of information flow under different physical coupling relationships.

[0153] Receive the set of material response priority vectors generated by S4.1 and the semantic embedding node representation matrix output by S3, and extract the initial feature vectors and corresponding stress-sensitive weight coefficients of the current center node and all its neighboring nodes in its second-order neighborhood.

[0154] A nonlinear gating function based on the attention mechanism is constructed, which concatenates the semantic label encoding vector of the central node with the attribute vectors of the neighboring nodes and inputs it into the multilayer perceptron to calculate the original attention score.

[0155] The material response priority vector is introduced as a bias term or multiplication factor in the attention score calculation, and the information acceptance level of neighboring nodes is corrected using the following formula:

[0156]

[0157] in, Indicates the central node For neighboring nodes The adaptive gating coefficient, For the original attention score, This refers to the priority weight of the material response corresponding to the semantic type obtained from the knowledge base query.

[0158] The calculated gating coefficient matrix is ​​normalized to ensure that the sum of the gating coefficients of all second-order neighbors of the same central node is 1, so as to maintain the energy conservation property in the information aggregation process.

[0159] The normalized adaptive gating coefficient matrix is ​​used to weight the feature vectors of neighboring nodes in the second-order neighborhood element by element to generate a gating feature tensor that reflects the difference in local stress sensitivity.

[0160] By significantly amplifying the gating coefficient of high-priority semantic nodes (such as impact clusters) while suppressing the information injection intensity of low-priority semantic nodes (such as steady-state aging segments), the key failure causes in physical coupling relationships can be accurately captured.

[0161] For example, when the central node is of the "high temperature gradual change" type, the knowledge base query yields a crystal region relaxation weight coefficient of 0.6, while its second-order neighbors include an "impact cluster" node with a crack initiation weight coefficient of 0.9. When calculating the gating coefficient, the system uses 0.9 as a high-priority factor to amplify the attention score of this neighboring node. This increases the proportion of information from the "impact cluster" neighbor in the final gating coefficient from the original 0.2 to 0.65, while the proportion of information from ordinary "high humidity retention" neighbors decreases accordingly. This dynamic filtering mechanism effectively avoids interference from irrelevant environmental noise on the prediction of microcrack propagation in PETG bottles, significantly improving the robustness of feature extraction in complex dynamic environments.

[0162] S4.3: Based on the adaptive gating coefficient matrix and the edge weights of the original graph structure, the semantic embedded node representations of adjacent nodes are subjected to element-wise weighted aggregation processing to generate an intermediate-state neighborhood aggregation feature tensor that integrates local stress sensitivity features, thereby completing information transmission and noise suppression under the second-order topological structure.

[0163] The adaptive gating coefficient matrix and the weighted directed edge data in the original graph structure are obtained as input control parameters for neighborhood information aggregation. The second-order neighborhood set of each central node in the stress semantic graph is traversed, and the semantic embedding node representation vectors of adjacent nodes are extracted.

[0164] The representation vectors of adjacent nodes are multiplied element-wise with the Jensen-Shannon divergence weights of the corresponding edges to achieve preliminary feature weighting based on distributional dissimilarity. An adaptive gating coefficient matrix is ​​introduced to perform Hadamard product operations on the pre-weighted adjacent features, dynamically suppressing the injection intensity of low-priority semantic types.

[0165] Perform a summation and aggregation operation on the gated neighbor feature vectors to generate an intermediate-state neighborhood aggregation feature tensor that reflects the local topology.

[0166] This processing method transforms the gating coefficients from the previous step into specific feature filtering actions, thereby achieving noise suppression and key stress feature enhancement under second-order topology.

[0167] For example, for the central node labeled "impact cluster," its second-order neighborhood includes "high-temperature gradual change" and "condensation cycle" nodes. The response priority coefficient for crack initiation in the "impact cluster" is set to 0.9, and the gating coefficient for "high-temperature gradual change" is calculated to be 0.2. If the embedding vector of the neighbor node of "high-temperature gradual change" is [0.5, 0.3], and the edge weight is 0.1, then the weighted vector is [0.05, 0.03]. After modulation with a gating coefficient of 0.2, the final contribution value is reduced to [0.01, 0.006], significantly weakening the interference of non-critical stresses. Conversely, if the neighbor is another "impact cluster" node, the gating coefficient is 0.8, preserving most of the damage evolution information. This mechanism significantly improves the model's ability to capture key failure inductions in complex dynamic environments, effectively reducing the accumulation of prediction errors.

[0168] S4.4: Perform residual connection fusion operation on the intermediate neighborhood aggregated feature tensor and the initial semantic embedded node representation of the current node, and perform dimension alignment through a linear transformation layer to output the final topology propagation node representation matrix after local stress sensitivity adaptive weighting optimization, which serves as the input data for subsequent temporal joint pooling.

[0169] Receive the intermediate-state neighborhood aggregation feature tensor generated by S4.3, which contains second-order neighborhood environment stress interaction information after stress-sensitive gating. Obtain the initial semantic embedding node representation of the current node output by S3, which preserves the independent physical properties and semantic label priors of the original stress fragment.

[0170] The residual connection fusion operation is performed by adding the intermediate neighborhood aggregated feature tensor to the initial semantic embedded node representation element-wise. This operation aims to preserve the key identity information of the original nodes, prevent oversmoothing during the propagation of deep graph neural networks, and ensure that local stress mutation features are not overwhelmed by the global averaging effect.

[0171] A linear transformation layer is constructed, initializing the learnable weight matrix and bias vector. The feature vector after residual fusion is input into the linear transformation layer, performing matrix multiplication and bias addition operations to achieve dimensionality mapping and alignment of the feature space. The node representation after topology propagation is calculated using the following formula:

[0172]

[0173] in, For the first The topology propagation node characterization of each node after adaptive weighted optimization based on local stress sensitivity. It is the ReLU activation function. The weight matrix of the linear transformation layer. For the intermediate state neighborhood aggregated feature tensor, This serves as the initial semantic embedding node representation. This is the bias vector.

[0174] The ReLU activation function is applied to the linear transformation output to introduce nonlinearity, thereby enhancing the model's ability to fit complex stress coupling relationships and suppressing negative noise interference. A final topology propagation node representation matrix, adaptively weighted by local stress sensitivity, is generated, which includes both global topological dependencies and preserves local stress sensitivity.

[0175] By using residual connections and linear transformations, the aggregated features from the previous step are transformed into node representation data with high discriminative power, achieving the expected technical effects of feature dimension alignment and lossless information transfer, and providing standardized input for subsequent temporal joint pooling.

[0176] For example, the node feature dimension is set to 64. For nodes labeled as "impact clusters," their initial representation vector is added to the aggregated neighborhood feature vector, and after transformation by a 64x64 dimensional weight matrix, a new node representation is output. If a node is suppressed by the gating mechanism in S4.2 due to non-critical semantics, its aggregated features approach zero. The residual connections ensure that its original "high-temperature gradual change" features are preserved, avoiding the loss of important aging information. In the final output representation matrix, the norm of each node vector is stable within a reasonable range, significantly improving the accuracy of subsequent LSTM units in capturing long-term dependencies.

[0177] Step S5: Apply temporal order constraints to the topology propagation node representation matrix, sequentially concatenate the node representations according to the order of occurrence of stress semantic segments, and input them into a lightweight long short-term memory network unit for processing. Extract cross-segment dependency features and output a single temporal health decay trend vector. Specifically, this includes:

[0178] S5.1: The timestamp information associated with each node in the topology propagation node representation matrix is ​​parsed and processed. Based on the chronological order of stress semantic fragments, a strictly monotonically increasing time index sequence is generated to establish the causal logic basis for subsequent time series modeling and obtain an ordered time index sequence.

[0179] It should be noted that the temporal order constraint means that when converting the node representation in the graph structure into temporal features, the order in which the nodes are arranged on the time axis must strictly follow the causal logic of the occurrence of each stress semantic segment, and any information flow that violates the temporal causality law is prohibited, such as using stress information from future moments to represent features at the current moment.

[0180] S5.2: Based on the ordered time index sequence, the topology propagation node representation matrix is ​​rearranged in rows to arrange the discrete node representations in an ordered manner according to the time flow, so as to construct a continuous time-series feature tensor that reflects the evolution law of the entire service cycle.

[0181] Obtain the topology propagation node representation matrix generated by step S4. This matrix contains the feature vectors of each stress semantic segment and their corresponding timestamp index information. Analyze the timestamp data associated with each node in the matrix, extract the absolute timestamps of all nodes, and construct the original time series set.

[0182] The original time series set is sorted in ascending order to generate a strictly monotonically increasing time index queue, establishing the causal temporal logic of stress events. Based on the order of the time index queue, the corresponding node feature vectors are retrieved and extracted from the topology propagation node representation matrix.

[0183] The extracted feature vectors are linearly concatenated in chronological order to eliminate the spatial discrepancies caused by the graph structure, constructing a continuous temporal feature tensor that reflects the evolutionary pattern throughout the entire service life. Dimension checking and alignment are then performed on the continuous temporal feature tensor to ensure that its shape meets the input dimension requirements of lightweight long short-term memory network units.

[0184] By using row rearrangement and ordered concatenation, the discrete topological node representations from the previous step are transformed into continuous temporal feature tensors with strict temporal causal relationships, realizing the modal transformation from spatial graph structure to time series data, and providing standardized input data for capturing long-distance cross-segment dependencies.

[0185] For example, assume the topology propagation node representation matrix contains 5 stress semantic fragment nodes with timestamps t1=10min, t2=25min, t3=40min, t4=55min, and t5=70min, corresponding to a 64-dimensional feature vector. The system parses the timestamps to generate an index sequence [1,2,3,4,5]. The node feature vectors v1 to v5 are extracted in this order and concatenated to generate a continuous temporal feature tensor T of shape [5,64]. If there is a timestamp disorder, such as t3 being earlier than t2, the system automatically adjusts the extraction order to v1, v3, v2, v4, v5 to ensure correct temporal logic. The final output continuous temporal feature tensor T is directly used as input to the lightweight LSTM unit in S5.3, effectively preserving the time dependence of stress evolution and significantly improving the response accuracy of the subsequent lifetime prediction model to dynamic environmental changes.

[0186] S5.3: Utilizes predefined lightweight long short-term memory network units to receive continuous temporal feature tensors as input data, and performs temporal state update operations through a collaborative gating mechanism of forget gate, input gate, and output gate to capture long-distance cross-segment dependencies and suppress noise interference, thereby generating hidden layer state sequences.

[0187] Receive a continuous temporal feature tensor generated by S5.2 rearrangement. This tensor contains topological propagation node representation vectors arranged in time order, which serve as input data for lightweight long short-term memory network units.

[0188] Initialize the hidden layer state vector and the cell state vector, setting the initial value to zero vector to establish the starting reference point for the temporal recursive calculation.

[0189] The forgetting gate activation value is calculated by concatenating the current input vector with the hidden state vector from the previous time step, and then mapping it to the interval between 0 and 1 through the Sigmoid activation function to generate the forgetting coefficient matrix.

[0190] Based on the forgetting coefficient matrix, element-wise multiplication is performed on the cell state vector of the previous time step to selectively retain long-term dependency information in the historical service environment and remove irrelevant noise interference.

[0191] The input gate activation value and candidate cell state vector are calculated, and the combination of the current input and the hidden state at the previous time step is processed by the Sigmoid function and the Tanh function respectively to generate the information stream to be updated.

[0192] The input gate activation value is multiplied element-wise with the candidate cell state vector to obtain the new feature components that need to be injected into the current cell state.

[0193] The historical cell states that have been filtered through forgetting are added to the newly added feature components to update the cell state vector at the current moment, thus completing the iteration of the internal memory unit.

[0194] Calculate the output gate activation value, use the Sigmoid function to process the combination of the current input and the hidden state of the previous time step, and determine the proportion of information that needs to be output in the current cell state.

[0195] The updated cell state vector is compressed to the range of -1 to 1 using the Tanh function, and then multiplied element-wise with the output gate activation value to generate the hidden layer state vector at the current time step.

[0196] By performing temporal state update operations through a collaborative gating mechanism of forget gate, input gate, and output gate, the continuous temporal features of the previous step are transformed into a hidden layer state sequence containing cross-segment dependencies, thereby achieving accurate capture and noise suppression of material damage evolution trajectory under dynamic environmental stress.

[0197] S5.4: Perform linear projection transformation on the state vector at the end of the hidden layer state sequence, and use a fully connected layer to map the high-dimensional hidden state to a low-dimensional physical meaning space to quantify the degree of material cumulative damage and output a single temporal health decay trend vector.

[0198] Obtain the hidden state sequence output by step S5.3, and extract the hidden state vector corresponding to the end of the current service cycle from the sequence as high-dimensional input data to characterize the cumulative damage features of the PETG bottle throughout the entire service cycle.

[0199] Construct a linear projection transformation module containing a single fully connected layer, and initialize trainable weight matrix and bias vector. The dimension of the weight matrix is ​​mapped from the number of hidden layer units to a low-dimensional physical space dimension, and the dimension of the bias vector is consistent with the dimension of the output space.

[0200] Perform matrix multiplication, dot product operation between the hidden state vector at the end time step and the weight matrix, and superimpose the bias vector to achieve a linear mapping from the high-dimensional abstract feature space to the low-dimensional damage quantization space, and output a single time-series health decay trend vector.

[0201] Apply a linear activation function to the linear mapping result or directly retain the linear output to ensure that the output value can continuously reflect the progressive damage evolution trajectory of the material from the initial intact state to the critical failure state.

[0202] By using linear projection transformation, the high-dimensional hidden state from the previous step is transformed into a single temporal health decay trend vector with clear physical meaning, thereby achieving accurate quantification of the cumulative damage level of PETG bottles and providing a direct basis for subsequent remaining life calculation.

[0203] For example, the hidden layer state vector is set to have a dimension of 32, and the output health decay trend vector has a dimension of 1. The weight matrix W is 32×1 dimension, and the bias b is a scalar. When the terminal hidden state h T Upon input, the calculated D value was 0.75. This value falls between 0 and 1, where 0 represents a brand-new state and 1 represents complete failure. A value of 0.75 indicates that the PETG bottle is in a severely damaged stage, approaching the failure threshold. Compared to the gradient saturation or information loss caused by traditional nonlinear activation, linear projection maintains the monotonicity and interpretability of damage evolution, significantly improving the response sensitivity and assessment accuracy of the lifetime prediction model in dynamic environments.

[0204] Step S6: Based on the parameters of the life prediction model trained using a single time-series health degradation trend vector and the course learning strategy, calculate the predicted remaining service life of the PETG bottle under the current dynamic environmental stress, and generate a crack resistance life assessment report including the failure risk level and prediction confidence level. Specifically, this includes:

[0205] S6.1: Obtain the weight matrix and bias vector of the fully connected regression layer in the parameters of the life prediction model obtained from the course learning strategy training, and perform a linear mapping operation with the single time-series health decline trend vector as input data to generate a normalized damage index scalar that characterizes the current cumulative damage level of the PETG bottle.

[0206] The system receives a single time-series health decay trend vector output from step S5, which characterizes the cumulative damage state of the PETG bottle under the current dynamic environmental stress sequence. It then calls upon a pre-stored lifetime prediction model parameter library to extract the weight matrix and bias vector of the fully connected regression layer, optimized by the course learning strategy, as the basis for calculating the linear mapping.

[0207] A linear regression calculation logic is constructed, which performs matrix multiplication on the input health decline trend vector and weight matrix, and superimposes the bias vector to quantify the current degree of damage, and outputs a normalized damage index scalar.

[0208] The calculated damage index is processed using a Sigmoid activation function, mapping the values ​​to a range of 0 to 1. This ensures the consistency of the physical meaning of the damage index, where 0 represents a completely new state and 1 represents a complete failure state. Nonlinear activation suppresses outlier fluctuations, enhancing the stability of the damage assessment.

[0209] The final output is a normalized damage index scalar, which accurately reflects the real-time cumulative damage level of the PETG bottle under the coupled effects of complex dynamic environment. This provides a standardized input data basis for the subsequent inverse-proportional degradation rate conversion of remaining life, realizing the accurate transformation from abstract feature space to specific physical damage index.

[0210] For example, the weight matrix W of the fully connected regression layer is set to the transpose of [0.85, -0.12, 0.05], and the bias vector b is 0.02. When the input single-time-series health decay trend vector H is [0.6, 0.3, 0.1], the matrix product is first calculated to obtain 0.85*0.6+(-0.12)0.3+0.050.1=0.51-0.036+0.005=0.479. Then, the bias value is added to obtain 0.479+0.02=0.499. After processing by the Sigmoid function, the normalized damage index D is approximately 0.623. This value indicates that the PETG bottle is currently in a moderate damage stage, significantly improving the resolution of damage quantification, avoiding the boundary truncation error caused by traditional linear mapping, and ensuring the accuracy of subsequent lifetime prediction.

[0211] S6.2: Based on the normalized damage index scalar combined with the critical failure threshold database of PETG materials, the current damage evolution trajectory is projected into the time domain using the inverse proportional degradation rate conversion algorithm to calculate the time series of the predicted remaining service life of the PETG bottle under the current dynamic environmental stress.

[0212] Obtain the normalized damage index scalar and the degradation rate benchmark parameters corresponding to the current service stage from the PETG material critical failure threshold database.

[0213] The inverse proportional degradation rate conversion algorithm is invoked to calculate the instantaneous degradation rate coefficient based on the tangent slope of the current damage evolution trajectory.

[0214] The instantaneous degradation rate coefficient is calculated using the following formula:

[0215]

[0216] in, for The instantaneous degradation rate at any given moment, for Normalized damage index at time step This represents the sampling time interval.

[0217] Based on the nonlinear cumulative damage characteristics of PETG materials under dynamic environments, an inverse proportional remaining lifetime prediction model is constructed.

[0218] The remaining useful life prediction is calculated using the following formula for time-domain projection processing:

[0219]

[0220] in, This is the predicted remaining useful life. This represents the critical failure threshold for PETG materials. The environmental stress correction factor is derived by weighting the mean stress amplitude and the slope of change of the current semantic segment. for Normalized damage index at time step.

[0221] The calculated remaining useful life predictions are mapped to a discrete time axis to generate a time series of remaining useful life predictions covering multiple future time steps.

[0222] Smoothing filtering is applied to the predicted time series to eliminate instantaneous fluctuations caused by sensor noise and ensure the monotonically decreasing characteristics of the evolution trajectory.

[0223] By using an inverse proportional degradation rate conversion algorithm, the normalized damage index scalar from the previous step is transformed into a time series of predicted remaining service life, enabling accurate quantitative assessment of the crack resistance life of PETG bottles under dynamic conditions.

[0224] For example, a critical failure threshold D for PETG bottles is set. crit The normalized damage exponent D(t) at the current time t is 0.42, the sampling interval Δt is 300 seconds, and the damage exponent D(t+Δt) at the next time step is 0.425. The calculated instantaneous degradation rate v(t) is 1.67 × 10⁻⁶. -5 / s. If the current environmental stress correction factor α(E) is calculated to be 0.9 after weighting, the remaining service life T can be calculated by substituting it into the formula. RUL The process takes approximately 25,628 seconds. This process effectively overcomes the error accumulation problem of static models under scenarios of sudden temperature and humidity changes, and significantly improves the prediction robustness under dynamic service conditions.

[0225] S6.3: Obtain the time series of the remaining service life prediction value and call the preset risk classification judgment rule library. Discretize and classify the prediction results according to the crack propagation rate range corresponding to different service stages to generate a failure risk level label that identifies the current safety status.

[0226] The system obtains the time series of predicted remaining service life and calls upon a pre-built risk grading rule base. This rule base is constructed based on the crack propagation kinetics model of PETG material and divides the service life into three stages: safe period, early warning period, and dangerous period.

[0227] Based on the crack propagation rate range corresponding to different service stages, the prediction results are discretized and classified. The crack propagation risk index is quantified based on the difference between the normalized damage index at the current time and the previous time, the predicted remaining service life, and the critical failure threshold of PETG material, and this index is compared with a preset threshold.

[0228] The calculated crack propagation risk index R is compared with a preset threshold. When R is less than 0.05, the crack is determined to be in the steady-state slow propagation stage and is marked as a green safety level; when R is between 0.05 and 0.15, the crack is determined to be in the initial stage of accelerated propagation and is marked as a yellow warning level; when R is greater than 0.15, the crack is determined to be in the unstable rapid propagation stage and is marked as a red danger level.

[0229] A failure risk level label identifying the current safety status is generated, and this label is linked to the remaining service life prediction value. Through the above discretization classification process, continuous service life prediction values ​​are transformed into intuitive risk level indicators, enabling real-time hierarchical monitoring and accurate early warning of the crack resistance safety status of PETG bottles.

[0230] For example, the critical failure threshold for PETG material is set to 1.0. In a certain monitoring, the normalized damage index at the current moment is 0.82, and at the previous moment it was 0.80, with a difference of 0.02. The predicted remaining service life is 48 hours. After quantifying the risk index (the difference divided by the product of the remaining service life and the critical failure threshold), the risk index is approximately 0.0004167, which is less than 0.05. The system determines that the current level is green and safe, and outputs the label "Level_1_Safe". If, in another scenario, the normalized damage index suddenly increases to 0.98 at the current moment, compared to 0.90 at the previous moment, the difference increases to 0.08, and the remaining service life is shortened to 0.48 hours (approximately 29 minutes), the calculated risk index is 0.1667, which is greater than 0.15. The system determines the level to red and outputs the label "Level_3_Danger", indicating that immediate intervention is required.

[0231] S6.4: Based on the variance distribution characteristics of a single time-series health decay trend vector and the error statistics model of the historical validation set, Monte Carlo Dropout sampling technology is used to perform multiple forward inference operations to quantify the uncertainty range of the model output and generate a numerical index of prediction confidence.

[0232] Obtain the single time-series health decay trend vector output by S5.4 and use it as the input baseline data for Monte Carlo Dropout uncertainty quantization.

[0233] During the inference phase, the activation state of the Dropout layer in the model is maintained, the dropout probability is set to 0.2, and N independent forward propagation samplings are performed on the same input vector.

[0234] Each sampling randomly masks some neuron connections to simulate the predicted response under different sub-network structures, generating a set of remaining lifetime distributions containing N predicted values.

[0235] Calculate the mean and variance of this distribution set, where the mean represents the best estimated lifetime and the variance characterizes the uncertainty of the model's understanding of the current dynamic environmental stress.

[0236] The prediction confidence index is calculated using the following formula:

[0237]

[0238] in, To normalize the confidence level, The standard deviation of the predicted value. To predict the mean, To prevent division by zero smoothing constants.

[0239] High uncertainty intervals are identified through analysis of variance. When the confidence level is lower than the preset threshold of 0.85, the prediction result is marked as low confidence and a manual review mechanism is triggered.

[0240] By using Monte Carlo sampling and statistical inference, the point estimation results are transformed into probability distribution indicators with confidence intervals, enabling a quantitative assessment of prediction reliability in dynamic environments.

[0241] For example, let's set the sampling number N to 50 times and the discard probability to 0.2. Inferring the decay vector of a PETG bottle after experiencing an 'impact cluster' yields 50 predicted remaining lifetime values, with a mean of 1200 hours and a standard deviation of 150 hours. Substituting these values ​​into the formula, and using a smoothing constant ε of 1e-6, we obtain a confidence level C of approximately 0.984. If, in another scenario, the standard deviation increases to 400 hours, the confidence level drops to 0.901. The system automatically adjusts the risk level weight based on the confidence level value; high confidence levels result in a direct report output, while low confidence levels increase the safety redundancy coefficient, significantly improving decision-making robustness under complex conditions.

[0242] S6.5: Obtain the time series of remaining useful life prediction values, failure risk level labels, and prediction confidence indexes. Use a structured report encapsulation engine to assemble the above multi-source assessment data in a standardized format to generate a crack resistance life assessment report containing complete failure risk levels and prediction confidence.

[0243] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0244] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0245] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the crack resistance life of PETG bottles, characterized in that, include: Multi-source environmental time-series data during the service life of PETG bottles are acquired, and semantic segmentation is performed on the multi-source environmental time-series data based on the environmental response characteristics of PETG material to generate a set of stress segments with semantic tags. Based on the set of stress segments, a stress correlation graph structure is constructed. Each stress segment is mapped to a graph node and given node attributes that include stress statistical features and inter-segment correlation features. Directed edges are established between segments that satisfy temporal proximity and physical coupling, and the stress mutation degree between segments is quantified to generate a stress correlation graph object with topological connection relationship. The semantic labels and node attributes in the stress correlation graph object are fused and encoded to generate a feature matrix representing the initial state of each stress segment. Based on the feature matrix, a graph topology propagation operation is performed. During the neighborhood information aggregation process, the intensity of adjacency information injection is dynamically adjusted according to the material response priority corresponding to the semantic type of each fragment, thereby generating a propagation feature matrix that has been adaptively weighted and optimized. A temporal order constraint is applied to the propagation feature matrix, and the segments are processed in an orderly manner according to their occurrence order. Cross-segment dependency features are extracted, and a temporal feature vector characterizing the material health degradation trend is output. Based on the time-series feature vectors and the parameters of the life prediction model obtained through training, the remaining life of PETG bottles under the current environmental stress is calculated, and an assessment result including risk level and confidence level is generated.

2. The method for predicting the crack resistance life of a PETG bottle according to claim 1, characterized in that, Based on the set of stress segments, a stress correlation graph structure is constructed. Each stress segment is mapped to a graph node and assigned node attributes containing stress statistical characteristics and inter-segment correlation characteristics. Directed edges are established between segments that satisfy temporal proximity and physical coupling, and the degree of stress abrupt change between segments is quantified. A stress correlation graph object with topological connections is generated, including: Each stress semantic segment in the stress segment set is taken as the processing object. Based on the statistical distribution characteristics of the temperature sequence, humidity sequence and vibration acceleration sequence inside the segment, four key indicators are calculated respectively: mean stress amplitude, slope of change, duration and adjacent segment transition entropy. The above four key indicators are combined and encapsulated to generate a node attribute vector containing complete physical state information. The system obtains the timestamp information of all generated node attribute vectors and their corresponding stress semantic segments, filters out candidate segment pairs with time intervals less than the preset threshold based on the preset time interval threshold, and further judges whether there is a specific physical coupling relationship between the candidate segment pairs that is high humidity retention pointing to condensation cycle or impact cluster pointing to high temperature slow change based on the physical coupling relationship rule library, thereby generating a set of potential connection edges that meet the spatiotemporal constraints. Obtain the node attribute vector of each pair of connected stress semantic segments in the potential connection edge set, use the Jensen-Shannon divergence algorithm to calculate the degree of difference between the two node attribute vectors in the probability distribution, quantify the calculated degree of difference into edge weight values ​​that characterize the severity of the sudden change in the stress parameters of the two segments, and generate weighted directed edge data with quantitative weight information. The process involves obtaining the mapped stress semantic fragments as graph nodes and the weighted directed edge data with quantitative weight information as graph edge connections. Then, a graph structure assembly operation is performed to integrate the graph nodes and weighted directed edge data with the topological connection logic in chronological order, ultimately generating a stress semantic graph object with complete topological connections and containing node attributes and edge weight information.

3. The method for predicting the crack resistance life of a PETG bottle according to claim 1, characterized in that, Based on the feature matrix, a graph topology propagation operation is performed. During the neighborhood information aggregation process, the intensity of adjacency information injection is dynamically adjusted according to the material response priority corresponding to the semantic type of each fragment, generating an adaptively weighted and optimized propagation feature matrix, including: Based on the discrete semantic labels in the semantic embedding node representation matrix, the pre-set PETG material failure mechanism knowledge base is called for matching query, and the crack initiation weight coefficient and crystal region relaxation weight coefficient corresponding to each semantic type are extracted to generate a material response priority vector set containing material response priority quantification index. Using the material response priority vector set as a control factor, a nonlinear gating function operation is performed on the feature vectors of adjacent nodes within the second-order neighborhood range to calculate an adaptive gating coefficient matrix that reflects the degree of acceptance of neighbor information by the current node, so as to realize dynamic filtering of information flow under different physical coupling relationships. Based on the adaptive gating coefficient matrix and the edge weights of the original graph structure, the semantic embedded node representations of adjacent nodes are subjected to element-wise weighted aggregation processing to generate an intermediate-state neighborhood aggregation feature tensor that integrates local stress sensitivity features, thereby completing information transmission and noise suppression under the second-order topological structure. The intermediate-state neighborhood aggregated feature tensor is fused with the initial semantic embedded node representation of the current node using a residual connection operation, and then dimensional alignment is performed through a linear transformation layer. The resulting topology propagation node representation matrix, which is finally optimized by local stress sensitivity adaptive weighting, is used as input data for subsequent temporal joint pooling.

4. The method for predicting the crack resistance life of a PETG bottle according to claim 1, characterized in that, Multi-source environmental time-series data should include at least temperature, humidity and vibration acceleration sequences.

5. The method for predicting the crack resistance life of a PETG bottle according to claim 1, characterized in that, The semantic segmentation process includes: The temperature sequence is initially screened for high-temperature gradual change segment and condensation cycle segment using the glass transition temperature range determination logic of PETG material. The humidity sequence is identified for high humidity retention segment according to the moisture absorption induced plasticity threshold rule. The spectral energy mutation feature of the vibration acceleration sequence is extracted to generate impact cluster candidate segments. The ultraviolet radiation superposition determination rule is introduced to perform multi-physics field coupling analysis to merge and generate a complete stress segment set containing six basic semantic tags.

6. The method for predicting the crack resistance life of a PETG bottle according to claim 1, characterized in that, The fusion encoding includes using a learnable discrete label encoder to perform vector mapping on the semantic labels and fusing the mapping result with the node attributes.

7. The method for predicting the crack resistance life of a PETG bottle according to claim 1, characterized in that, The temporal order constraint includes using long short-term memory network units to update the propagation feature matrix in a temporal state.

8. The method for predicting the crack resistance life of a PETG bottle according to claim 1, characterized in that, The physical coupling relationships include a first coupling relationship between the high-humidity retention segment and the condensation cycle segment, and a second coupling relationship between the impact cluster segment and the high-temperature gradual change segment.

9. The method for predicting the crack resistance life of a PETG bottle according to claim 1, characterized in that, The adjacency information injection intensity is determined based on the material response priority corresponding to the semantic type to which the stress fragment belongs, and the material response priority is determined by the category of the semantic tag.

10. The method for predicting the crack resistance life of a PETG bottle according to claim 1, characterized in that, The risk level is obtained by classifying the time-series feature vector by calling a pre-set risk grading judgment rule base, and the confidence level is obtained by performing multiple forward inferences on the time-series feature vector based on Monte Carlo Dropout sampling technology.

Images