Tracing analysis method for blank bottom whitening and yellowing defect in PET bottle blank production

Abstract

The invention provides a PET bottle embryo production embryo bottom whitening and yellowing defect traceability analysis method, which comprises the following steps: synchronous acquisition and high-precision timestamp binding of multichannel original sensing signals, systematic denoising, normalization and local contrast sensing mapping, and dynamic representation of sensing activation intensity is realized; a cross-channel perception consistency index is constructed through calculation of sliding window mutual information and a phase locking value, and collaborative optimization adjustment of gain and time sequence is realized by combining a double-layer feedback structure and a distributed adaptive controller and utilizing an LSTM prediction and online reinforcement learning mechanism. The robustness of the system to interference and dynamic change is enhanced, and high-reliability intelligent self-control of the multi-channel biological sensing system is realized.

Description

Technical Field

[0001] This invention relates to the field of quality analysis technology in polymer material processing and intelligent manufacturing, and in particular to a method for tracing the source of defects such as whitening and yellowing at the bottom of PET preforms. Background Technology

[0002] In the current field of polymer material processing and injection molding intelligent manufacturing, process quality analysis and defect attribution technology are increasingly becoming important technical directions for preform quality control, digital process optimization, and intelligent equipment operation and maintenance, particularly for defects such as whitening and yellowing at the bottom of PET preforms. Traditional production processes largely rely on experienced process engineers to adjust molding parameters, material batches, and equipment operating status based on their accumulated production experience and statistical data to reduce quality fluctuations. However, with the increasing automation and intelligence of PET preform production lines, the scale and complexity of the production data involved have increased dramatically. The complex coupling between different batches, unlabeled, and multi-source process, material, and equipment variables is becoming increasingly prominent. Attribution paths relying on manual experience are no longer sufficient to accurately pinpoint the root cause of defects and cannot meet the quality control requirements of large-scale, multi-batch, and complex variable coupling. Current technologies are based on statistical quality management, correlation analysis, and empirical rule bases. Common engineering practices involve regression analysis, single-factor hypothesis testing, or multiple linear regression on production batch data and defect detection results. Some intelligent manufacturing systems have introduced end-to-end black-box machine learning models, such as using deep neural networks to directly identify defects and predict root causes from multi-source raw data. Simultaneously, some research attempts to establish causal chains between equipment, materials, and processes using knowledge graphs to achieve certain inference and defect cause induction. Some advanced manufacturers are using big data platforms and MES (Manufacturing Execution System) for traceable management and statistical quantitative monitoring of the production process, enabling source analysis and risk warnings for defect occurrences. However, in real-world scenarios with frequent operational disturbances, missing data, or the influence of latent variables, problems such as broken causal chains or incomplete attribution can easily arise. While end-to-end deep learning defect detection models have a certain level of automation, their attribution process lacks an interpretable and transparent path, making it difficult to meet the requirements of traceable reasoning and physical semantics in process optimization. Furthermore, they often rely on large-scale pre-labeled samples, resulting in a high barrier to practical deployment. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method for tracing and analyzing the defects of white and yellowing bottom of PET bottle preforms.

[0004] The technical solution of this invention is implemented as follows: A method for tracing and analyzing the source of defects such as whitening and yellowing at the bottom of PET preforms during production, comprising: S1: Acquire multi-source time series data related to the whitening and yellowing defects of the preform bottom during the PET injection molding process, including mold temperature, injection pressure, and cooling time in process parameters, raw material moisture content and melt viscosity in material properties, and screw speed and clamping force fluctuations in equipment status. Simultaneously collect optical inspection images of the bottom of each batch of preforms and their defect scoring results as the original input dataset for subsequent analysis. S2: Perform sliding window segmentation on the original input dataset, and calculate the time-varying mutual information values ​​between the three types of variable pairs: process-material, material-equipment, and equipment-process based on the segmented time series segments. Combined with the conditional Granger causality strength analysis method, extract the dynamic response consistency characteristics between each variable before and after the disturbance, and generate a dynamic coupling strength matrix that reflects the nonlinear cooperative change trend between variables. S3: Based on the dynamic coupling strength matrix, construct a ternary hypergraph topology with 'process-material-equipment' as nodes. Each node contains a set of sub-parameters in the corresponding dimension. The edge weights are determined by the corresponding coupling strength values ​​in the dynamic coupling strength matrix. This ternary hypergraph is used as a structured prior knowledge representation to support subsequent message passing and attribution reasoning. S4: The defect feature vector extracted from the optical detection image by the convolutional neural network is used as the initial message input to the corresponding node in the ternary hypergraph. Based on the attention mechanism-driven ternary message passing algorithm, the information flow related to the context is circulated and aggregated between the three nodes of process, material and equipment, and the implicit contribution vector of each dimension to the current defect mode is output. S5: Based on the implicit contribution vector, identify the dominant influence dimension, and use counterfactual attribution analysis technology to generate a virtual intervention scenario on this dimension. By simulating the change trajectory of the defect probability when the key parameters deviate from the standard range, form an interpretable set of causal inference paths. S6: Perform Bayesian confidence interval evaluation on the set of causal inference paths, and combine it with the attribution conclusions under similar working conditions matched in the historical similar defect case library to generate an attribution conclusion report with statistical significance labels and case support, as the final defect cause determination result; S7: Determine whether a single main cause with high confidence has been identified in the attribution conclusion report; if so, output the corresponding process optimization suggestion instruction; if not, trigger the multi-factor collaborative disturbance analysis process and call the incremental data acquisition mechanism to supplement the observation samples under specific working conditions to enhance the model's discrimination ability. S8: Store the dynamic coupling strength matrix update record, attribution conclusion report and feedback execution result of this reasoning process into the process knowledge evolution database, and perform periodic calibration of the ternary hypergraph structure and message passing weights based on the newly accumulated data to achieve closed-loop iterative optimization of the defect attribution model.

[0005] The present invention provides a method for tracing and analyzing the defects of white and yellow bottom in PET bottle preform production, which has the following beneficial effects: (1) This invention significantly improves the accuracy and physical interpretability of defect root cause localization by constructing a progressive analysis architecture of "dynamic coupling strength perception - ternary collaborative reasoning - attribution interpretability enhancement". By introducing a sliding window mutual information and conditional Granger causality fusion mechanism and designing a "perturbation response consistency" metric function, it can effectively identify the transient collaborative change trend between key parameters without pre-setting the causal direction, accurately capture the system-level response pattern caused by small perturbations, thereby overcoming the problem of misjudgment caused by static correlation analysis, and providing reliable and dynamically updated prior structure support for subsequent high-order reasoning; (2) At the inference modeling level, this invention proposes a ternary collaborative inference engine driven by hypergraph topology and attention mechanism, which realizes fine-grained analysis of nonlinear interactions between cross-domain parameters. A hypergraph structure with physical semantics is constructed with "process-material-equipment" as the three nodes. The edge weights are dynamically assigned by the time-varying coupling strength extracted in the previous stage to ensure that the network topology reflects the influence path strength in the real production environment. A message passing algorithm based on attention mechanism is designed so that the defect feature vector is cyclically propagated between the three domain nodes and adaptively focuses on the most contributing information channel. This not only enhances the model's sensitivity to key abnormal paths, but also realizes the quantitative ranking output of multi-dimensional contributions. This mechanism greatly improves the attribution accuracy of multi-factor coupled defects under complex working conditions, especially when facing new defect patterns or parameter offset combinations, it shows stronger generalization ability and robustness. (3) This invention integrates a dual verification mechanism of counterfactual attribution analysis and historical case matching to construct an attribution conclusion output system with statistical credibility and operational guidance. By generating virtual intervention scenarios (such as "if the moisture content of raw materials decreases to 0.02% and the mold temperature remains stable, the defect probability decreases by 76%), and combining Bayesian confidence intervals to evaluate the significance of inferences, the attribution results are no longer isolated probability judgments, but operational suggestions with clear intervention directions. At the same time, a similar case retrieval module is introduced to cross-validate the current inference using historical successful handling records, giving each attribution conclusion a traceable confidence label. The entire process does not require additional sensor deployment or large-scale labeled sample accumulation, is compatible with the existing production line data acquisition level, and has good engineering feasibility. The resulting closed-loop adaptive attribution system avoids the limitations of poor scalability of traditional rule systems and makes up for the lack of interpretability of pure data-driven models, providing a scientific and practical intelligent decision support tool for process optimization. Attached Figure Description

[0006] Figure 1This is a flowchart of a method for tracing and analyzing the source of defects such as whitening and yellowing at the bottom of PET preforms produced according to the present invention; Figure 2 This is a sub-flowchart of a method for tracing and analyzing the source of defects such as whitening and yellowing at the bottom of PET preforms produced according to the present invention; Figure 3 This is another sub-flowchart of the present invention, which is a method for tracing and analyzing the defects of white and yellow bottom in PET bottle preform production. Detailed Implementation

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

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

[0009] like Figure 1 As shown, this invention provides a method for tracing and analyzing the source of defects such as whitening and yellowing at the bottom of PET preforms during production, specifically including: S1: Acquire multi-source time series data related to the whitening and yellowing defects of the preform bottom during the PET injection molding process, including mold temperature, injection pressure, and cooling time in process parameters, raw material moisture content and melt viscosity in material properties, and screw speed and clamping force fluctuations in equipment status. Simultaneously collect optical inspection images of the bottom of each batch of preforms and their defect scoring results as the original input dataset for subsequent analysis. S2: Perform sliding window segmentation on the original input dataset, and calculate the time-varying mutual information values ​​between the three types of variable pairs: process-material, material-equipment, and equipment-process based on the segmented time series segments. Combined with the conditional Granger causality strength analysis method, extract the dynamic response consistency characteristics between each variable before and after the disturbance, and generate a dynamic coupling strength matrix that reflects the nonlinear cooperative change trend between variables. S3: Based on the dynamic coupling strength matrix, construct a ternary hypergraph topology with 'process-material-equipment' as nodes. Each node contains a set of sub-parameters in the corresponding dimension. The edge weights are determined by the corresponding coupling strength values ​​in the dynamic coupling strength matrix. This ternary hypergraph is used as a structured prior knowledge representation to support subsequent message passing and attribution reasoning. S4: The defect feature vector extracted from the optical detection image by the convolutional neural network is used as the initial message input to the corresponding node in the ternary hypergraph. Based on the attention mechanism-driven ternary message passing algorithm, the information flow related to the context is circulated and aggregated between the three nodes of process, material and equipment, and the implicit contribution vector of each dimension to the current defect mode is output. S5: Based on the implicit contribution vector, identify the dominant influence dimension, and use counterfactual attribution analysis technology to generate a virtual intervention scenario on this dimension. By simulating the change trajectory of the defect probability when the key parameters deviate from the standard range, form an interpretable set of causal inference paths. S6: Perform Bayesian confidence interval evaluation on the set of causal inference paths, and combine it with the attribution conclusions under similar working conditions matched in the historical similar defect case library to generate an attribution conclusion report with statistical significance labels and case support, as the final defect cause determination result; S7: Determine whether a single main cause with high confidence has been identified in the attribution conclusion report; if so, output the corresponding process optimization suggestion instruction; if not, trigger the multi-factor collaborative disturbance analysis process and call the incremental data acquisition mechanism to supplement the observation samples under specific working conditions to enhance the model's discrimination ability. S8: Store the dynamic coupling strength matrix update record, attribution conclusion report and feedback execution result of this reasoning process into the process knowledge evolution database, and perform periodic calibration of the ternary hypergraph structure and message passing weights based on the newly accumulated data to achieve closed-loop iterative optimization of the defect attribution model.

[0010] Step S1: Acquire multi-source time-series data related to the whitening and yellowing defects at the bottom of the preform during the PET injection molding process. This includes process parameters such as mold temperature, injection pressure, and cooling time; material properties such as raw material moisture content and melt viscosity; and equipment status such as screw speed and clamping force fluctuations. Simultaneously, optical inspection images of the bottom of each batch of preforms and their defect scoring results are collected as the raw input dataset for subsequent analysis. Specifically, this includes: S1.1: Obtain time series data of process parameters related to preform quality in the PET injection molding production line, including mold temperature, injection pressure, and cooling time. Based on the real-time acquisition of process variables within the execution cycle of the industrial sensor system, synchronously record them at a standard sampling frequency to obtain the raw process data stream, which serves as the input condition for the process dimension. Based on the real-time monitoring environment of the PET injection molding production line, mold temperature, injection pressure and cooling time are taken as the core process variables input conditions in the process dimension, all of which come from the raw signal output of a high-precision industrial sensor array. Multi-channel thermocouples are used to collect temperature data at key locations of the mold (parameters: sampling frequency 1 Hz, resolution 0.1℃) to achieve a complete record of the thermal field distribution of the mold. The temperature values ​​of each channel are synchronized to the data acquisition server within the PLC control cycle to construct a continuous temperature time series. Furthermore, the pressure curve of the injection cylinder is monitored by a pressure sensor module (range 0~200 MPa, accuracy ±0.5%FS). A low-pass FIR filter (cutoff frequency is 0.45 times the sampling frequency) in the digital signal processing algorithm is used to remove high-frequency noise, thereby realizing the extraction of a stable injection pressure signal and recording it as a high-fidelity pressure time series. Furthermore, the cooling time monitoring unit obtains the process cycle length from the completion of mold cavity filling to demolding based on the production line logic control signal, and uses the event-triggered timestamp differential method to calculate the duration of the cooling stage of a single cycle, generating a cooling time sequence corresponding to the batch. By using a data time synchronization mechanism, the three types of process parameters—temperature, pressure, and cooling time—are aligned on the same sampling time axis. The missing points are filled in using a linear interpolation method to ensure that the time series length of the three types of parameters is completely consistent with the timestamp. Through the above three measurement and processing methods, the raw sensor signals acquired in real time are transformed into noise-free, aligned, and computable process parameter time series, thereby achieving accurate quantification of the process dimension status during production. For example, during the production of a certain batch of PET preforms, thermocouples were deployed at 8 measurement points with a sampling frequency of 1 Hz. During the mold temperature stabilization phase, a temperature field distribution with an average value of 278.6℃ was detected. The pressure sensor sampling frequency was 10 Hz, and the FIR filter length was 21 points. After filtering, the average injection pressure sequence was 135.4 MPa, with a peak value of 142.7 MPa. The cooling time monitoring unit recorded an average cooling time of 15.2 seconds and a standard deviation of 0.4 seconds for this batch. The three types of time series data were linearly interpolated and aligned to the 1 Hz time axis, resulting in a total sequence length of 900 points (corresponding to a 900-second production cycle). Using this data, it can be directly input into the dynamic coupling strength calculation module, significantly improving the matching accuracy and interpretability of process dimension variables in the causal analysis of this batch. S1.2: Obtain time series data of material parameters related to raw material properties, including raw material moisture content and melt viscosity. Based on the output signals of near-infrared moisture meter and online viscometer, measure the material state before each batch of material is fed. Use the data alignment mechanism to match the timestamps with the process data of the corresponding production cycle to generate a structured material property sequence as the input condition for the material dimension. The input conditions are the raw material moisture content and melt viscosity measurement signals before each batch of material is fed in the PET injection molding production line. The acquisition equipment includes a near-infrared moisture meter and an online viscometer. The output format is a material property data stream serialized by timestamp. The near-infrared moisture detection method (parameters: wavelength range selected as 900nm~1700nm, integration time 0.5s) is used to realize rapid non-destructive measurement of the moisture content inside raw material particles and output the initial value sequence of raw material moisture content. Furthermore, measurements were performed using an online rotational rheovis meter (parameters: rotor size Φ25mm, shear rate range 10~1000 s). -1 This allows for the real-time acquisition of the viscosity curve of the melt at a set temperature, and the generation of the original melt viscosity data sequence. Furthermore, a signal denoising algorithm (parameters: third-order low-pass Butterworth filter, cutoff frequency 0.8 times sampling frequency) is used to smooth the raw material moisture content and melt viscosity measurement signals in the time domain to reduce the influence of transient jitter and generate a steady-state material parameter sequence. Furthermore, a timestamp synchronization alignment algorithm is applied (parameters: maximum alignment deviation Δt is set to 0.1s, and the interpolation method is cubic spline interpolation) to achieve accurate registration of the material parameter sequence with the corresponding process parameter sequence in the production cycle in the time dimension, and obtain a batch-level matching index; Furthermore, based on the batch ID mapping table of the production management system, batch association operation is performed to encapsulate the aligned raw material moisture content and melt viscosity sequence into a structured material attribute sequence data object, whose fields include timestamp, batch ID, moisture content value, viscosity value and data quality label; Through the above timestamp alignment and signal denoising processing, the original measurement signal is transformed into a structured material property input that can be directly used for dynamic coupling analysis and subsequent ternary hypergraph modeling, thereby ensuring high precision and stability of material dimension data. For example, in a PET preform production batch number B20230512, the near-infrared moisture analyzer was set with a wavelength range of 900nm~1700nm, an integration time of 0.5s, and a sampling frequency of 10Hz. The measured raw material moisture content sequence showed an average value of 0.025 and a standard deviation of 0.001 within the time window [0s, 300s]. The online rotational rheovister was set with a shear rate of 100 s⁻¹. -1The operating temperature was 280℃, the sampling frequency was 5Hz, and the average value of the original melt viscosity sequence was 0.74 Pa·s, with a standard deviation of 0.02 Pa·s. A third-order low-pass Butterworth filter was applied to smooth both types of data, with the cutoff frequency set to 0.8 times the sampling frequency, significantly reducing the peak noise amplitude. Cubic spline interpolation was used to align the timestamps, with a maximum deviation of no more than 0.08s. The data was then registered with the corresponding process parameter sequence under the index B20230512. The final generated structured material property sequence contains 3000 records, each including a timestamp, batch ID, moisture content, viscosity, and mass label fields. This sequence can be directly entered into the dynamic coupling strength matrix calculation module. Verification results show that the time alignment accuracy of the material property input is significantly improved, signal fluctuations are reduced, and this effectively supports the improvement of subsequent causal modeling accuracy. S1.3: Obtain time series data of equipment parameters related to equipment operating status, including screw speed and clamping force fluctuation values. Based on the PLC controller and vibration sensor, collect dynamic response signals of key components of the injection molding machine, use sliding mean filtering to remove high-frequency noise interference, extract the effective amplitude sequence of steady state segment, form equipment status characterization data, and use it as input conditions for the equipment dimension. S1.4: Simultaneously acquire high-resolution optical inspection images of the bottom area of ​​each batch of preforms, trigger a line scan camera at a fixed position on the conveyor belt based on the machine vision system to capture images with uniform grayscale distribution, and combine the image quality discrimination algorithm to exclude blurry or occluded samples, generating an effective image set that can be used for defect feature extraction. S1.5: Based on expert annotation and automated scoring model, generate defect scoring results for each batch of preforms. Apply a pre-trained convolutional neural network to the optical inspection image to extract local texture abnormality features, output a whitening and yellowing defect severity score in the range of 0 to 1, and integrate the score with the aforementioned three types of time series data of process, material and equipment according to batch ID to form a unified indexed original input dataset.

[0011] Step S2: The original input dataset is segmented using a sliding window. Based on the segmented time-series segments, the time-varying mutual information values ​​between three types of variable pairs—process and materials, materials and equipment, and equipment and process—are calculated. Combined with conditional Granger causality strength analysis, the dynamic response consistency characteristics between variables before and after the disturbance are extracted, generating a dynamic coupling strength matrix reflecting the nonlinear cooperative change trend between variables. Specifically, this includes: S2.1: Perform sliding window segmentation processing on the acquired multi-source time series data. Based on the preset time window length and step size, divide the key variables of process parameters, material properties and equipment status in the PET injection molding process into overlapping time segments to capture local dynamic characteristics and generate a set of time series segments for time-varying coupling analysis. For the multi-source time series data such as process parameters, material properties and equipment status collected in step S1, a sliding window segmentation method (parameters: time window length W, step size S, window overlap rate R) is adopted to divide the continuous time series data into a set of discrete segments with local dynamic characteristics. Furthermore, through the time window parameter setting module (based on: production cycle time and critical event response time), variables such as mold temperature, injection pressure, cooling time, raw material moisture content, melt viscosity, screw speed and clamping force fluctuation are synchronized in segments according to a unified time base, so that the timestamps of different dimensions of variables in each segment are consistent, thereby ensuring the time alignment accuracy of subsequent coupling strength calculation; Furthermore, a normalization preprocessing algorithm (parameters: mean μ, variance σ) is adopted to standardize the amplitude of each variable after segmentation, eliminate the influence of differences in different dimensions and ranges on local fluctuation analysis, and obtain a set of time segments with normalized amplitude. Furthermore, through a sliding window vectorization process, the multidimensional original sampling point sequence within each time window is mapped into a matrix form to support mutual information and Granger causality modeling; the matrix rows correspond to time sampling points, and the columns correspond to the normalized amplitude sequence of each variable, outputting a structured time series segment matrix set; Through the window boundary marking and fragment index construction process, the time-series fragment matrix of the previous step is transformed into a structured set with batch index, window number, and variable label, realizing the time-varying coupling analysis input that can be directly called by S2.2 to S2.4; For example, in a certain operating condition of a PET preform production line, the time window length W is set to... seconds, step size S is seconds, window overlap rate R is %. For the mold temperature sequence (sampling frequency) Hz, measurement range (℃) Perform segmented processing to obtain the contents of each window. Each sampling point. For the injection pressure sequence (sampling frequency) Hz, range MPa), raw material moisture content sequence (sampling frequency) Hz, range %), screw speed sequence (sampling frequency) Hz, range All variables (rpm, etc.) are segmented and aligned according to the same time window length and step size. After Z-score normalization, the mean and variance of each variable within the window are unified to a zero-mean and unit-variable scale. For example, one row of the mold temperature segment matrix corresponds to a temperature sampling point at a certain second, and the column values ​​are the standardized temperature amplitudes. The segmented matrices of all variables are then compiled using window numbers and batch ID indices to form a matrix containing... The set of time-series segments of window segments exhibits significant local dynamic change patterns in subsequent mutual information and causal strength analysis, effectively supporting the extraction of multidimensional coupling features and dynamic response consistency measurement. S2.2: Based on the set of time segments, the sliding window mutual information algorithm is used to calculate the nonlinear correlation between the time series data of process parameters and material properties, so as to obtain the time-varying mutual information value of the process-material variable pair in different time periods, so as to quantify the degree of information sharing between the two in the local time window and output the process-material time-varying mutual information sequence. S2.3: Based on the aforementioned time series fragment set, the conditional Granger causality test method is used to perform causal strength analysis on the material property and equipment state variable pair. After controlling for the influence of other covariates, it is determined whether there is a statistically significant predictive causal relationship, and its F-statistic value is calculated as the conditional causal strength index of the material-equipment variable pair, generating a material-equipment time-varying causal strength sequence. S2.4: Based on the set of time segments, conditional Granger causality modeling is performed on the cross-direction of equipment state and process parameter variables. It is then identified whether equipment fluctuations before and after disturbance events (such as parameter mutations or alarm triggers) cause process deviations or vice versa. Combining the directional causal strength and the magnitude of mutual information change, a two-way time-varying coupling index sequence of equipment-process is constructed. S2.5: Integrate the time-varying mutual information sequence of process-material, the time-varying causal strength sequence of material-equipment, and the bidirectional time-varying coupling index sequence of equipment-process. Generate a three-dimensional coupling strength vector based on a normalized weighted fusion strategy, and arrange it into a symmetric matrix form to form a dynamic coupling strength matrix that reflects the nonlinear co-evolution relationship among process, material, and equipment. This matrix serves as the structured prior input for the construction of the ternary hypergraph in the next stage. The process-material time-varying mutual information sequence, the material-equipment time-varying causal strength sequence, and the equipment-process bidirectional time-varying coupling index sequence are received as input conditions. All of these are time-series numerical sequences obtained through sliding window segmentation and feature calculation. The normalization method (parameter: min-max scaling) is used to linearly scale the values ​​of each time period of the above three types of coupled sequences to achieve the unification of the units of different metrics and the standardization of the numerical range. Furthermore, through a weighted fusion algorithm (parameter: set of weight factors) , , This achieves a linear combination of the three types of normalized coupling values ​​and yields the three-dimensional coupling strength vector corresponding to each time slice, as shown in the following formula:

[0012] in, Let be the three-dimensional coupling strength vector for time slice t. The process-material mutual information value for time slice t. The material-device causal strength value for time slice t. The device-process bidirectional coupling value for time slice t; Furthermore, by using a matrix construction method, the three-dimensional coupling strength vector of each time slice is arranged among the three axes of process (P), material (M), and equipment (E) to generate a symmetric dynamic coupling strength matrix. The matrix elements... This represents the normalized weighted coupling value between dimension X and dimension Y; Furthermore, by employing time alignment and matrix serialization techniques, the coupling strength matrices corresponding to consecutive time slices are spliced ​​into a set of time-series matrices, thereby achieving a global characterization of the nonlinear co-evolutionary relationship among process, materials, and equipment. By using normalized weighted fusion and matrix construction, the results of the multi-type coupling sequence in the previous step are transformed into a set of structured dynamic coupling strength matrices, thus providing optimized prior input for the construction of ternary hypergraphs. For example, during the production batch number B202311 of PET preforms, the normalized range of the process-material mutual information value sequence is [0,1], the normalized range of the material-equipment causality strength value is [0,1], the normalized range of the equipment-process coupling value is [0,1], and weighting factors are selected. =0.4, =0.35, =0.25. At time slice t=15, the normalized values ​​are respectively =0.72, =0.65, =0.58, substitute into the formula: Calculations yielded =(0.288, 0.2275, 0.145). Arranging this vector into a symmetric matrix form: process-material unit value = 0.288, material-equipment unit value = 0.2275, equipment-process unit value = 0.145, with other matrix elements filled symmetrically. After performing time serialization, a set of dynamic coupling strength matrices containing 35 time slices is obtained. In the subsequent ternary hypergraph construction, this significantly improves the interpretability and accuracy of defect attribution calculations for batch B202311.

[0013] like Figure 2As shown, step S3 involves constructing a ternary hypergraph topology with 'process-material-equipment' as nodes based on the dynamic coupling strength matrix. Each node contains a set of sub-parameters for its corresponding dimension, and the edge weights are determined by the corresponding coupling strength values ​​in the dynamic coupling strength matrix. This ternary hypergraph serves as a structured prior knowledge representation supporting subsequent message passing and attribution reasoning. Specifically, this includes: S3.1: Based on the dynamic coupling strength matrix output by S2, extract the time-varying coupling strength values ​​between three types of variable pairs: process-material, material-equipment, and equipment-process, and use them as the initial weight input for cross-dimensional edges in the ternary hypergraph; use the response consistency features obtained by jointly calculating the sliding window mutual information and conditional Granger causality to form a weighted set of relationships with directional perception capabilities to characterize the dynamic influence potential between parameters of different dimensions. Based on the dynamic coupling strength matrix output from stage S2, a matrix element sieving method (parameter: matrix size = 3×3, including three types of cross-dimensional variable pairs: process-material, material-equipment, and equipment-process) is used to extract the time-varying coupling strength values ​​between each cross-dimensional variable pair from the matrix. The extracted time-varying coupling strength sequence is then processed jointly using a sliding window mutual information calculation algorithm (parameters: window length w, step size s, consistent with stage S2) and a conditional Granger causality test algorithm (parameters: lag order p, significance level α = 0.05) to achieve a dual characterization of nonlinear correlation and predictive causality. The formula used in the mutual information calculation process is:

[0014] in, For time-varying mutual information values, For cross-dimensional variable pairs, Let X be the joint probability distribution of variable X taking the value x and variable Y taking the value y. These are the marginal probability distributions of variable X taking the value x and variable Y taking the value y, respectively, to quantify the degree of information sharing between the two variables within the window.

[0015] In the conditional Granger causality test, a constrained regression model is constructed and the statistic is calculated. To measure the predictive contribution of a variable after controlling for covariates, a directional causal strength index is obtained. Furthermore, the mutual information and causal strength sequence are subjected to directional weighting through a perturbation response consistency metric function (the weight coefficients are generated by normalizing the response consistency values) to form a set of weighted relationships with directional awareness. By encapsulating the weighted set in a structured manner, the initial edge weight of each pair of cross-dimensional nodes is defined as the product of the time-varying coupling strength and the directional weight, thereby realizing a quantitative prior input that reflects the potential dynamic influence between parameters of different dimensions. The above algorithm transforms the coupling strength matrix extracted in the previous step into weighted relational data that can be used for the initialization of cross-dimensional edges in a ternary hypergraph, thereby realizing the dynamic and directional representation of cross-dimensional connections and providing an accurate edge weight benchmark for the construction of the internal sub-parameter network and the overall topology shaping in subsequent sub-steps of S3. For example, in a certain PET preform production batch, the dynamic coupling strength matrix calculation yields a process-material unilateral coupling strength of 0.62, a material-equipment unilateral coupling strength of 0.47, and equipment-process bilateral coupling strengths of 0.55 and 0.51, respectively. The sliding window length is set to 50 seconds, with a step size of 10 seconds. Mutual information calculations determine the process-material coupling strength within a certain window. Materials and Equipment The conditional Granger causality test yields the material → equipment direction when the lag order p=3. Significantly, equipment → process direction Significantly, in the disturbance response consistency metric, the weights for process-materials are 0.87, materials-equipment are 0.74, and equipment-process are 0.79. After normalization, these weights are multiplied by the corresponding coupling strengths to form edge weights: process-materials 0.5394, materials-equipment 0.3478, and equipment-process 0.4345. This weighted set of relationships is used as the initial weights for cross-dimensional edges and input into the ternary hypergraph construction module. This verifies that in subsequent message passing, these weights can significantly improve the effectiveness of cross-dimensional information flow and enhance the accuracy of the final defect attribution. S3.2: Construct internal sub-parameter networks for the three dimensions of process, material, and equipment: construct a process sub-graph based on process parameters such as mold temperature, injection pressure, and cooling time; construct a material sub-graph based on material properties such as raw material moisture content and melt viscosity; and construct an equipment sub-graph based on equipment state parameters such as screw speed and clamping force fluctuation. Within each sub-graph, the static correlation degree between sub-parameters is calculated using Pearson correlation coefficient and time delay alignment strategy to generate the internal connection weights of nodes, forming the local topological basis of the ternary hypergraph. The cross-dimensional coupling strength values ​​extracted in step S3.1 are classified by dimension to determine the set of sub-parameters under the three top-level nodes of process, material, and equipment. The parameter selection rules are adopted (based on the following criteria: the stability coefficient of the parameter in the past N batches is ≥0.85 and the signal sampling integrity is ≥95%). Mold temperature, injection pressure and cooling time are selected as input variables for the process sub-graph, raw material moisture content and melt viscosity are selected as input variables for the material sub-graph, and screw speed and clamping force fluctuation are selected as input variables for the equipment sub-graph. The Pearson correlation coefficient calculation method (parameters: number of data samples m, significance level α=0.05) is used to perform correlation calculation on the time series of any two sub-parameters under the same dimension, and obtain the correlation coefficient matrix reflecting their linear static dependence. The matrix elements take values ​​in the range of [-1,1], with positive values ​​indicating positive correlation and negative values ​​indicating negative correlation. Furthermore, through a time-delay alignment strategy (parameter: maximum delay) =5 sampling period) to perform peak alignment correction on the correlation coefficient sequence, identify the time offset pattern of different sub-parameter responses, and correct the correlation coefficient according to the offset pattern to reflect the true causal lag correlation; Furthermore, a correlation coefficient significance test (method: two-tailed t-test, parameters: degrees of freedom m-2, critical value lookup table method) is used to eliminate insignificant correlated connections at a given confidence level, and the correlation coefficient values ​​of significant connections are mapped to the connection weights within the nodes; Through the above processing, the correlation relationship of each dimension sub-parameter is transformed into the adjacency matrix structure of the corresponding subgraph, which serves as the local topological basis of the ternary hypergraph, realizing the binding of the internal connection weight of the node with the physical correlation of the parameters. For example, in the process dimension of a PET preform production line, the temperature sensor sampling frequency is configured at 10Hz, the pressure sensor sampling frequency at 50Hz, and the cooling time recording accuracy is 0.01s; in the material dimension, the moisture meter measurement accuracy is ±0.001%, and the viscometer output frequency is 1Hz; in the equipment dimension, the screw speed sensor accuracy is ±0.5rpm, and the clamping force fluctuation meter sampling frequency is 10Hz. The Pearson correlation coefficient is calculated for the sequence of mold temperature and injection pressure (sample size m=500) in the process sub-graph, and the correlation coefficient value is obtained. The t-value of the two-tailed t-test is The corresponding p-value is less than 0.001, indicating a significant positive correlation; the correlation coefficient was calculated for the mold temperature and cooling time series. The p-value was 0.004, indicating a significant negative correlation. The correlation category was mapped to a weighted absolute value of 0.65. Time-delay alignment analysis revealed that the response delay of the screw speed to the mold clamping force fluctuation was 2 sampling periods. The original correlation coefficient was then... Revised to The adjacency matrices of the three subgraphs—process, materials, and equipment—are generated in the end, each containing strongly correlated edges with weight values ​​ranging from 0.62 to 0.85. These edges serve as inputs to the internal structure of nodes in the S3.3 hypergraph construction, ensuring that the local topology accurately reflects the static dependency patterns of physical variables in the PET preform production process. S3.3: Integrate the above cross-dimensional coupling strength and subgraph internal association structure to construct a ternary hypergraph topology framework with 'process-material-equipment' as the top-level node; each top-level node encapsulates the sub-parameter set of the corresponding dimension and its internal connection relationship, and the hypergraph edges are determined by the values ​​in the dynamic coupling strength matrix, forming a structured prior knowledge representation that reflects the nonlinear and time-varying interaction characteristics between multi-source variables, which is used for subsequent information propagation modeling under the attention mechanism; Based on the three types of cross-dimensional coupling strength input values ​​(process-materials, materials-equipment, and equipment-process) and the internal sub-parameter network structure of each dimension, a hypergraph topology mapping method (parameters: dynamic coupling strength matrix, sub-parameter node set, internal connection weight) is adopted to realize the structured integration of multi-dimensional coupled data and the encapsulation of top-level nodes. Furthermore, by using a cross-dimensional edge weight injection algorithm (parameters: normalized coupling value matrix, directional response coefficient), the connection edges between process nodes, material nodes, and equipment nodes are assigned corresponding weights, and the sub-parameter network is embedded into the data encapsulation area of ​​each top-level node, realizing the mapping from time-varying nonlinear correlation to multi-dimensional structured node representation; Furthermore, by utilizing the intra-node association aggregation mechanism (parameters: Pearson correlation coefficient matrix, time delay alignment parameter τ), the static association degree of each sub-parameter is weighted and aggregated within the top-level node to generate an internal connection subgraph, while maintaining independence from cross-dimensional edge weights, to ensure that local parameter interactions are independently modeled in the global topology; Furthermore, a hierarchical relationship encoding algorithm (parameters: top-level node identifier, sub-parameter index, edge weight matrix) is applied to encode the hierarchical structure of the three nodes of process-material-equipment and their internal sub-parameter networks, so that the ternary hypergraph can reflect the multi-layer interaction path from top-level coupling to local parameter linkage in the message passing of the graph neural network; Through the above hypergraph topology construction method, the cross-dimensional coupling strength and internal sub-parameter association structure are transformed into composite graph data containing top-level nodes and internal local topology, realizing the output of structured prior knowledge that supports cross-domain information propagation under the attention mechanism. For example, in the PET preform production process, the process dimension includes mold temperature (sampling frequency 1Hz), injection pressure (sampling frequency 2Hz), and cooling time (sampling frequency 0.5Hz); the material dimension includes raw material moisture content (batch sampling) and melt viscosity (sampling once per minute); and the equipment dimension includes screw speed (sampling frequency 5Hz) and clamping force fluctuation (sampling frequency 2Hz). The calculated average coupling strength between process and material is... Materials and equipment are Equipment and process are The directional response coefficient (process-affected material) is Materials affect equipment Using the Pearson correlation coefficient, the static correlation between mold temperature and injection pressure is calculated to be: The static correlation between raw material moisture content and melt viscosity is: These values ​​are used as top-level edge weights and node intra-connection weights, respectively, to form process node encapsulations (including temperature, pressure, and cooling sub-networks), material node encapsulations (including moisture content and viscosity sub-networks), and equipment node encapsulations (including rotational speed and clamping force fluctuation sub-networks) in the hierarchical relationship encoding, generating a ternary hypergraph data structure. This structure significantly improves the ability to mine multi-dimensional implicit associations in the subsequent attention-driven ternary message passing algorithm input, as evidenced by a significant increase in the confidence of the main cause identification on the validation set for the defect attribution model. S3.4: Perform sparsification on the constructed ternary hypergraph and filter weakly coupled connections based on a significance threshold: using the coupling strength distribution obtained statistically under historical steady-state conditions, set a 95% confidence lower bound as the pruning threshold, and remove edge connections below this threshold to reduce noise interference and improve model interpretability; retain strong correlation paths to ensure that the hypergraph structure focuses on high-contribution variable interaction patterns; S3.5: The sparsified ternary hypergraph is serialized and encoded into an adjacency tensor and node feature matrix that can be parsed by a graph neural network. The node features are aggregated from the time series mean and variance of the sub-parameters in the corresponding dimension, and the adjacency tensor is filled with the retained edge weights. The output is standardized graph structure data, which serves as the prior topological input to support the attention-driven ternary message passing algorithm in S4.

[0016] like Figure 3 As shown, step S4 involves: using the defect feature vector extracted from the optical inspection image via a convolutional neural network as the initial message input to the corresponding node in the ternary hypergraph; and, based on an attention-driven ternary message passing algorithm, cyclically propagating and aggregating context-related information flows among the process, material, and equipment nodes, outputting the implicit contribution vector of each dimension to the current defect pattern. Specifically, this includes: S4.1: Based on the pre-trained convolutional neural network model, feedforward computation is performed on the optical inspection images of the bottom of each batch of preforms to extract deep spatial features and pool them into a fixed-dimensional defect feature vector, which serves as the initial semantic representation of the current defect pattern for subsequent cross-modal alignment with the ternary hypergraph structure. S4.2: Perform a linear projection transformation on the defect feature vector, embed it into a latent space dimension compatible with the ternary hypergraph node, and assign the projected features to the most relevant initial node among the three types of nodes 'process', 'material', and 'equipment' based on the prior knowledge of the defect type, forming an initial message injection point to initiate the cross-domain information propagation process; For the defect feature vectors extracted and pooled into fixed-dimensional vectors by a convolutional neural network, a linear projection transformation algorithm is used (parameter: projection matrix size is...). × The weight initialization method is Xavier initialization) to achieve a compatible mapping from the feature space to the hypergraph latent space, transforming the original high-dimensional visual features into low-dimensional vectors that can be directly fused with the feature matrix of the ternary hypergraph nodes; Furthermore, by using a feature normalization method (parameters: mean vector and standard deviation vector are from the training dataset, and Z-score normalization is used), the dimensionality of each feature in the latent space is unified, resulting in a projection feature vector that eliminates the scaling effect. Furthermore, based on prior knowledge labels of defect types (derived from the one-hot encoding of defect types output by the defect detection model), a cosine similarity calculation method is used (parameter: similarity threshold). Establish a correlation assessment between the defect feature vector and the center vector of each node category, and output the index of the most relevant node category; Furthermore, utilizing node category indexing, a multi-category feature allocation algorithm is employed (parameter: the allocation ratio for a single node does not exceed...). The projected feature vectors are distributed into the initial node feature matrices of the three types of nodes, 'process', 'material', and 'equipment', according to the weight of their matching degree with the node type, thereby generating the initial message injection points; By aligning vectorized features with the ternary hypergraph structure, the results of the previous step are transformed into cross-domain compatible initial message input data, thus realizing the initiation condition for the subsequent attention mechanism-driven ternary message passing process. For example, in a PET preform defect attribution task, the defect feature vector output by the convolutional neural network has a dimension of 128, which is then projected by a linear projection matrix (scale). × After transformation (using the Xavier method with weights initialized), a 64-dimensional latent space vector is obtained. This vector is then Z-score normalized, with the mean derived from the 64-dimensional mean vector of the training set and the standard deviation from the corresponding 64-dimensional standard deviation vector. The defect type label is a one-hot encoded yellowing type, and its cosine similarity with the process node center vector is calculated. The similarity value with the material node center vector is The similarity value with the center vector of the device node is Based on similarity threshold The most relevant class is determined to be the process node, and the allocation ratio is determined according to that node. The feature vector elements are directly written into the initial feature matrix of the process node, and the remaining parts are written into the material and equipment nodes respectively according to the matching degree. This process makes the embedding space of the initial message completely consistent with the node space of the ternary hypergraph, and significantly improves the aggregation efficiency of relevant information in the subsequent message passing stage; S4.3: Based on the edge weights constructed by the dynamic coupling strength matrix, perform message passing operations based on the attention mechanism in the ternary hypergraph topology: apply learnable attention coefficients to the message set passed in by adjacent nodes, calculate the importance weights of each connection path, and then aggregate the context information from the other two dimensions in a weighted manner to generate an updated node state vector, reflecting the degree of synergistic influence between multidimensional variables; Under the condition of receiving the projected defect feature vector from the output of S4.2 and mapping it to the initial state of the ternary hypergraph node, an adjacency weighted message passing algorithm based on the attention mechanism (parameters: adjacency tensor A, node feature matrix X, dynamic coupling strength matrix C) is adopted to realize the correlation propagation and importance assessment of cross-dimensional information. Furthermore, the attention coefficient from node j to node i is calculated. Using the scaled dot product attention method (parameter: query vector) Key vector Scaling factor This enables adaptive adjustment of path weights and obtains the degree of attention node i pays to messages from node j.

[0017] in Let i be the query vector for node i. Let be the key vector of node j. The dimension of the key vector; Furthermore, combining the weights in the dynamic coupling strength matrix C The attention coefficients mentioned above are weighted and corrected to inject the influence of cross-dimensional edge strength on message importance assessment, and the corrected attention coefficients are generated. ;

[0018] in, This represents the coupling strength between node i and node j within the current time window. Furthermore, a parallel weighted aggregation of the message sets received by each node is performed through a multi-head attention mechanism (parameters: number of attention heads h, dimension of each attention head). This enables context feature extraction from multiple perspectives and yields the node state update vector after multi-head aggregation. ; Furthermore, a nonlinear activation function (parameter: activation type ReLU or LeakyReLU) is applied to the updated node state vector to realize the nonlinear transformation of node features in high-dimensional space, enhance the feature expression capability, and form the input state for a new round of message passing; By using the above message passing processing method based on attention mechanism weighting, multi-head feature aggregation and nonlinear transformation, the defect features of the previous step are transformed into node state vectors containing cross-dimensional collaborative influence information in the ternary hypergraph, thereby achieving dynamic modeling effect of implicit correlation between process, material and equipment dimensions. For example, in the defect analysis of a batch of PET preforms, the initial input conditions are: process node state vector [0.12, 0.45, 0.33], material node state vector [0.25, 0.38, 0.41], and equipment node state vector [0.31, 0.29, 0.40]. The process-material edge weight in the adjacency tensor is 0.78, the material-equipment edge weight is 0.65, and the equipment-process edge weight is 0.72. A three-head attention mechanism is used, with the query, key, and value vector dimensions set to 64 and a scaling factor √64 = 8. The attention coefficient of the process node on the material node is calculated. for:

[0019] After softmax normalization, the value is approximately 0.21. Combined with the coupling strength weight... =0.78, after correction =0.21×0.78≈0.1638. Multi-head aggregation yields the process node update vector. =[0.28,0.47,0.36]. After the node state vector is updated, it becomes the input for the next round of propagation after ReLU activation. Finally, a stable node state convergence value is obtained in stage S4.4, realizing the information fusion and collaborative relationship characterization of defect features in the ternary domain, which significantly improves the interpretability and accuracy of the subsequent attribution stage; S4.4: Perform a three-stage cyclic propagation process on the updated node state vector, that is, complete a complete information flow in the order of process → material → equipment → process. In each transmission process, continuously optimize the attention weight and accumulate cross-dimensional dependencies until the node state converges or reaches the preset maximum number of iterations, thereby fully exploring potential indirect influence paths. Perform a three-stage cyclic propagation calculation on the state vectors of the three nodes (process, materials, and equipment) updated from S4.3. The input conditions are the weighted updated node state vectors and the sparsed adjacency tensor of the ternary hypergraph and its attention weight coefficient matrix. A sequentially dependent cyclic propagation algorithm is adopted (parameters: propagation sequence is process → material → equipment → process, iteration step size is 1). At the current iteration step, the process node passes its own state vector to the material node after element-wise multiplying the edge weight coefficient and the attention weight coefficient, thereby realizing the cross-dimensional information flow in the first stage. Furthermore, the message vector received by the material node from the process node is fused with its own state vector using a weighted aggregation method (parameter: the aggregation function is a weighted summation, and the weight is the product of the edge weight and the attention coefficient), and then transmitted to the device node in the same way to generate the cross-dimensional information integration result of the second stage. Furthermore, the directional propagation algorithm (parameter: propagation direction is device → process, aggregation rule is the same as before) is used to transmit the fused state vector of the device node back to the process node. During the fusion process, the attention weight coefficient matrix is ​​updated in real time to optimize the contribution ratio of cross-dimensional connection paths and form the closed-loop information propagation result of the third stage. Furthermore, based on the state convergence criterion, the rate of change of the node state vector in two adjacent iterations is calculated. This rate of change is obtained by comparing the node state vector in step t and step t-1. This is used to determine whether the node state has converged. At the same time, it is checked whether the current iteration count has reached the preset maximum iteration count: if the node state meets the convergence condition (the rate of change is lower than the convergence threshold), the loop propagation is terminated; if the convergence condition is not met and the maximum iteration count has not been reached, a new round of process → material → equipment → process propagation iteration is started until the termination condition is met. By using a three-stage loop propagation algorithm and a dynamic attention weight update mechanism, the node state vector of the previous step is transformed into a multi-dimensional state representation containing potential indirect influence paths, thereby fully mining cross-dimensional implicit dependencies. For example, in a single PET preform defect attribution process, the input process node state vector has a dimension of 8, the material node has 5, and the equipment node has 6. The edge weights retained by the sparse adjacency tensor range from 0.65 to 0.92, and the average value of the initial attention weight coefficient matrix is ​​0.48. The loop propagation algorithm is set to an upper limit of 10 iterations, and the convergence threshold is... After the fourth round of propagation, the rate of change of process node state decreased to The termination condition was met. The final output of the process, material, and equipment node status all incorporated contextual information from the other two dimensions. Verification results showed that this status significantly improved the accuracy and stability of indirect path identification in subsequent defect cause contribution calculations. S4.5: The state vectors output by the three nodes of process, materials and equipment that have finally stabilized are normalized by L2 norm and converted into implicit contribution indicators with relative comparability. A three-dimensional contribution vector is generated as the output result to quantify the comprehensive influence of each dimension on the current defect mode and support the subsequent identification of dominant factors and attribution decision.

[0020] Step S5: Based on the implicit contribution vector, the dominant influence dimension is identified, and counterfactual attribution analysis is used to generate a virtual intervention scenario on this dimension. By simulating the change trajectory of defect probability when key parameters deviate from the standard range, an interpretable set of causal inference paths is formed. Specifically, this includes: S5.1: Based on the implicit contribution vectors of the three dimensions of process, materials and equipment output from the previous stage, calculate the normalized weight value of the contribution of each dimension, identify the dominant influence dimension according to the maximum weight criterion, and if the highest weight exceeds the preset threshold (e.g. 0.6) and the difference between it and the second highest weight is greater than 0.2, then the dimension is determined to be a single dominant factor and is used as the object of counterfactual analysis. S5.2: For the identified dominant influencing dimensions, extract the corresponding key parameter variables (such as mold temperature, raw material moisture content or screw speed), and determine the benchmark operating range based on the parameter distribution characteristics under historical normal working conditions. Construct a set of virtual intervention variable sequences that systematically deviate from this range to simulate the scenario of abnormal parameter fluctuations. S5.3: Input the virtual intervention variable sequence into the trained defect probability response model. This model is constructed based on graph neural network and attention mechanism. It can receive the modified process-material-equipment ternary input and output the corresponding defect occurrence probability estimate, thereby obtaining the defect probability response curve that changes with parameter offset. S5.4: Perform differential analysis on the defect probability response curve to identify the critical inflection point where the probability increases significantly and the corresponding parameter offset amplitude. Combine the collaborative response characteristics between variables in the dynamic coupling strength matrix to deduce the trend path of the chain change of other variables caused by the parameter disturbance, and form a preliminary causal inference path candidate set. S5.5: Based on the physical constraint rule base, the feasibility screening of the candidate set of causal inference paths is carried out, paths that violate the thermodynamic or rheological laws of the injection molding process are excluded, and explainable causal chains that conform to industrial mechanisms are retained. The key disturbance nodes and propagation directions of each path are marked to generate the final set of explainable causal inference paths.

[0021] Step S6: Perform Bayesian confidence interval evaluation on the set of causal inference paths, and combine this with attribution conclusions from similar working conditions matched in the historical database of similar defect cases to generate an attribution conclusion report with statistical significance labels and case support, which serves as the final defect cause determination result. Specifically, this includes: S6.1: Based on the set of causal inference paths output by S5, extract the key intervention variables corresponding to each path and their defect probability change sequences under simulated perturbations as observation inputs for Bayesian posterior distribution modeling; use non-parametric Bayesian estimation methods to fit the probability density of the effect strength of each causal path, calculate the confidence interval width and central tendency offset at the 95% confidence level, so as to quantify the statistical stability of the attribution results; Based on the set of interpretable causal inference paths output by S5, the key intervention variables of each path and the corresponding defect probability change sequence under simulated perturbation conditions are used as the original observation input for modeling. The variable parsing module is used to structure and encode key intervention variables, and their identifiers, perturbation magnitudes, perturbation directions and defect probability response values ​​are used to form standardized data records for sample item construction in the Bayesian estimation process; Furthermore, a probability density fit is performed on the effect intensity of each causal path using a non-parametric Bayesian estimation method (the kernel density estimation type is selected as Gaussian kernel, and the bandwidth parameter is set according to the Silverman rule of thumb). The effect intensity is defined as the difference between the defect probability under the intervention and benchmark comparison conditions. Furthermore, confidence intervals are calculated for the effect intensity distribution function of each path, and the interval range at the 95% confidence level is generated using the following formula. :

[0022] in, The sample mean of the effect intensity. The sample standard deviation of the effect strength. This represents the quantile value at the cumulative probability of 0.975 in the standard normal distribution. For sample size; Furthermore, the width of the obtained confidence interval is calculated. offset from the center trend The interval width is calculated using the formula:

[0023] in This is the upper limit of the confidence interval. This represents the lower limit of the confidence interval; The center trend offset is calculated using the formula:

[0024] in The sample mean of the effect intensity. This represents the mean effect intensity under the baseline condition; Through the density fitting and interval calculation described above, the statistical stability of each causal path is quantified into two indicators: interval width and center offset, thereby achieving stability assessment of the attribution results. For example, in the production of a certain batch of PET preforms, the dominant intervention variable included in the counterfactual inference path set is the mold temperature. The decline and the moisture content of raw materials The increase The intervention condition was input into the defect probability response model, yielding a sequence of defect probability sample values ​​for the mold temperature path: {0.41, 0.46, 0.43, 0.45, 0.44}. The sample mean was then calculated. Sample standard deviation Sample size From the table, we can find Substitute into the confidence interval formula to calculate the upper limit. and lower limit , obtain the interval width The central trend offset is 0.412 of the baseline value. The results indicate that this path effect is statistically stable and deviates significantly from the baseline state, and can be identified as a high-confidence causal path in subsequent steps, providing quantitative support for defect attribution determination. S6.2: Perform significance discrimination processing on the Bayesian confidence interval evaluation results, set the confidence interval overlap threshold to 0.15. If the confidence interval of a certain causal path does not significantly overlap with other competing paths and its central effect value is higher than the preset attribution strength threshold, then the path is marked as a 'high confidence causal chain' and assigned a first-level statistical significance label (p<0.05). S6.3: Retrieve the historical similar defect case library from the process knowledge evolution database, and execute the K-nearest neighbor matching algorithm (K=7) based on the current working condition feature vector (including the average mold temperature, raw material moisture content range, and screw speed fluctuation standard deviation) to retrieve the Top-5 historical cases with the smallest Euclidean distance; extract their verified attribution conclusions and corresponding production adjustment feedback results to form a case support evidence set; S6.4: The high-confidence causal chain is fused with the matched historical case supporting evidence set, and a weighted voting mechanism is used to calculate the comprehensive support score: the Bayesian significance contribution weight is 0.6, and the historical case matching consistency contribution weight is 0.4; if the comprehensive score exceeds the decision threshold θ=0.82, the causal path is upgraded to 'strong support attribution conclusion' and the case evidence number is marked. Based on the case support evidence set described in S6.3 and the high-confidence causal chain identified in S6.2, a weighted voting mechanism (parameters: Bayesian significance contribution weight = 0.6, historical case matching consistency weight = 0.4) is used to calculate the support from the two sources and form a comprehensive support score to ensure the comparability of information from different sources under a unified scale. Furthermore, the Bayesian saliency score sequence and the case matching consistency score sequence were scaled to [values] respectively through normalization processing. For each interval, construct a weighted summation formula to calculate the overall support:

[0025] in, Bayesian significance score for high-confidence causal chains. Match consistency scores to the corresponding cases; Furthermore, a floating-point comparison method is used to... A precise comparison is performed with a decision threshold, which is set to 0.82. If the comparison satisfies... If so, the attribution level of the causal path is upgraded from a high-confidence causal chain to a "strong support attribution conclusion" to enhance its reference priority in downstream process optimization decisions; Furthermore, the upgraded causal path is bound to the supporting evidence number in the historical case library. The number comes from the Top-5 historical case set matched by S6.3, and the consistency between the number and the case content is ensured through index mapping. Through the weighted voting fusion and threshold determination process described above, the high-confidence causal chain results of the previous step are transformed into strong supporting attribution conclusions with case support indicators and higher decision reference value, thereby enhancing the multi-source evidence of the defective attribution results. For example, in a batch analysis of PET preform production, the Bayesian significance score for the high-confidence causal chain was 0.87, and the case-match consistency score was 0.78. After normalization, both were within the range of... Interval. Substitute the parameters into the comprehensive support formula: The calculated overall support is .because The system determined that the causal chain was upgraded to a "strongly supported attribution conclusion." It assigned a case verification number "C-20230915-003," and the entry with the same number in the case library contained the same process conditions, material parameters, and equipment status characteristics. This processing result significantly improved the reference value of decision-making in the subsequent process optimization suggestion generation stage, while reducing the possibility of the attribution conclusion being questioned. S6.5: Generate a structured attribution conclusion report, including the dominant influence dimension, key parameter names, defect probability change trajectory diagram, Bayes confidence interval labels, matching historical case numbers, and comprehensive support score; output this report as the final defect cause determination result to the downstream decision module to support the generation of process optimization instructions or trigger the multi-factor collaborative disturbance analysis process.

[0026] Step S7: Determine whether a single main cause with high confidence has been identified in the attribution conclusion report; if so, output the corresponding process optimization suggestion instruction; if not, trigger the multi-factor collaborative perturbation analysis process, and call the incremental data acquisition mechanism to supplement the observation samples under specific operating conditions to enhance the model's discriminative ability. Specifically, this includes: S7.1: Based on the Bayesian confidence intervals and case support labels of each causal dimension in the attribution conclusion report, construct a confidence quantification scoring function, and use a weighted fusion method to calculate the comprehensive confidence score of each potential main cause, which serves as the decision-making basis for determining whether a single main cause with high confidence has been formed. Based on the Bayesian confidence intervals and case support labels for the three causal dimensions (process, materials, and equipment) included in the attribution conclusion report, a weighted fusion scoring method was used (parameter: significance weight). =0.6, Case Support Weight =0.4), enabling multi-source reliable quantification of each potential main cause; Furthermore, using the Bayesian confidence interval significance label transformation formula, the effect values ​​of the "high confidence causal chain" marked in S6.2 are mapped to normalized significance scores. ,in This is the path center effect value. and These are the minimum and maximum effect values ​​for the entire path set, respectively, to achieve normalized quantification of effect strength; Furthermore, the overall support score output by S6.4 is converted into a case consistency score using the case support label mapping function. ,in The overall support for the current path, and These represent the minimum and maximum support for the entire set, enabling standardized processing of case information. Furthermore, the aforementioned significance scores and case consistency scores were weighted accordingly. and Perform linear weighted calculation ,in To calculate the overall confidence score, The significance score is... To assign a consistency score to cases, and to achieve quantitative fusion of information from multiple sources; By using the weighted fusion method described above, the statistical inference results and empirical case verification results of the previous step attribution conclusion report are transformed into a comprehensive confidence score that can be used for threshold determination, thereby improving the accuracy and stability of S7.2 in single cause identification. For example, under certain production conditions, the Bayesian central effect value is E=0.74, and the total set... =0.52, =0.85, obtained using the significance score calculation formula. =0.667. The overall support of cases for this defect path is... =0.88, complete series =0.75, =0.93, obtained through the case consistency score formula. =0.722. By weight =0.6、 =0.4 Perform weighted fusion calculation =0.692, generating a comprehensive confidence score for this causal dimension. In the attribution determination stage, this score is used to compare with a threshold to confirm whether it constitutes a single principal cause with high confidence. When the threshold is set to 0.65, the score of this case meets the principal cause identification condition, providing a quantitative decision-making basis for subsequent generation of process optimization suggestions; S7.2: Perform a threshold comparison operation on the comprehensive confidence score, and perform binary classification judgment based on the preset high confidence judgment threshold: if there is a single cause dimension score that exceeds the threshold and is significantly higher than the other dimensions, then a single main cause with high confidence is identified; otherwise, it is judged as a multi-factor coupling interference state. S7.3: Under the premise of confirming that a single main cause with high confidence is established, based on the deviation direction and magnitude of the key parameters corresponding to the main cause, combined with the historical process optimization strategy knowledge base, process optimization suggestion instructions with operational semantics are generated and encapsulated into an executable control command format and output to the production management system. S7.4: In the absence of a single main cause with high confidence, activate the multi-factor collaborative perturbation analysis process. Based on the process-material-equipment combination with multiple strongly coupled edges in the current ternary hypergraph topology, generate a targeted multivariate perturbation experiment design matrix to guide the direction and scope of incremental data acquisition. S7.5: Invoke the incremental data acquisition mechanism to implement the working condition adjustment specified by the multivariate perturbation experimental design matrix in the controlled production environment, and simultaneously acquire the newly added multi-source time series data and defect detection results as the input sample set for subsequent dynamic coupling strength matrix reconstruction and attribution model enhancement training.

[0027] Step S8: The dynamic coupling strength matrix update record, attribution conclusion report, and feedback execution result of this inference process are stored in the process knowledge evolution database. Based on the newly accumulated data, the ternary hypergraph structure and message passing weights are periodically calibrated to achieve closed-loop iterative optimization of the defect attribution model. Specifically, this includes: S8.1: The dynamic coupling strength matrix update record, attribution conclusion report and corresponding process optimization suggestion execution feedback results generated during this inference process are structured and encapsulated. Key fields including coupling strength change trajectory, dominant influence dimension identification results, counterfactual inference path set and its Bayes confidence interval are extracted to form a standardized knowledge tuple data package to support the incremental storage and index retrieval of the knowledge base in the future. S8.2: Based on the knowledge tuple data packet, write it into the corresponding time batch and production condition classification directory in the process knowledge evolution database, use the database transaction mechanism to ensure the consistency and integrity of data writing, and use timestamp markers and batch IDs to achieve traceable association of multi-source heterogeneous knowledge items, and generate historical knowledge snapshots with spatiotemporal context attributes; S8.3: Based on historical knowledge snapshots accumulated in the process knowledge evolution database, the dynamic calibration process of the ternary hypergraph topology is triggered periodically. First, the statistical distribution trend of the edge weights between each 'process-material-equipment' node is calculated to identify the standard coupling patterns with significant deviations. Then, the sliding window weighted average algorithm is used to fuse the dynamic coupling strength matrices of the latest N batches to generate an updated prior coupling strength benchmark matrix, which serves as the input basis for adjusting the edge weights of the ternary hypergraph. S8.4: Based on the updated prior coupling strength benchmark matrix, the weight parameters of the three types of connection edges in the current ternary hypergraph, namely process-material, material-equipment, and equipment-process, are adaptively corrected. Combined with the historical attribution accuracy evaluation results, the gradient descent strategy is used to optimize the message passing weight coefficient in the attention mechanism, so that the ternary message passing algorithm is more inclined to activate high-frequency verified effective causal paths in subsequent inference, and outputs the calibrated ternary hypergraph structure and inference parameter set. S8.5: Deploy the calibrated ternary hypergraph structure and inference parameter set to the online defect attribution system, replace the original model configuration, and set a version control identifier and effective time window; at the same time, start the backtracking verification task, call the historical defect samples in the recent period for re-inference test, compare the performance differences between the new and old models in terms of attribution consistency, confidence stability and case matching, generate a model iteration evaluation report, and complete the closed-loop optimization verification process of the defect attribution system.

[0028] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0029] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for tracing and analyzing the source of defects such as whitening and yellowing at the bottom of PET preforms during production, characterized in that, Includes the following steps: S1: Obtain multi-source time-series data of PET injection molding process parameters related to the whitening and yellowing defects at the bottom of the preform, as well as optical inspection images of the bottom of each batch of preforms and their defect scoring results, as the original input dataset; S2: Perform sliding window segmentation on the original input dataset, and calculate the time-varying mutual information values ​​between three types of variable pairs: process parameters and materials, material properties and equipment status, and equipment and process based on the segmented time series segments. Extract the dynamic response consistency features between each variable before and after the disturbance, and generate a dynamic coupling strength matrix. S3: Based on the dynamic coupling strength matrix, construct a ternary hypergraph with process-material-equipment as nodes; S4: Input the defect feature vector extracted from the optical detection image by the convolutional neural network into the corresponding node in the ternary hypergraph, propagate it in a loop between nodes and aggregate the context-related information flow, and output the implicit contribution vector of each dimension to the current defect mode. S5: Based on the implicit contribution vector, identify the dominant influence dimension, generate a virtual intervention scenario on the dominant influence dimension, and form a set of causal inference paths by simulating the change trajectory of defect probability when key parameters deviate from the standard range; S6: Perform Bayesian confidence interval assessment on the set of causal inference paths, and generate an attribution conclusion report by combining the attribution conclusions of similar working conditions matched in the historical similar defect case library.

2. The method for tracing and analyzing the source of defects such as whitening and yellowing at the bottom of PET preforms according to claim 1, characterized in that, Following step S6, the following is also included: S7: Determine whether a single main cause with high confidence has been identified in the attribution conclusion report. If so, output the corresponding process optimization suggestion instruction; if not, trigger the multi-factor collaborative disturbance analysis process and call the incremental data acquisition mechanism to supplement the observation samples under specific working conditions. S8: Store the dynamic coupling strength matrix update record, attribution conclusion report and feedback execution result of this reasoning process into the process knowledge evolution database, and periodically calibrate the ternary hypergraph structure and message passing weights based on the newly accumulated data.

3. The method for tracing and analyzing the source of defects such as whitening and yellowing at the bottom of PET preforms according to claim 1, characterized in that, Step S1 specifically includes: Acquire time-series data of process parameters related to preform quality in PET injection molding production line. Based on the process variables within the execution cycle of the industrial sensor real-time acquisition system, synchronously record them at a standard sampling frequency to obtain the raw process data stream. Obtain time series data of material parameters related to raw material properties. Based on the output signals of near-infrared moisture meter and online viscometer, measure the material state before each batch of material is fed. Use the data alignment mechanism to match the timestamps with the process data of the corresponding production cycle to generate a structured material property sequence. Acquire time series data of equipment parameters related to equipment operating status, collect dynamic response signals of key components of injection molding machine based on PLC controller and vibration sensor, use sliding mean filtering to remove high frequency noise interference, extract the effective amplitude sequence of steady state segment, and form equipment status characterization data; High-resolution optical inspection images of the bottom area of ​​each batch of preforms are acquired synchronously. Based on the machine vision system, a line scan camera is triggered at a fixed position on the conveyor belt to capture images with uniform grayscale distribution. The images are then combined with an image quality discrimination algorithm to exclude blurry or occluded samples and generate a valid image set. The defect score results for each batch of preforms are generated by combining expert annotation and automated scoring model. A pre-trained convolutional neural network is applied to the optical inspection image to extract local texture anomaly features and output the severity score of whitening and yellowing defects. The severity score of whitening and yellowing defects is then associated and integrated with the original process data stream, the structured material attribute sequence, and the equipment status characterization data by batch ID to form a unified indexed original input dataset.

4. The method for tracing and analyzing the source of defects such as whitening and yellowing at the bottom of PET preforms according to claim 3, characterized in that, The process parameter time series data includes mold temperature, injection pressure and cooling time; the material parameter time series data includes raw material moisture content and melt viscosity; and the equipment parameter time series data includes screw speed and clamping force fluctuation value.

5. The method for tracing and analyzing the source of defects such as whitening and yellowing at the bottom of PET preforms according to claim 1, characterized in that, Step S2 specifically includes: Sliding window segmentation is performed on multi-source time series data. Based on the preset time window length and step size, key variables of process parameters, material properties and equipment status in PET injection molding are divided into overlapping time segments to generate a time series segment set. Based on the set of time segments, the sliding window mutual information algorithm is used to calculate the nonlinear correlation between the time series data of process parameters and material properties, to obtain the time-varying mutual information values ​​of process-material variable pairs in different time periods, and output the process-material time-varying mutual information sequence. Based on the aforementioned time series set, the conditional Granger causality test method is used to analyze the causal strength of the material property and equipment state variable pairs. After controlling for the influence of other covariates, it is determined whether there is a statistically significant predictive causal relationship, and its F-statistic value is calculated as the conditional causal strength index of the material-equipment variable pair, generating a material-equipment time-varying causal strength sequence. Based on the set of time segments, conditional Granger causality modeling is performed on the cross-direction of equipment state and process parameter variables to identify whether equipment fluctuations before and after disturbance events cause process deviations or vice versa. Combining the directional causal strength and the magnitude of mutual information change, a two-way time-varying coupling index sequence of equipment-process is constructed. By integrating the time-varying mutual information sequence of the process-material, the time-varying causal strength sequence of the material-equipment, and the bidirectional time-varying coupling index sequence of the equipment-process, a three-dimensional coupling strength vector is generated and arranged into a symmetric matrix form to form a dynamic coupling strength matrix.

6. The method for tracing and analyzing the source of defects such as whitening and yellowing at the bottom of PET preforms according to claim 1, characterized in that, Each node in the ternary hypergraph contains a set of sub-parameters in the corresponding dimension, and the edge weights are determined by the corresponding coupling strength values ​​in the dynamic coupling strength matrix.

7. The method for tracing and analyzing the source of defects such as whitening and yellowing at the bottom of PET preforms according to claim 1, characterized in that, Step S4 specifically includes: The pre-trained convolutional neural network model performs feedforward computation on the optical inspection images of the bottom of each batch of preforms, extracts deep spatial features and pools them into a fixed-dimensional defect feature vector. The defect feature vector is subjected to linear projection transformation and embedded into a latent space dimension compatible with the ternary hypergraph node. Based on the prior knowledge of the defect type, the projected features are assigned to the most relevant initial node among the three types of nodes: process, material, and equipment, forming the initial message injection point. Based on the edge weights constructed by the dynamic coupling strength matrix, a message passing operation based on the attention mechanism is performed in the ternary hypergraph topology. Learnable attention coefficients are applied to the message set passed in by neighboring nodes, the importance weights of each connection path are calculated, and then the context information from the other two dimensions is aggregated in a weighted manner to generate an updated node state vector. The updated node state vector is subjected to a three-stage cyclic propagation process. In each propagation process, attention weights are continuously optimized and cross-dimensional dependencies are accumulated until the node state converges or the preset maximum number of iterations is reached. The state vectors output by the three nodes of process, materials and equipment that are finally stabilized are normalized by L2 norm and converted into implicit contribution indicators to generate a three-dimensional implicit contribution vector.

8. The method for tracing and analyzing the source of defects such as whitening and yellowing at the bottom of PET preforms according to claim 7, characterized in that, The three-stage cyclic propagation process uses a sequentially dependent cyclic propagation algorithm. The propagation sequence is process → material → equipment → process, with an iteration step size of 1. In the current iteration step, the process node passes its own state vector to the material node after element-wise multiplying the edge weight coefficient and the attention weight coefficient.

9. The method for tracing and analyzing the source of defects such as whitening and yellowing at the bottom of PET preforms according to claim 1, characterized in that, Step S5 specifically includes: Based on the output three-dimensional implicit contribution vector, the normalized weight value of the contribution of each dimension is calculated. The dominant influence dimension is identified according to the maximum weight criterion. If the highest weight exceeds the preset threshold and the difference between the highest weight and the second highest weight is greater than 0.2, then the dimension is determined to be a single dominant factor. For the identified dominant influencing dimensions, the corresponding key parameter variables are extracted, and the baseline operating range is determined based on the parameter distribution characteristics under historical normal operating conditions, and a virtual intervention variable sequence is constructed. The virtual intervention variable sequence is input into the trained defect probability response model, which receives the modified process-material-equipment ternary input and outputs the corresponding defect occurrence probability estimate to obtain the defect probability response curve. Differential analysis is performed on the defect probability response curve to identify the critical inflection point where the probability increases significantly and the corresponding parameter offset. Combined with the collaborative response characteristics between variables in the dynamic coupling strength matrix, the trend path of the parameter disturbance causing chain changes in other variables is deduced, forming a candidate set of causal inference paths. The feasibility of the candidate causal inference path is screened based on the physical constraint rule base. Paths that violate the thermodynamic or rheological laws of the injection molding process are excluded, while explainable causal chains that conform to industrial mechanisms are retained. The key perturbation nodes and propagation directions of each path are marked to generate the final set of explainable causal inference paths.

10. The method for tracing and analyzing the source of defects in the bottom of PET preforms as described in claim 9, characterized in that, The defect probability response model is constructed based on a graph neural network and an attention mechanism.

Images