Supercritical foaming material production quality management method based on big data

Abstract

This invention discloses a big data-based method for quality control in the production of supercritical foamed materials, belonging to the field of manufacturing quality control technology. This big data-based method acquires time-series data and secondary mapping data on barrel temperature deviation, melt pressure fluctuation, and screw speed fluctuation during continuous extrusion foaming. Based on these two types of data, it determines the correspondence between the carbon dioxide dissolution rate distribution and the residence time distribution, and generates concentration gradient field characterization data. After fusion, it obtains fused feature data, inputs it into a degradation trend identification model, outputs the degradation trend of dissolution uniformity inside the barrel, and determines the risk of finished product quality defects. This invention simultaneously collects process time-series data and secondary mapping data reflecting the internal state of the melt, generates concentration gradient field characterization data, and fuses and models it to identify the degradation trend and risk status of dissolution uniformity. It locates the data source deviation before defects form, providing a traceable decision-making basis for production quality control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of manufacturing quality control, in particular to a supercritical foaming material production quality control method based on big data. BACKGROUND

[0002] Supercritical foaming material is a kind of lightweight polymer material with micro-porous structure formed in polymer melt by using supercritical carbon dioxide or supercritical nitrogen as a physical foaming agent, which is widely used in sports shoe midsole, automotive interior, cushion packaging and aviation thermal insulation fields. Continuous extrusion foaming is one of the main process routes for industrial production of supercritical foaming material. The basic process is as follows: the polymer raw material is melted and plasticized by the extruder, then the supercritical fluid is injected into the cylinder and mixed to form a homogeneous system, and then it is extruded through the die of the machine head and rapidly depressurized to induce bubble nucleation and growth in the melt, and finally solidified and shaped into a foamed product with micro-porous structure.

[0003] In the continuous extrusion foaming production process, the main process parameters affecting product quality include barrel temperature, die melt pressure, screw speed and supercritical fluid injection flow rate. In order to maintain the stability of the production process, the existing production equipment is usually equipped with online monitoring devices such as temperature sensors, pressure sensors and speed encoders, which can collect time series data of the above parameters in real time, and adjust the process set value through programmable logic controller or distributed control system. Some production enterprises introduce statistical process control method to monitor the control chart of key parameters such as temperature deviation and pressure fluctuation, and issue an alarm prompt when the parameter exceeds the preset control limit.

[0004] There are related researches and practices in the industry around the data analysis method of continuous extrusion foaming quality control. The common practice is to collect temperature, pressure, speed and other process time series data during extrusion, and use multivariate linear regression, partial least squares regression or artificial neural network modeling methods to establish a statistical mapping relationship between process parameters and macro performance indicators such as foamed product density, resilience and compression permanent deformation. This method can provide a reference for process parameter setting and monitor the long-term drift trend of product quality in production scenarios where process parameters are relatively stable and raw material batch quality fluctuates little.

[0005] The limitations of existing technologies include at least the following problems: Existing technologies mainly collect time-series data related to main process parameters, such as barrel temperature deviation, melt pressure fluctuation, and screw speed fluctuation. This collection framework assumes that the material is always in thermodynamic equilibrium. In actual production, slight changes in the rheological properties of raw material batches can cause deviations in the distribution of carbon dioxide dissolution rate and residence time in the melt, forming a concentration gradient field that evolves along the extrusion direction. Temperature and pressure data are difficult to reflect the distribution of this physical field. The lack of synchronous collection of secondary mapping data of the invisible state inside the melt makes it difficult for the control system to identify the deterioration trend of dissolution uniformity inside the barrel when the process parameters show a steady state. There is a dimensional mismatch between the collected information and the evolution process of the product's microstructure. Big data analysis based on such surface data can only establish a broad statistical correlation between process parameters and finished product quality. When faced with uneven cell size and excessive density fluctuations caused by abnormal internal physical fields, it is difficult to trace the cause of defects from the source of data and to make accurate intervention decisions before defects form. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a big data-based method for quality control in the production of supercritical foamed materials. This method solves the problem that existing technologies, due to insufficient data collection dimensions, make it difficult to perceive the deterioration trend of melting uniformity within the melt, thus making it difficult to predict and trace quality defects.

[0007] To achieve the above objectives, this invention provides the following technical solution: a big data-based method for quality control in the production of supercritical foamed materials, comprising the following steps: acquiring time-series data and secondary mapping data related to the main process parameters during the continuous extrusion production of supercritical foamed materials, including barrel temperature deviation, melt pressure fluctuation, and screw speed fluctuation; determining the correspondence between the carbon dioxide dissolution rate distribution and residence time distribution within the melt based on the time-series data and secondary mapping data related to the main process parameters, and generating concentration gradient field characterization data evolving along the extrusion direction; fusing the concentration gradient field characterization data and the time-series data related to the main process parameters to obtain fused feature data; inputting the fused feature data into a dissolution uniformity degradation trend identification model, and outputting the dissolution uniformity degradation trend identification result within the barrel; determining the finished product quality defect risk monitoring result corresponding to the internal physical field anomaly based on the dissolution uniformity degradation trend identification result; when the quality defect risk monitoring result indicates a risk, performing causal tracing analysis based on the fused feature data to locate the data source deviation that causes the dissolution uniformity degradation trend, and generating production quality control tracing and early warning decision information based on the located data source deviation.

[0008] Further, the specific steps for generating concentration gradient field characterization data evolving along the extrusion direction are as follows: Feature extraction is performed on the secondary mapping data to obtain the feature sequence of carbon dioxide dissolution state inside the melt, and the time-series data related to the main process parameters are aligned with the time axis to construct a multi-channel input matrix; the multi-channel input matrix is ​​input into a physical information neural network, which, constrained by the melt rheological constitutive equation and the carbon dioxide dissolution and diffusion equation, outputs the correspondence between the carbon dioxide dissolution rate distribution and the residence time distribution at each discrete cross-sectional position along the extrusion direction inside the melt; spatial interpolation and gradient calculation are performed on the correspondence between the carbon dioxide dissolution rate distribution and the residence time distribution at each discrete cross-sectional position to generate concentration gradient field characterization data evolving along the extrusion direction.

[0009] Further, the specific steps to obtain the characteristic sequence of carbon dioxide dissolution state inside the melt are as follows: perform wavelet packet decomposition on the power compensation micro-variable time series in the secondary mapping data to extract the energy entropy characteristic sequence; perform frequency band energy integration on the acoustic emission energy attenuation spectrum time series in the secondary mapping data to extract the acoustic signature characteristic sequence; and concatenate the energy entropy characteristic sequence and the acoustic signature characteristic sequence to obtain the characteristic sequence of carbon dioxide dissolution state inside the melt.

[0010] Furthermore, the specific steps of the physical information neural network with the melt rheological constitutive equation and the carbon dioxide dissolution and diffusion equation as constraints are as follows: construct the prediction error term and the physical constraint penalty term in the loss function of the physical information neural network; during the forward propagation of the physical information neural network, calculate the prediction error term and the physical constraint penalty term simultaneously, and minimize the degree of violation of the prediction error and physical constraint through back propagation.

[0011] Further, the specific steps to obtain the fused feature data are as follows: the spatial distribution sequence of the concentration gradient field characterization data along the extrusion direction is dimensionally spliced ​​with the time-aligned barrel temperature deviation time series, melt pressure fluctuation time series, and screw speed fluctuation time series to obtain the fused feature matrix; principal component analysis is performed on the fused feature matrix to reduce its dimensionality, and the principal component components with a cumulative contribution rate exceeding a preset threshold are retained as the fused feature data.

[0012] Furthermore, the specific steps for identifying the trend of dissolution uniformity degradation inside the barrel are as follows: the fused feature data is divided into multiple feature subsequences according to a time sliding window, and each feature subsequence corresponds to a production state at a time step; the multiple feature subsequences are sequentially input into the dissolution uniformity degradation trend identification model, and the dissolution uniformity state category corresponding to each time step is output; the dissolution uniformity state categories of multiple consecutive time steps are combined to form a state transition sequence, which is used as the identification result of the trend of dissolution uniformity degradation inside the barrel.

[0013] Further, the specific steps for outputting the dissolution uniformity state category corresponding to each time step are as follows: The dissolution uniformity degradation trend identification model is a temporal classification model based on a long short-term memory network; this temporal classification model controls the degree of information discarded in the hidden state of the previous time step through forget gate units; this temporal classification model controls the degree of information written in the input features of the current time step through input gate units; this temporal classification model controls the amount of information passed from the hidden state of the current time step to the next time step through output gate units; the hidden state of the last time step of this temporal classification model is extracted, mapped through a fully connected layer, and the probability value of each dissolution uniformity state category is calculated through a normalized exponential function; the dissolution uniformity state category corresponding to the maximum probability value is taken as the output state category of that time step.

[0014] Furthermore, the specific steps for determining the finished product quality defect risk monitoring results are as follows: Analyze the state transition sequence in the identification result of the deterioration trend of the dissolution uniformity inside the barrel; when a unidirectional transition from a slightly deteriorated state to a moderately deteriorated state occurs in the state transition sequence, or when a state category of moderate deterioration or above occurs, the finished product quality defect risk monitoring result is determined to be risky; when all state categories in the state transition sequence remain normal, or when only an isolated slightly deteriorated state exists and subsequently recovers to a normal state, the finished product quality defect risk monitoring result is determined to be risk-free.

[0015] Furthermore, the specific steps for locating the data source deviation that causes the deterioration trend of dissolution uniformity are as follows: Extract the fusion feature subsequence corresponding to the time step identified as having risk from the fusion feature data; calculate the contribution score of each dimension feature in the fusion feature subsequence to the output of the dissolution uniformity deterioration trend identification model using SHAP values; filter feature dimensions whose contribution scores exceed a preset significant threshold, and map the filtered feature dimensions back to the corresponding original data source type. The original data source types include barrel temperature deviation, melt pressure fluctuation, screw speed fluctuation, power compensation micro-variation features, and acoustic signature features; when the contribution score ratio of power compensation micro-variation features or acoustic signature features in the mapped original data source type exceeds a preset proportion, the data source deviation is located as an anomaly in the internal physical field of the melt corresponding to the secondary mapped data; when the contribution score ratio of barrel temperature deviation, melt pressure fluctuation, or screw speed fluctuation in the mapped original data source type exceeds a preset proportion, the data source deviation is located as an abnormal disturbance of the corresponding process parameters.

[0016] Furthermore, the specific steps for generating production quality control traceability and early warning decision information based on the data source deviation are as follows: When the data source deviation is an anomaly in the physical field inside the melt corresponding to the secondary mapping data, quality control early warning information containing the anomaly identifier of the physical field inside the melt and the risk level indication is generated; when the data source deviation is an abnormal disturbance of the corresponding process parameter, quality control analysis information containing the abnormality type of the corresponding process parameter and the abnormality traceability conclusion is generated.

[0017] The present invention has the following beneficial effects:

[0018] (1) This big data-based method for quality control of supercritical foamed material production collects time-series data of main process parameters such as barrel temperature deviation, melt pressure fluctuation, and screw speed fluctuation, and simultaneously collects power compensation micro-variation time-series data and acoustic emission energy attenuation spectrum time-series data as secondary mapping data to achieve quantitative characterization of carbon dioxide dissolution state inside the melt. Existing technologies mostly rely on surface process parameters for quality control, making it difficult to detect the deterioration of dissolution uniformity caused by small changes in the rheological properties of raw material batches. This invention extracts energy entropy features and acoustic features through wavelet packet decomposition and frequency band energy integration to construct concentration gradient field characterization data evolving along the extrusion direction. When process parameters remain stable, it can identify changes in dissolution gradient inside the barrel, expand the data perception dimension of the production process, and make up for the shortcomings of existing quality control in perceiving the physical state inside the melt.

[0019] (2) The big data-based supercritical foaming material production quality control method integrates concentration gradient field characterization data with main process parameter time series data, and inputs it into a dissolution uniformity deterioration trend identification model based on long short-term memory network to realize continuous monitoring and risk identification of the evolution trend of dissolution state inside the barrel. The integrated feature data simultaneously carries the equipment operating status and the physical field information inside the melt. The long short-term memory network can mine time series features, capture the deterioration trajectory of dissolution uniformity over time, form a state transition sequence, and identify quality defect risks accordingly. Compared with the traditional single threshold alarm method, it can capture the critical change of dissolution uniformity deterioration earlier and output risk supervision results before the formation of macroscopic quality defects, providing advance risk warning for production management.

[0020] (3) The big data-based supercritical foaming material production quality control method, when determining the existence of quality defect risk, locates the source deviation of the data source causing the deterioration of the dissolution uniformity through SHAP value contribution analysis, and generates corresponding production quality control traceability and early warning decision information. By quantifying the influence of each feature dimension on the deterioration trend, high contribution features can be mapped to the original data source to distinguish whether the deterioration cause is the physical field change inside the melt or the abnormal fluctuation of the main process parameters. Based on the location results, the abnormality identifier, risk level, abnormality type and traceability conclusion generated can quickly locate the root cause of quality fluctuation, reduce the misjudgment caused by relying on experience to investigate, and shorten the defect cause analysis cycle.

[0021] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method for quality control of supercritical foamed material production based on big data, as described in this invention.

[0023] Figure 2 This is a flowchart illustrating the specific steps involved in obtaining fused feature data in the big data-based method for quality control of supercritical foamed material production.

[0024] Figure 3 This is a flowchart illustrating the specific steps involved in determining the risk monitoring results of finished product quality defects in the big data-based supercritical foaming material production quality control method of this invention. Detailed Implementation

[0025] Please see Figure 1 This invention provides a technical solution: a big data-based method for quality control in the production of supercritical foamed materials, comprising the following steps: acquiring time-series data and secondary mapping data related to the main process parameters during the continuous extrusion production of supercritical foamed materials, including barrel temperature deviation, melt pressure fluctuation, and screw speed fluctuation; determining the correspondence between the carbon dioxide dissolution rate distribution and residence time distribution inside the melt based on the time-series data and secondary mapping data related to the main process parameters, and generating concentration gradient field characterization data evolving along the extrusion direction; fusing the concentration gradient field characterization data and the time-series data related to the main process parameters to obtain fused feature data; inputting the fused feature data into a dissolution uniformity degradation trend identification model, and outputting the dissolution uniformity degradation trend identification result inside the barrel; determining the finished product quality defect risk monitoring result corresponding to the internal physical field anomaly based on the dissolution uniformity degradation trend identification result; when the quality defect risk monitoring result indicates a risk, performing a cause tracing analysis based on the fused feature data to locate the data source deviation that causes the dissolution uniformity degradation trend, and generating production quality control traceability and early warning decision information based on the located data source deviation.

[0026] Specifically, the secondary mapping data consists of barrel heating power feedback timing data (i.e., power compensation micro-variation timing) and head acoustic emission signal timing data (i.e., acoustic emission energy attenuation spectrum timing). The acquisition frequency of the timing data and secondary mapping data related to the main process parameters is uniformly 100Hz, and the acquisition duration is consistent with the continuous extrusion production duration. During the acquisition process, the sliding mean filtering method is used to remove accidental abnormal data, and the filtering window is set to 5 acquisition points. The input dimension of the dissolution uniformity degradation trend identification model is consistent with the dimension of the fused feature data, and the output dimension is consistent with the number of dissolution uniformity state categories, which are divided into 4 categories: normal state, slightly deteriorated state, moderately deteriorated state, and severely deteriorated state. Production quality control traceability and early warning decision-making information need to be linked to production data at the corresponding time step.

[0027] Specifically, the steps for generating characterization data of the concentration gradient field evolving along the extrusion direction are as follows: Feature extraction is performed on the secondary mapping data to obtain the feature sequence of carbon dioxide dissolution state inside the melt. This sequence is then aligned with the time-series data related to the main process parameters to construct a multi-channel input matrix, specifically as follows: The secondary mapping data includes power compensation micro-variation time series and acoustic emission energy attenuation spectrum time series. After feature extraction, energy entropy feature sequence and acoustic signature feature sequence are obtained respectively. The length of the feature sequence is one-tenth of the total number of collection points, that is, one feature value is extracted for every 10 collection points. The timing data related to the main process parameters, including barrel temperature deviation timing, melt pressure fluctuation timing, and screw speed fluctuation timing, all retain the original acquisition length. The time axis alignment uses linear interpolation to stretch the two feature sequences to the same length as the timing data of the main process parameters, ensuring that the time axis is completely consistent. The multi-channel input matrix is ​​a matrix with 5 columns and 6 rows for the total number of acquisition points. The columns correspond to the barrel temperature deviation, melt pressure fluctuation, screw speed fluctuation, and the interpolated energy entropy feature value and acoustic feature value, respectively. Each row corresponds to all feature data at one acquisition time. The multi-channel input matrix is ​​fed into the physical information neural network. The physical information neural network, constrained by the melt rheological constitutive equation and the carbon dioxide dissolution and diffusion equation, outputs the correspondence between the carbon dioxide dissolution rate distribution and the residence time distribution at each discrete cross-sectional position along the extrusion direction within the melt. Specifically: The physical information neural network consists of an input layer, three hidden layers, and an output layer. The number of neurons in the three hidden layers are 64, 32, and 16, respectively. The input layer receives a multi-channel input matrix, and the output layer outputs the carbon dioxide dissolution rate and residence time corresponding to 20 discrete cross-sectional positions uniformly distributed along the extrusion direction. The discrete cross-sections are uniformly distributed from the extrusion start end (0 mm) to the end end (2000 mm). The melt rheological constitutive equation adopts a power-law equation, the core of which is the relationship between melt viscosity and shear rate. The carbon dioxide dissolution and diffusion equation is the core of which is the relationship between the change of carbon dioxide concentration over time and the diffusion coefficient. The two equations are used as physical constraints and embedded in the loss function of the neural network. Finally, the output is the set of correspondences between carbon dioxide dissolution rate and residence time at each discrete cross-sectional position. Spatial interpolation and gradient calculation are performed on the correspondence between the carbon dioxide dissolution rate distribution and the residence time distribution at each discrete cross-section location to generate concentration gradient field characterization data evolving along the extrusion direction, specifically as follows: Spatial interpolation uses cubic spline interpolation, with 20 discrete cross-sectional locations as interpolation nodes, to interpolate the carbon dioxide dissolution rate and obtain a continuous dissolution rate distribution at any location in the extrusion direction; The concentration gradient was calculated using the first-order central difference method, combined with the relationship between dissolution rate and concentration (concentration is the integral of the ratio of dissolution rate to residence time along the extrusion direction), to obtain the concentration gradient at each position along the extrusion direction. The concentration gradient field characterization data is a 20-row, 2-column matrix, with each row corresponding to the location of a discrete cross section and the concentration gradient value at that location, forming the concentration gradient field characterization data that evolves along the extrusion direction.

[0028] The specific steps to obtain the characteristic sequence of the carbon dioxide dissolution state inside the melt are as follows: Wavelet packet decomposition is performed on the power compensation micro-variable time series in the secondary mapping data to extract the energy entropy feature sequence, specifically as follows: The power compensation micro-variable timing sequence is decomposed into 3-level wavelet packet decomposition using the db4 wavelet basis, and the wavelet packet coefficients of 8 frequency bands are obtained after decomposition. Calculate the energy and total energy of each frequency band, and then calculate the energy entropy feature of each feature window based on the energy ratio. Each feature window contains 10 sampling points, and the sliding step size is 10 sampling points. The length of the obtained energy entropy feature sequence is one-tenth of the total number of sampling points; For example, when the total number of collection points is 10,000, the feature sequence length is 1,000, which means it consists of 1,000 energy entropy feature values. Frequency band energy integration is performed on the acoustic emission energy attenuation spectrum time series in the secondary mapping data to extract the acoustic signature feature sequence, specifically as follows: The frequency range of the acoustic emission energy attenuation spectrum timing is 10kHz-100kHz, which is divided into 5 characteristic frequency bands: 10-28kHz, 28-46kHz, 46-64kHz, 64-82kHz, and 82-100kHz. Energy integration is performed on each feature frequency band, and the energy integration values ​​of the five frequency bands are normalized to obtain the voiceprint features of each feature window. The feature window settings are consistent with the energy entropy feature extraction. The final voiceprint feature sequence length is consistent with the energy entropy feature sequence. Each voiceprint feature is a 5-dimensional vector, reflecting the activity of carbon dioxide bubble prenucleation. The energy entropy feature sequence and the acoustic signature feature sequence are concatenated to form a feature sequence of the carbon dioxide dissolution state inside the melt, specifically as follows: By using a vertical splicing method, the energy entropy feature sequence in row 1 (feature sequence length column) is spliced ​​with the voiceprint feature sequence in row 5 (feature sequence length column), resulting in a feature sequence of carbon dioxide dissolution state inside the melt in row 6 (feature sequence length column). For example, when the feature sequence length is 1000, the feature sequence is a matrix of 6 rows and 1000 columns. Each column corresponds to the dissolution state feature of a feature window. The first row is the energy entropy feature, and the second to sixth rows are the normalized energy values ​​of the five frequency bands of the acoustic signature feature, which fully characterizes the dissolution state of carbon dioxide inside the melt.

[0029] In this implementation scheme, by embedding the melt rheological constitutive equation and the carbon dioxide dissolution and diffusion equation into the physical information neural network as physical constraint penalty terms of the loss function, the carbon dioxide dissolution rate and residence time distribution output by the network must simultaneously meet the dual requirements of data fitting accuracy and physical laws. The energy entropy feature extracted by wavelet packet decomposition of the power-compensated micro-variable time series reflects the fluctuation state of melt viscosity along the extrusion direction, and the acoustic signature feature extracted by frequency band energy integration of the acoustic emission energy attenuation spectrum reflects the activity of carbon dioxide bubble prenuclearization. These two types of features, together with barrel temperature deviation, melt pressure fluctuation, and screw speed fluctuation, constitute a multi-channel input matrix. After the network outputs the dissolution rate and residence time of twenty discrete sections in the extrusion direction, the concentration gradient field characterization data is obtained by cubic spline interpolation and first-order central difference calculation. Thus, the dissolution and diffusion process of carbon dioxide inside the melt is transformed into a numerical distribution with spatial location labels, providing a quantitative input for identifying the deterioration trend of dissolution uniformity inside the barrel, and avoiding the lack of information dimension caused by relying solely on surface process parameters for statistical inference.

[0030] Specifically, the physical information neural network uses the melt rheological constitutive equation and the carbon dioxide dissolution and diffusion equation as constraints as follows: In the loss function of the physical information neural network, a prediction error term and a physical constraint penalty term are constructed, specifically as follows: The total loss function of a neural network consists of a prediction error term and a physical constraint penalty term, which are weighted and summed by a penalty coefficient (0.8, used to balance the prediction accuracy and the degree of satisfaction of physical constraints). The prediction error term uses mean squared error loss to calculate the deviation between the carbon dioxide dissolution rate and residence time output by the neural network and the true values ​​obtained from the small-scale experiment. The physical constraint penalty term consists of the melt viscosity deviation penalty term and the carbon dioxide diffusion deviation penalty term, which correspond to the constraint deviations of the melt rheological constitutive equation and the carbon dioxide dissolution and diffusion equation, respectively. During the forward propagation of the physical information neural network, the prediction error term and the physical constraint penalty term are calculated simultaneously. Then, through backpropagation, both the prediction error and the degree of violation of physical constraints are minimized simultaneously. Specifically: During forward propagation, the input multi-channel input matrix is ​​processed by the ReLU activation function in the hidden layer (only input values ​​greater than 0 are retained, and input values ​​less than or equal to 0 are discarded). The output layer then outputs the predicted values ​​of carbon dioxide dissolution rate and residence time through a linear activation function. Simultaneously calculate the prediction error term, and then calculate the physical constraint penalty term based on the predicted melt viscosity and carbon dioxide concentration, thereby obtaining the total loss; During backpropagation, the Adam optimizer (learning rate set to 0.001, decay factor set to 0.9) is used to update the weights and biases of the neural network. The update direction is the negative gradient direction of the total loss. Iteration training is performed for 500 epochs until the total loss converges (convergence threshold set to 10). -4 ).

[0031] The pre-training steps for a physical information neural network are as follows: The pre-training dataset uses sample data collected from small-scale supercritical foaming production experiments. It covers time-series data of main process parameters and secondary mapping data under different production conditions, as well as the corresponding actual measured values ​​of carbon dioxide dissolution rate and residence time inside the melt. The actual measured values ​​of carbon dioxide dissolution rate and residence time inside the melt were obtained by sampling in segments along the extrusion direction and measuring the carbon dioxide content in each segment of the melt by gravimetric method. A total of 1000 valid samples were collected and divided into training set, validation set and test set in a ratio of 7:2:1. During the pre-training process, the neural network is first trained on the training set. The model performance is verified using the validation set every 50 training rounds. If the loss on the validation set does not decrease for three consecutive rounds, the pre-training is stopped using an early stopping strategy to avoid overfitting. After pre-training, the model's prediction accuracy is verified using a test set to ensure that the prediction errors for dissolution rate and residence time do not exceed 3%.

[0032] In this implementation scheme, the loss function of the physical information neural network consists of a prediction error term and a physical constraint penalty term. The physical constraint penalty term calculates the prediction deviation of melt viscosity and the diffusion deviation of carbon dioxide concentration based on the melt rheological constitutive equation and the carbon dioxide dissolution and diffusion equation, and sums them with the prediction error in a weighted manner. During forward propagation, a multi-channel matrix is ​​input, and after processing by the hidden layer activation function, the predicted values ​​of dissolution rate and residence time are output. The total loss is calculated synchronously, and the network weights and biases are updated through backpropagation. The Adam optimizer is used to iterate until the loss converges. During the pre-training stage, the sample data collected from small-scale experiments are divided into training set, validation set and test set according to the proportion. An early stopping strategy is introduced during the training process to prevent overfitting. After the pre-training is completed and the prediction error is verified to be no more than three percent, the model parameters are saved for real-time production use.

[0033] Specifically, such as Figure 2 As shown, the specific steps to obtain the fused feature data are as follows: The spatial distribution sequence of the concentration gradient field characterization data along the extrusion direction is concatenated with the time-aligned barrel temperature deviation time series, melt pressure fluctuation time series, and screw speed fluctuation time series to obtain the fused feature matrix, which is as follows: The spatial distribution sequence of the concentration gradient field characterization data along the extrusion direction consists of the positions of 20 discrete cross sections and their corresponding concentration gradient values, which are flattened into a spatial feature vector of 1 row and 40 columns. The timing data of the master process parameters after time alignment are all vectors with 1 row and a total number of acquisition points, for a total of 3. The spatial feature vector is concatenated with the time-series vectors of the three main process parameters using a vertical concatenation method. The resulting fusion feature matrix has 44 rows (4 rows of basic vectors plus 40 rows of spatial features) and the number of columns is the total number of collection points. For example, when the total number of collection points is 10,000, the fusion feature matrix is ​​44 rows and 10,000 columns, with each row corresponding to a feature and each column corresponding to the fusion feature value at a collection time. Principal component analysis (PCA) is performed on the fusion feature matrix to reduce its dimensionality. Principal component components with a cumulative contribution rate exceeding a preset threshold are retained as fusion feature data. Specifically: The fusion feature matrix is ​​standardized to eliminate the influence of dimensions. The standardization method is to subtract the mean of the corresponding column from each element and then divide it by the standard deviation of the corresponding column. Principal component analysis is performed on the standardized matrix to calculate its covariance matrix. The eigenvalues ​​and corresponding eigenvectors of the covariance matrix are then solved, and the contribution rate of each principal component and the cumulative contribution rate are calculated. The preset threshold is set to 85%. The top principal component components with a cumulative contribution rate of 85% or higher are selected as fused feature data, which not only preserves the core information of the original features but also reduces the data dimensionality.

[0034] The specific steps for identifying the deterioration trend of dissolution uniformity inside the output barrel are as follows: The fused feature data is divided into multiple feature subsequences according to a time sliding window. Each feature subsequence corresponds to the production status at a time step, specifically as follows: The number of rows in the fused feature data is equal to the total number of collected points, and the number of columns is equal to the number of principal component components. The length of the time sliding window is set to 100 data points, corresponding to 1 second of production time, and the sliding step is set to 50 data points, corresponding to 0.5 seconds. The segmentation method is to start from the first collection point, slide 50 collection points and then cut out a window of 100 collection points, until all fused feature data is segmented. The number of feature subsequences obtained is calculated based on the total number of collection points. Each feature subsequence has 100 rows, 100 columns, and corresponds to a production state at a time step (1 second). For example, when the total number of collection points is 10,000 and the number of principal components is 20, each feature subsequence is 20 rows and 100 columns, with a total of 199 feature subsequences, which correspond to the production status of 199 time steps respectively; Multiple feature subsequences are sequentially input into the dissolution uniformity degradation trend identification model, which outputs the dissolution uniformity state category corresponding to each time step, specifically: The dissolution uniformity degradation trend identification model is a time-series classification model based on Long Short-Term Memory (LSTM) network. The model input is a feature subsequence, which is processed by an LSTM layer (64 hidden neurons and a dropout coefficient of 0.2) to extract time-series features. After passing through a fully connected layer (with 32 neurons) and a softmax layer, four probability values ​​are output, corresponding to the probabilities of the normal state, slightly deteriorated state, moderately deteriorated state, and severely deteriorated state, respectively. For example, if the probability of a moderately deteriorated state is the highest after inputting a certain feature subsequence, then the dissolution uniformity state category corresponding to that time step is a moderately deteriorated state. By sequentially inputting all feature subsequences, the state category at each time step is obtained, forming a state category sequence. The state categories are represented by 0, 1, 2, and 3, respectively, indicating the normal state, slightly deteriorated state, moderately deteriorated state, and severely deteriorated state. The state transition sequence is constructed by classifying the dissolution uniformity states over multiple consecutive time steps, and is used as the result of identifying the deterioration trend of dissolution uniformity inside the barrel. Specifically: The state transition sequence is based on the state category sequence, retaining the state category of each time step and the state transition relationship between adjacent time steps. That is, a transition pair is formed by the state categories of two adjacent time steps, and all transition pairs are arranged in sequence to form the state transition sequence. For example, if the state category sequence is normal, normal, slightly deteriorated, slightly deteriorated, moderately deteriorated, moderately deteriorated, and severely deteriorated, then the state transition sequence is (normal, normal), (normal, slightly deteriorated), (slightly deteriorated, slightly deteriorated), (slightly deteriorated, moderately deteriorated), (moderately deteriorated, moderately deteriorated), (moderately deteriorated, severely deteriorated). This state transition sequence fully reflects the change process of the dissolution uniformity inside the barrel from normal to deterioration, and can clearly show the deterioration trend, serving as the result of identifying the deterioration trend of dissolution uniformity inside the barrel.

[0035] The specific steps for outputting the dissolution uniformity state category for each time step are as follows: The model for identifying the deterioration trend of dissolution uniformity is a time-series classification model based on a long short-term memory network, specifically as follows: This time series classification model consists of an input layer, an LSTM layer, a fully connected layer, and an output layer; the input layer has the same dimension as the feature subsequence and is used to receive the feature subsequence. The LSTM layer contains 64 hidden units and uses the tanh activation function to extract the temporal dependencies of feature subsequences; The fully connected layer contains 32 neurons and uses the ReLU activation function to perform nonlinear mapping on the temporal features extracted by the LSTM layer; The output layer contains 4 neurons and uses the softmax activation function to output the probabilities of the 4 dissolution uniformity state categories; During model training, the cross-entropy loss function and Adam optimizer were used, with 300 training epochs and a batch size of 32 to ensure that the model classification accuracy reached over 95%. This time-series classification model controls the degree of information discarded from the hidden state of the previous time step through forget gate units, specifically: The input to the forget gate unit is the input features of the current time step and the hidden state of the previous time step. The output is the forget gate coefficient, which ranges from 0 to 1 and is calculated by the sigmoid activation function. When the forget gate coefficient is close to 1, all information in the hidden state of the previous time step is retained; when the forget gate coefficient is close to 0, all information in the hidden state of the previous time step is discarded. For example, when the solubility uniformity is in a stable and normal state, the forgetting gate coefficient is close to 0.9, retaining most of the historical stable information; When there is a sudden change in the uniformity of dissolution, the forgetting gate coefficient approaches 0.2, discarding some invalid historical information; This time-series classification model controls the degree to which information is written into the input features at the current time step through input gate units, specifically as follows: The input gate unit consists of two parts. One part calculates the input gate coefficients through the sigmoid activation function to control the writing ratio of information. Another part generates candidate information for the current time step through the tanh activation function; The output of the input gate unit is the element-wise product of the input gate coefficient and the candidate information, that is, the candidate information is written into the cell state proportionally according to the input gate coefficient. For example, when the input features at the current time step contain key information about the deterioration of dissolution uniformity, the input gate coefficient is close to 0.8, and most of the candidate information is written in; This time-series classification model controls the amount of information passed from the hidden state of the current time step to the next time step through output gate units, specifically: The input to the output gate unit is the input features of the current time step and the hidden state of the previous time step. First, the output gate coefficients are calculated using the sigmoid activation function. The cell state is then processed using the tanh activation function; The output of the output gate unit is the hidden state of the current time step, which is the element-wise product of the output gate coefficient and the processed cell state. Based on the output gate coefficient, the proportion of information in the current cell state is passed to the next time step is controlled. For example, when the state category of the current time step is moderately deteriorated, the output gate coefficient is close to 0.9, which passes most of the deterioration-related information to the next time step; The hidden states of the last time step in the time-series classification model are extracted, mapped through a fully connected layer, and then the probability values ​​of each dissolution uniformity state category are calculated using a normalized exponential function. Specifically: For each feature subsequence containing 100 time steps, after inputting into the LSTM layer, 100 hidden states are obtained. The hidden state (64 dimensions) of the last time step is extracted. The hidden state is input into a fully connected layer and processed by the ReLU activation function to obtain a 32-dimensional feature vector. The feature vector is then input into the output layer, and the probability values ​​of the four state categories are calculated by the normalized exponential function (softmax function). The sum of the probability values ​​is 1. For example, after inputting a certain feature subsequence, the probability of a moderately degraded state is close to 0.84, which is the maximum probability value, corresponding to a moderately degraded state; The dissolution uniformity state category corresponding to the maximum probability value is taken as the output state category for this time step, specifically as follows: The four probability values ​​output for each feature subsequence are compared, and the state category corresponding to the highest probability value is selected as the output state category for that time step. The normal state corresponds to a dissolution uniformity deviation of no more than 5%, the slightly deteriorated state corresponds to a dissolution uniformity deviation between 5% and 15%, the moderately deteriorated state corresponds to a dissolution uniformity deviation between 15% and 30%, and the severely deteriorated state corresponds to a dissolution uniformity deviation exceeding 30%. For example, if the maximum probability value corresponds to a moderately deteriorated state, then the output state category for that time step is a moderately deteriorated state. And so on, to obtain the dissolution uniformity state category for each time step.

[0036] The pre-training steps for the dissolution uniformity degradation trend identification model (LSTM time series classification model) are as follows: The pre-training dataset uses fused feature data collected during historical production processes and corresponding dissolution uniformity state category labels. A total of 800 time-series samples were collected, covering four scenarios: normal state, slightly deteriorated state, moderately deteriorated state, and severely deteriorated state. Each sample contains a complete feature subsequence and a corresponding state category label. Before pre-training, the sample data is preprocessed to remove abnormal samples and the feature subsequences are standardized. During pre-training, the cross-entropy loss function and Adam optimizer are used. Samples are input in batches (batch size 32), and the training rounds are 300. Every 30 rounds, the model classification accuracy is verified using a validation set (accounting for 20% of the total samples). If the classification accuracy of the validation set does not improve for 5 consecutive rounds, the pre-training is stopped. After pre-training, the model performance is validated using a test set (accounting for 10% of the total samples) to ensure that the classification accuracy is not less than 95%.

[0037] In this implementation scheme, the concentration gradient field characterization data, after being flattened, is longitudinally spliced ​​with the time-aligned barrel temperature deviation, melt pressure fluctuation, and screw speed fluctuation to form a fusion feature matrix that simultaneously carries spatial physical field information and equipment operating status. Redundant dimensions are eliminated through principal component analysis, and principal component components with a cumulative contribution rate exceeding a preset threshold are retained as fusion feature data. This fusion feature data is divided into multiple feature subsequences according to a time sliding window and then sequentially input into a time-series classification model based on a long short-term memory network. The model uses forget gates, input gates, and output gates to control the selection and updating of historical state information, extracts the evolution law of dissolution uniformity in the time dimension, and outputs the state category corresponding to each time step. The state categories of consecutive time steps constitute a state transition sequence. When there is a unidirectional progression from slight deterioration to moderate deterioration or a direct occurrence of a state above moderate deterioration, it is determined that there is a risk of quality defects.

[0038] Specifically, such asFigure 3 As shown, the specific steps for determining the results of finished product quality defect risk monitoring are as follows: The state transition sequence in the analysis of the deterioration trend identification results of the dissolution uniformity inside the barrel is as follows: The state transition sequence consists of the state categories of two adjacent time steps, and the parsing process mainly includes: Statistically analyze the frequency and duration of each state category; Analyze the transition direction and frequency between adjacent state categories; For example, statistics show that the normal state has 80 time steps, the slightly deteriorated state has 30 time steps, the moderately deteriorated state has 50 time steps, and the severely deteriorated state has 39 time steps. The transfer directions mainly include normal state to normal state, normal state to slightly deteriorated state, slightly deteriorated state to normal state, etc. The occurrence frequency of each transfer direction is counted to clarify the deterioration path of dissolution uniformity. When a unidirectional transition from a slightly deteriorated state to a moderately deteriorated state occurs in the state transition sequence, or when a state category of moderate deterioration or above occurs, the finished product quality defect risk monitoring result is determined to be risky. Specifically: Unidirectional transfer refers to the process from a slightly deteriorated state to a moderately deteriorated state without any reverse transfer from the moderately deteriorated state back to the slightly deteriorated state or the normal state. For example, if a transfer sequence shows a slight degradation state to a moderate degradation state, a moderate degradation state to a moderate degradation state, or a moderate degradation state to a severe degradation state, it is considered a unidirectional transfer. The state categories of moderate degradation and above refer to moderate degradation and severe degradation. As long as either of the above two situations appears in the state transition sequence, the finished product quality defect risk monitoring result is determined to be risky. For example, if a transition from a slightly deteriorated state to a moderately deteriorated state or a severely deteriorated state occurs directly in the state transition sequence, it is considered a risk. This indicates that the uniformity of dissolution inside the melt has deteriorated significantly during the current production process, which may lead to quality defects in the finished product, such as uneven foaming and excessive deviation in cell size. When all state categories in the state transition sequence remain in a normal state, or only an isolated slightly deteriorated state exists and subsequently recovers to a normal state, the finished product quality defect risk monitoring result is determined to be risk-free. Specifically: "All state categories remain in a normal state" means that the state category is in a normal state at all time steps, indicating that the dissolution uniformity is always in a stable and normal state. An isolated slightly degraded state refers to a slightly degraded state that occurs for only 1-2 consecutive time steps and then immediately transitions back to the normal state; For example, the state transition sequence includes a normal state to a slightly deteriorated state and a slightly deteriorated state to a normal state, and this type of transition occurs only once, with no other deterioration transitions; If either of the above two conditions is met, the result of the finished product quality defect risk monitoring is determined to be risk-free, indicating that the internal dissolution uniformity of the melt is stable during the current production process and will not lead to quality defects in the finished product. The specific steps for locating the data source deviation that causes the deterioration of dissolution uniformity are as follows: The fusion feature subsequences corresponding to the time steps identified as having risk are extracted and determined from the fusion feature data, specifically as follows: All time steps with risks are identified to form a risk time step set; each risk time step corresponds to fused feature data from 100 collection points, and feature subsequences corresponding to each risk time step are extracted accordingly. The feature subsequences corresponding to all risk time steps are concatenated to obtain the risk fusion feature matrix. The number of rows in this matrix is ​​the number of principal component components, and the number of columns is the total number of collection points corresponding to all risk time steps. The contribution score of each dimension of the fused feature subsequence to the output of the dissolution uniformity degradation trend identification model is calculated using the SHAP value. Specifically: The contribution of each feature dimension is calculated using SHAP values. KernelExplainer is selected as the model-independent SHAP interpreter to adapt to the dissolution uniformity degradation trend identification model. The risk fusion feature matrix and the dissolution uniformity degradation trend identification model are input, and the SHAP value of each feature dimension at each risk time step is output. The contribution score is the average absolute value of the SHAP values ​​of this feature dimension across all risk time steps. The larger the contribution score, the greater the influence of this feature dimension on the deterioration trend of dissolution uniformity. For example, if the contribution score of one feature dimension is 0.8 and that of another feature dimension is 0.2, then the former has a much greater impact on the degradation trend than the latter. Feature dimensions whose contribution scores exceed a preset significance threshold are selected and mapped back to their corresponding original data source types. These original data source types include barrel temperature deviation, melt pressure fluctuation, screw speed fluctuation, power compensation micro-variation characteristics, and acoustic signature characteristics, specifically: The pre-defined significance threshold is set as the mean of all contribution scores. Feature dimensions whose contribution scores exceed this mean are selected to form a set of significant impact features. Based on the construction rules for fused feature data, each feature dimension in the feature set that significantly affects the data will be mapped back to the original data source type: The feature dimensions related to the concentration gradient field are mapped to power compensation micro-variation features or acoustic features (since the concentration gradient field is derived from the secondary mapping data), the mapping related to the barrel temperature deviation time series is the barrel temperature deviation, the mapping related to the melt pressure fluctuation time series is the melt pressure fluctuation, and the mapping related to the screw speed fluctuation time series is the screw speed fluctuation. For example, if three of the selected significant feature dimensions are related to power compensation micro-variation features and two are related to barrel temperature deviation, then the original data source type after mapping is power compensation micro-variation features and barrel temperature deviation. When the contribution score of power compensation micro-variation features or acoustic features in the original data source type after mapping exceeds the preset ratio, the deviation of the location data source is the anomaly of the physical field inside the melt corresponding to the secondary mapping data, which is as follows: Calculate the total contribution score of each original data source type after mapping, and then calculate the total contribution ratio of power compensation micro-variation features and voiceprint features, that is, the ratio of the total contribution score of the two to the total contribution score of all data source types after mapping. The preset ratio is set to 50%. When this ratio exceeds 50%, it indicates that the secondary mapping data has a dominant influence on the deterioration trend of dissolution uniformity. The deviation of the source of the location data is the anomaly of the physical field inside the melt corresponding to the secondary mapping data. For example, if the percentage is 65%, then the deviation of the data source is determined to be an anomaly in the physical field inside the melt, such as abnormal fluctuations in melt viscosity or abnormal prenucleation of carbon dioxide. When the contribution percentage of barrel temperature deviation, melt pressure fluctuation, or screw speed fluctuation in the mapped original data source type exceeds a preset ratio, the source deviation is identified as an abnormal disturbance of the corresponding process parameter, specifically: The total contribution percentage of the computer barrel temperature deviation, melt pressure fluctuation, and screw speed fluctuation, i.e., the proportion of the total contribution score of the three to the total contribution score of all mapped data source types. The preset ratio is set to 50%. When this ratio exceeds 50%, it indicates that the main process parameters have a dominant influence on the deterioration trend of dissolution uniformity, and the deviation of the source of the location data is an abnormal disturbance of the corresponding process parameters. For example, if the barrel temperature deviation accounts for 35% of the contribution and the melt pressure fluctuation accounts for 20%, with a total contribution of 55%, then the source deviation is the abnormal disturbance of the barrel temperature deviation and melt pressure fluctuation, such as the barrel temperature deviation exceeding ±2℃ and the melt pressure fluctuation exceeding ±0.5MPa.

[0039] The specific steps for generating production quality control traceability and early warning decision information based on the deviation of the data source are as follows: When the data source deviation corresponds to an anomaly in the physical field inside the melt corresponding to the secondary mapping data, a quality control early warning message is generated, which includes an identifier of the anomaly in the physical field inside the melt and a risk level indication. Specifically: The anomaly markers for physical field anomalies within the melt are determined based on the type of anomaly, and are divided into melt viscosity anomalies and carbon dioxide prenucleation anomalies; if the contribution of power compensation micro-variation characteristics is higher (more than 50%), it is marked as “melt viscosity anomaly”. If the contribution of voiceprint features is higher, it is marked as "abnormal movement of carbon dioxide prenucleation". The risk level is determined based on the proportion of time steps in which the risk exists, and is divided into three levels: Level 1 risk corresponds to a risk time step proportion of no more than 30%, Level 2 risk corresponds to a proportion between 30% and 60%, and Level 3 risk corresponds to a proportion of more than 60%. Quality control early warning information must clearly indicate the anomaly, risk level, risk time step, and source tracing prompts; For example, "Abnormality indicator: Abnormal fluctuation in melt viscosity;" Risk level: Level 2 risk; Risk time step: t5-t34; Source tracing tip: "The physical field inside the melt is abnormal, and the secondary mapping data acquisition process and the reaction state inside the melt need to be carefully checked." When the data source deviation is an abnormal disturbance of the corresponding process parameter, quality control analysis information containing the abnormality type of the corresponding process parameter and the conclusion of the abnormality tracing is generated, specifically as follows: The corresponding process parameter anomaly type is determined based on the contribution ratio of each process parameter. If the barrel temperature deviation has the highest contribution ratio, then the anomaly type is "barrel temperature deviation anomaly". If the contribution of melt pressure fluctuation is the highest, then it is considered "abnormal melt pressure fluctuation". If the screw speed fluctuation accounts for the largest proportion, then it is considered "abnormal screw speed fluctuation". The conclusions of the anomaly tracing are based on the abnormal range of process parameters and the collected data, which clarifies the time range of the anomaly and the possible direction of tracing its source. For example, anomaly type: abnormal barrel temperature deviation; Abnormal time range: t12-t45; Conclusion: The temperature deviation in zone 3 of the barrel consistently exceeds ±2℃, which may lead to fluctuations in melt viscosity and consequently deterioration in dissolution uniformity. It is recommended to check the data acquisition accuracy and operating status of the barrel heating device.

[0040] In this implementation plan, after parsing the state transition sequence, the risk of finished product quality defects is determined based on whether there is a one-way progression from slight degradation to moderate degradation or a direct occurrence of moderate degradation or above. At the same time, isolated slight degradation fluctuations that recover to normal are classified as risk-free, thereby avoiding false alarms triggered by instantaneous disturbances that disrupt the production rhythm. After the risk is determined to exist, the feature subsequences of the corresponding time step are extracted from the fused feature data, and the SHAP contribution scores of each dimension feature are calculated. High contribution features are screened and mapped back to five types of original data sources: barrel temperature deviation, melt pressure fluctuation, screw speed fluctuation, power compensation micro-variation features, and acoustic features. When the contribution ratio of power compensation micro-variation features and acoustic features exceeds the preset ratio, it is located as an anomaly in the physical field inside the melt; otherwise, it is located as an abnormal disturbance of the corresponding process parameters. Based on the location results, quality control early warning information containing anomaly identification and risk level or quality control analysis information containing anomaly type and source tracing conclusion are generated respectively.

[0041] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0042] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for quality control in the production of supercritical foamed materials based on big data, characterized in that, Includes the following steps: Acquire time-series data and secondary mapping data related to the main process parameters during the continuous extrusion production of supercritical foamed materials. The time-series data related to the main process parameters include barrel temperature deviation, melt pressure fluctuation, and screw speed fluctuation. Based on the time-series data and secondary mapping data related to the main process parameters, the correspondence between the carbon dioxide dissolution rate distribution and the residence time distribution inside the melt is determined, and the concentration gradient field characterization data evolving along the extrusion direction is generated. By fusing concentration gradient field characterization data with time-series data related to main process parameters, fused feature data is obtained. The fused feature data is input into the dissolution uniformity degradation trend identification model, and the output is the dissolution uniformity degradation trend identification result inside the barrel; Based on the identification results of the deterioration trend of the internal dissolution uniformity of the barrel, the risk monitoring results of finished product quality defects corresponding to the internal physical field anomalies are determined; When the results of quality defect risk monitoring indicate that there is a risk, the cause tracing analysis is carried out based on the fused feature data to locate the data source deviation that causes the trend of deterioration of dissolution uniformity, and the production quality control traceability and early warning decision information is generated based on the located data source deviation.

2. The method for quality control of supercritical foamed material production based on big data according to claim 1, characterized in that, The specific steps for generating concentration gradient field characterization data evolving along the extrusion direction are as follows: Feature extraction is performed on the secondary mapping data to obtain the feature sequence of carbon dioxide dissolution state inside the melt, and time-axis alignment is performed with the time series data related to the main process parameters to construct a multi-channel input matrix. The multi-channel input matrix is ​​input into the physical information neural network. The physical information neural network is constrained by the melt rheological constitutive equation and the carbon dioxide dissolution and diffusion equation, and outputs the correspondence between the carbon dioxide dissolution rate distribution and the residence time distribution at each discrete cross-section position along the extrusion direction inside the melt. Spatial interpolation and gradient calculation are performed on the correspondence between the carbon dioxide dissolution rate distribution and the residence time distribution at each discrete cross-section location to generate concentration gradient field characterization data that evolves along the extrusion direction.

3. The method for quality control of supercritical foamed material production based on big data according to claim 2, characterized in that, The specific steps to obtain the characteristic sequence of the carbon dioxide dissolution state inside the melt are as follows: Wavelet packet decomposition was performed on the power compensation micro-variable time series in the secondary mapping data to extract the energy entropy feature sequence. Frequency band energy integration is performed on the acoustic emission energy attenuation spectrum time series in the secondary mapping data to extract the acoustic signature feature sequence; The energy entropy feature sequence and the acoustic signature feature sequence are concatenated to form a feature sequence of the carbon dioxide dissolution state inside the melt.

4. The method for quality control of supercritical foamed material production based on big data according to claim 2, characterized in that, The specific steps of the physical information neural network, using the melt rheological constitutive equation and the carbon dioxide dissolution and diffusion equation as constraints, are as follows: In the loss function of the physical information neural network, a prediction error term and a physical constraint penalty term are constructed; During the forward propagation of the physical information neural network, the prediction error term and the physical constraint penalty term are calculated simultaneously, and the prediction error and the degree of violation of physical constraints are minimized simultaneously through back propagation.

5. The method for quality control of supercritical foamed material production based on big data according to claim 1, characterized in that, The specific steps to obtain the fused feature data are as follows: The spatial distribution sequence of the concentration gradient field characterization data along the extrusion direction is dimensionally spliced ​​with the time-aligned barrel temperature deviation time series, melt pressure fluctuation time series, and screw speed fluctuation time series to obtain the fused feature matrix. Principal component analysis is performed on the fusion feature matrix to reduce its dimensionality. Principal component components with a cumulative contribution rate exceeding a preset threshold are retained as fusion feature data.

6. The method for quality control of supercritical foamed material production based on big data according to claim 5, characterized in that, The specific steps for identifying the deterioration trend of dissolution uniformity inside the output barrel are as follows: The fused feature data is divided into multiple feature subsequences by a time sliding window, and each feature subsequence corresponds to the production status at a time step. Multiple feature subsequences are sequentially input into the dissolution uniformity degradation trend identification model, and the dissolution uniformity state category corresponding to each time step is output. The state transition sequence is formed by classifying the dissolution uniformity states over multiple consecutive time steps, and is used as the result of identifying the deterioration trend of dissolution uniformity inside the barrel.

7. The method for quality control of supercritical foamed material production based on big data according to claim 6, characterized in that, The specific steps for outputting the dissolution uniformity state category for each time step are as follows: The model for identifying the deterioration trend of dissolution uniformity is a time-series classification model based on long short-term memory networks; This time-series classification model controls the degree of information discarded from the hidden state of the previous time step through forget gate units; This time-series classification model controls the degree to which information is written into the input features at the current time step through input gate units; This time series classification model controls the amount of information passed from the hidden state of the current time step to the next time step through the output gate unit; The hidden state of the last time step of the time series classification model is extracted, mapped through a fully connected layer, and the probability value of each dissolution uniformity state category is calculated by a normalized exponential function. The state category of dissolution uniformity corresponding to the maximum probability value is taken as the output state category for that time step.

8. The method for quality control of supercritical foamed material production based on big data according to claim 6, characterized in that, The specific steps for determining the results of finished product quality defect risk monitoring are as follows: Analyze the state transition sequence in the identification results of the deterioration trend of dissolution uniformity inside the barrel; When a unidirectional transition from a slightly deteriorated state to a moderately deteriorated state occurs in the state transition sequence, or when a state category of moderate deterioration or above occurs, the finished product quality defect risk monitoring result is determined to be risky. When all state categories in the state transition sequence remain in a normal state, or only an isolated slightly deteriorated state exists and subsequently recovers to a normal state, the finished product quality defect risk monitoring result is determined to be risk-free.

9. The method for quality control of supercritical foamed material production based on big data according to claim 8, characterized in that, The specific steps for locating the data source deviation that causes the deterioration of dissolution uniformity are as follows: Extract and identify fusion feature subsequences corresponding to time steps that are at risk from the fusion feature data; The contribution score of each dimension of the fused feature subsequence to the output of the dissolution uniformity degradation trend identification model is calculated using the SHAP value. Feature dimensions whose contribution scores exceed a preset significant threshold are selected and mapped back to their corresponding original data source types. The original data source types include barrel temperature deviation, melt pressure fluctuation, screw speed fluctuation, power compensation micro-variation features, and acoustic features. When the contribution score of power compensation micro-variation feature or acoustic feature in the original data source type after mapping exceeds the preset ratio, the deviation of the location data source is the anomaly of the physical field inside the melt corresponding to the secondary mapping data. When the contribution score of barrel temperature deviation, melt pressure fluctuation, or screw speed fluctuation in the mapped original data source type exceeds the preset ratio, the deviation of the location data source is identified as an abnormal disturbance of the corresponding process parameter.

10. The method for quality control of supercritical foamed material production based on big data according to claim 9, characterized in that, The specific steps for generating production quality control traceability and early warning decision information based on the deviation of the data source are as follows: When the deviation of the data source is the anomaly of the physical field inside the melt corresponding to the secondary mapping data, a quality control early warning information containing the anomaly identifier of the physical field inside the melt and the risk level prompt is generated. When the deviation at the data source is an abnormal disturbance of the corresponding process parameter, quality control analysis information containing the abnormality type of the corresponding process parameter and the conclusion of the abnormality tracing is generated.

Images