Die temperature feedforward compensation method based on polyvinyl chloride skin layer thickness prediction

Abstract

The present application relates to a die temperature feedforward compensation method based on polyvinyl chloride skin layer thickness prediction, aiming at the problems of inaccurate traditional process stage division, weak correlation of key variables and process weight adjustment lag, a die temperature dynamic weighted distribution method based on time sequence process parameter multi-scale semantic labeling, causal structure learning and storage is proposed. Through segmented preprocessing, key event detection and time sequence classification, the different process stages of the equipment are accurately identified and the weights are associated, combined with the ternary target observation data, the standardized causal graph of process target and die temperature is constructed, the smooth migration and conflict buffer mechanism of process weight at stage switching are realized. Finally, the traceability of the whole process of process weight distribution is guaranteed by the storage. The scheme improves the stage adaptability and target collaborative optimization effect of process control, and facilitates knowledge sedimentation and quality traceability.

Description

Technical Field

[0001] This invention relates to the field of multi-objective die temperature compensation control technology in polyvinyl chloride (PVC) extrusion processes, and in particular to a die temperature feedforward compensation method based on the prediction of PVC skin layer thickness. Background Technology

[0002] Die temperature compensation in PVC extrusion is a core element affecting multiple objectives, including product skin thickness, surface finish, and process stability. With increasing automation, the industry commonly employs multi-objective collaborative compensation strategies to address complex disturbances and nonlinear coupling in the production process. Current mainstream solutions generally use fixed-weight multi-objective optimization methods, where static weights for different objectives such as thickness, surface finish, and stability are set based on experience or limited experimental results at the initial production stage, serving as the priority for die temperature compensation. A few advanced technologies attempt to introduce rule-based fuzzy decision-making, analytic hierarchy process (AHP), non-dominated ranking evolutionary algorithm (NSGA), multi-objective particle swarm optimization (MOPSO), data-driven LSTM models, or automatic path optimization based on reinforcement learning frameworks (such as DQN) to improve the global optimization performance of multi-objective compensation.

[0003] This type of technical solution plays a positive role in the stable production process of PVC skin layer, reducing index fluctuations caused by human error, and is widely used in high-end manufacturing production lines and smart factory system integration. However, in actual continuous extrusion processes, the multi-objective weight requirements involved in die temperature compensation dynamically change with process stages (such as equipment startup, normal production, raw material change, and shutdown cleaning) and external disturbances (such as raw material batch switching and ambient temperature changes): some stages prioritize thickness uniformity, while others emphasize surface finish or system stability. Existing fixed weight allocation mechanisms struggle to maintain a flexible dynamic balance among the objectives, especially during stage switching, batch switching, or abnormal disturbances, easily leading to problems such as compensation response lag, decreased control accuracy, and mismatch of some objectives.

[0004] Typical "black box" methods based on nonlinear optimization algorithms or deep learning, while demonstrating good global optimization capabilities in some published literature, suffer from high barriers to practical deployment, significant user debugging difficulties, and heavy maintenance burdens due to the lack of interpretability in weight allocation and target switching processes, low parameter transparency, and high dependence on large-scale data, computing resources, and professional algorithm tuning teams. Furthermore, most data-driven methods often struggle to adapt to the complex conditions of real-world production lines, where stage identification is ambiguous and target priority changes frequently. Their weight switching often exhibits control problems such as jitter and overshoot, which are detrimental to ensuring batch stability and quality consistency.

[0005] The known multi-objective die temperature compensation weight setting methods in this field have the following main technical drawbacks: The fixed weight allocation mode leads to response lag and cannot adapt to the changing demands for multi-objective weights at different process stages, thus limiting the intelligence level of the compensation strategy. For production processes with frequent dynamic switching, the control results are unstable and prone to objective conflicts. Weight adaptation methods based on preset optimization algorithms or deep learning networks lack clear causal links, and their control behavior is unexplainable, making it difficult to provide process engineers with transparent and traceable auxiliary decision-making prompts. Existing solutions lack real-time weight transfer and dynamic smoothing mechanisms when dealing with changes in target priority at different stages, abnormal target achievement, and sudden process disturbances. As a result, compensation failures and target achievement degradation often occur. The management of system operation data mostly relies on traditional database storage, which makes it difficult to support distributed and tamper-proof knowledge accumulation and production audit backtracking, thus affecting the reuse of process knowledge and the accumulation of experience.

[0006] Therefore, the industry urgently needs a novel multi-objective die temperature compensation weight allocation method that features high-precision identification of process stages, a clear causal mechanism between target and compensation response, smooth switching of dynamic weights, and traceable control behavior. This method should not only cover the actual needs of all stages, multiple targets, and multiple operating conditions, but also reduce engineering deployment risks, achieve intelligent multi-objective compensation, and fundamentally improve process adaptability, thereby effectively solving the prominent problems of existing technologies in flexible weight switching, control interpretability, and knowledge accumulation. Summary of the Invention

[0007] This application provides a die temperature feedforward compensation method based on the prediction of polyvinyl chloride skin layer thickness, which aims to solve one of the problems or issues of the prior art mentioned in the background.

[0008] The die temperature feedforward compensation method based on the prediction of PVC skin layer thickness provided in this application specifically includes: S1: Obtain historical process parameters of the polyvinyl chloride extrusion process, align and segment them, and generate a stage tag sequence.

[0009] S2: Based on the stage label sequence, collect the crust thickness, surface roughness and PID deviation data within the corresponding time period to construct the target observation matrix.

[0010] S3: Perform causal structure learning on the observation matrix, eliminate spurious correlation paths, identify the intensity and direction of the direct causal effect of the mold temperature disturbance on the skin layer thickness, surface finish and process stability, and generate a compensated causal map.

[0011] S4: Obtain the real-time process parameter stream and input it into the timing classifier, output the current stage label and drift probability, and determine whether to trigger the stage switching signal based on the confidence level crossing the threshold.

[0012] S5: Activate the corresponding compensation causal graph based on the current stage label, extract the in-degree normalized value of the node as the initial weight, and perform weight migration when a switching signal is detected.

[0013] S6: Perform exponential smoothing calculation on the initial weights within the transition window to complete the smooth weight switching and generate a weight allocation vector.

[0014] S7: Based on the weight allocation vector, monitor the target achievement status. If the target does not reach the preset threshold and its causal strength coefficient is higher than the limit value, temporarily increase the target weight and trigger the target conflict buffer mechanism.

[0015] S8: Generate a switching record based on the stage switching signal, generate a migration log based on the weight migration process, and store the weight allocation vector, switching record and migration log together to form a weight allocation execution file.

[0016] The die temperature feedforward compensation method based on the prediction of PVC skin layer thickness provided in this application has the following beneficial effects: (1) By introducing a lightweight time-series classification and multi-scale segmentation mechanism based on key process anchor points, and combining real production line dynamic features such as die temperature change, current step, and pressure platform, the system achieves high-precision identification of four semantic stages of the production process (start-up stage, steady-state stage, material change transition stage, and shutdown stage), which significantly improves the system's perception and response sensitivity to complex working conditions. Compared with traditional static partitioning or fixed time window analysis methods, this scheme can adaptively capture the process state drift trend and predict the dynamic evolution path before stage switching, effectively overcoming the control delay problem caused by pattern recognition lag, ensuring the timeliness and context matching of subsequent multi-objective compensation strategies, especially showing stronger robustness and adaptability in non-steady-state scenarios such as frequent material changes and intermittent operation.

[0017] (2) A causal structure learning framework is constructed to accurately identify the direct causal effects of die temperature disturbance on skin thickness, surface roughness and temperature control deviation accumulation under controllable variables such as screw speed, melt pressure and cooling wind speed. The pseudo-correlation interference widely present in traditional correlation analysis is eliminated, and a dynamic causal map with physical interpretability is generated. Four maps divided according to process stages and with topological correlation respectively depict the driving relationship between quality targets under different operating conditions. The normalized value of the node in-degree is used to initialize the compensation weight, so that the initial decision is closer to the actual process mechanism. In the online stage, the weight transfer protocol is triggered by the confidence level of the detection stage crossing the threshold, and the map switching is completed by exponential smoothing within the transition window to avoid system oscillation caused by sudden changes in control parameters. This greatly reduces the instability risk in the multi-objective collaborative control process and realizes the paradigm shift from "experience-based parameter tuning" to "causal guidance".

[0018] (3) Design a target conflict buffer mechanism and a closed-loop system for evidence storage to further enhance the system's security, transparency, and human-machine collaboration capabilities: When a key quality indicator continuously deviates from the set range and is in a strong causal driving position in the current causal graph, the system automatically increases the weight of the target to prioritize correction, and pushes a "target mismatch warning" to the operation interface and highlights the relevant causal path, providing engineers with traceable decision-making basis and supporting manual intervention and knowledge feedback; All graph updates, weight migrations, and stage switching records are stored to ensure that the process optimization process is fully traceable and tamper-proof, meeting the requirements of industrial auditing and accumulating a reusable causal knowledge base, promoting the enterprise from "data-driven" to "knowledge-driven". The overall solution is based entirely on real production line causal modeling, does not rely on complex optimization or learning frameworks such as NSGA, MOPSO, DQN, and LSTM, does not require a large amount of hyperparameter adjustment and offline training, has low deployment cost and strong interpretability, and is particularly suitable for actual production environments with high mixed batches, small batches, and many changes, with good promotion value and engineering practicality.

[0019] In summary, this solution integrates temporal semantic segmentation, causal graph modeling, dynamic weight transfer, and conflict buffering mechanisms to construct an interpretable, adaptive, and robust multi-objective collaborative control system. This system not only significantly improves the quality stability and response agility of the process but also achieves a high degree of coupling between control logic and actual production. While ensuring automated operation, it retains space for human intervention, forming a closed-loop chain of "perception-reasoning-decision-audit," providing a new technical path for intelligent control in complex manufacturing scenarios. Attached Figure Description

[0020] Figure 1 This is the main flowchart of the die temperature feedforward compensation method based on the prediction of PVC skin layer thickness.

[0021] Figure 2 This is a sub-flowchart of the die temperature feedforward compensation method based on the prediction of PVC skin layer thickness.

[0022] Figure 3 This is another sub-flowchart of the die temperature feedforward compensation method based on the prediction of PVC skin layer thickness. Detailed Implementation

[0023] 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.

[0024] 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.

[0025] like Figure 1 As shown, this application provides a die temperature feedforward compensation method based on the prediction of polyvinyl chloride skin layer thickness, specifically including: S1: Obtain historical process parameters of the polyvinyl chloride extrusion process, align and segment them, and generate a stage tag sequence.

[0026] S2: Based on the stage label sequence, collect the crust thickness, surface roughness and PID deviation data within the corresponding time period to construct the target observation matrix.

[0027] S3: Perform causal structure learning on the observation matrix, eliminate spurious correlation paths, identify the intensity and direction of the direct causal effect of the mold temperature disturbance on the skin layer thickness, surface finish and process stability, and generate a compensated causal map.

[0028] S4: Obtain the real-time process parameter stream and input it into the timing classifier, output the current stage label and drift probability, and determine whether to trigger the stage switching signal based on the confidence level crossing the threshold.

[0029] S5: Activate the corresponding compensation causal graph based on the current stage label, extract the in-degree normalized value of the node as the initial weight, and perform weight migration when a switching signal is detected.

[0030] S6: Perform exponential smoothing calculation on the initial weights within the transition window to complete the smooth weight switching and generate a weight allocation vector.

[0031] S7: Based on the weight allocation vector, monitor the target achievement status. If the target does not reach the preset threshold and its causal strength coefficient is higher than the limit value, temporarily increase the target weight and trigger the target conflict buffer mechanism.

[0032] S8: Generate a switching record based on the stage switching signal, generate a migration log based on the weight migration process, and store the weight allocation vector, switching record and migration log together to form a weight allocation execution file.

[0033] Step S1: Obtain historical process parameters for the polyvinyl chloride extrusion process, align and segment them, and generate a stage tag sequence. Specifically, this includes: S1.1: Collect and preprocess the process parameters of the entire process of historical stable production batches in the polyvinyl chloride extrusion process, extract the original data streams of die temperature setpoint, main motor current and material pressure, and generate a standardized set of historical process parameters after noise filtering and outlier removal.

[0034] It should be noted that the alignment process refers to aligning the main motor current step point with the material pressure plateau starting point with the moment when the die temperature setpoint changes abruptly as the zero point of time; the segmentation process refers to treating the time interval between two adjacent anchor points as an independent process stage.

[0035] The input conditions for collecting and preprocessing the historical stable production batch process parameters of PVC extrusion are the raw output signal records of temperature sensors, main motor current sensors, and material pressure sensors stored in the production line data acquisition server. The die temperature setpoint data stream is calculated and extracted from the historical batch records, with its sampling frequency and timestamp information obtained from the temperature control system control board logs. The main motor current data stream is read from the motor drive control module interface to ensure synchronous acquisition of current values ​​and load status information. The material pressure data stream is obtained from the extrusion process pressure acquisition subsystem, with accompanying real-time sampling timestamps to support subsequent alignment operations. A bandpass filtering algorithm is applied to the above three types of data streams to eliminate high-frequency noise components; the filter parameters are set based on the sensor characteristic curves and the actual signal spectrum distribution. An outlier removal operation is performed on the filtered data streams, using the three-standard-deviation principle to identify and delete data points exceeding the normal fluctuation range. During the removal process, the mean and standard deviation are calculated, and the mean μ and fluctuation range are defined respectively. The remaining valid data are standardized by transforming the values ​​to a zero-mean, unit-variance space to eliminate differences in dimensions between different data sources. The standardization formula is as follows: ,in The sample mean. The standard deviation of the sample is 1. For the standardization results, the original multi-source signal is transformed into a standardized set of historical process parameters after noise filtering and outlier removal through the above filtering, outlier removal and standardization processes, so as to achieve high-precision input conditions for subsequent first-order differential mutation point detection.

[0036] For example, for 250 stable production batches of a continuously operating PVC extrusion line, the sampling frequency for the die temperature setpoint data stream is 2Hz, the sampling frequency for the main motor current data stream is 10Hz, and the sampling frequency for the material pressure data stream is 5Hz. A bandpass filter is used for the die temperature setpoint, with cutoff frequencies set at 0.1Hz and 1Hz. After filtering, the signal peak-to-valley fluctuation amplitude is significantly reduced. Outliers in the main motor current and material pressure data are removed using a three-times-standard-deviation rule. The mean values ​​μ are calculated to be 185.4A and 7.8MPa, respectively, and the standard deviations σ are calculated to be 4.2A and 0.6MPa, respectively. After removal, the data integrity remains above 98%. All signals are processed according to the formula... Standardization yields a data matrix with zero mean and unit variance, ranging from -2.5 to 2.3. This data matrix serves as a standardized set of historical process parameters, which is then input into the subsequent differential detection module. This significantly improves the sensitivity of abrupt change point identification, and the classification accuracy of production line process stages has been verified to be significantly enhanced.

[0037] S1.2: Based on the standardized historical process parameter set, the first-order difference algorithm is used to detect the abrupt change point of the die temperature setpoint, the step change point of the main motor current, and the characteristic point of the material pressure plateau period, so as to generate a key process event anchor point sequence containing accurate timestamps.

[0038] Based on the obtained standardized historical process parameter set, differential analysis input channels were established for three time series data: die temperature setpoint, main motor current, and material pressure. Fixed-length sliding window data blocks were constructed with second-level sampling intervals. First-order differential operations were performed on each data block, and the difference between adjacent sampling points was used as the rate of change sequence to eliminate the influence of absolute value offset on abrupt change detection. For the die temperature setpoint differential sequence, a temperature abrupt change judgment threshold was set, and the absolute value comparison method was used to filter out differential data points exceeding the threshold, recording the corresponding original timestamps as temperature abrupt change events. For the main motor current differential sequence, an amplitude step detection algorithm was used to identify events where the current value change within a single sampling period reached a set step amplitude, and the timestamps were extracted to form a current step change event set. For the original material pressure sequence, continuous zero values ​​or low rate of change intervals in the differential sequence were combined with variance threshold testing to identify plateau characteristics, and the start and end times of the plateau period were extracted to form a pressure plateau event set. The timestamps of the above three types of events were labeled according to event type and merged to form a key process event anchor point sequence containing event type and time precision identifiers. By employing differential detection and thresholding logic processing, the standardized historical process parameter set from the previous step is transformed into key process event anchor data with precise time indexes, enabling event-driven support for subsequent multi-source timestamp alignment and multi-scale segmentation processing. For example, in the historical batch data of a PVC extrusion production line, the die temperature setpoint sampling frequency is 1Hz, and the set temperature abrupt change threshold is 2℃. For this sequence, a first-order differential calculation formula is executed at each sampling point: in The change in temperature For the first The temperature setpoint is in seconds. Moments in the differential results where the absolute value of the temperature change is greater than 2 are recorded as temperature abrupt change events. The threshold for main motor current step detection is set to 5A, and the differential formula is: in The change in current For the first The current value per second is used, and the timestamp when the absolute value of the current change exceeds 5 is included in the step event set. In the detection of material pressure plateau periods, the variance threshold is set to 0.01 MPa². If the rate of change of the difference sequence is less than 0.002 MPa / s and the variance is less than the threshold within 30 consecutive seconds, then this interval is identified as a pressure plateau period, and its start and end times are extracted. The processing results show 15 temperature abrupt change events, 8 current step events, and 3 pressure plateau periods, forming a key process event anchor sequence with event type identifiers and second-level timestamps. This can provide accurate window positioning basis in subsequent multi-source data time alignment.

[0039] S1.3: Based on the key process event anchor point sequence, a dynamic time warping strategy is used to perform timestamp alignment processing on the multi-source sensor data, and a multi-scale sliding window segmentation operation is performed to generate a multi-scale process parameter segmented dataset with a unified time base.

[0040] S1.4: Apply a time-series classification model to the multi-scale process parameter segmented dataset for pattern recognition, and calculate the category belonging probability of each segment in combination with manual annotation rules to generate an initial process stage classification result with confidence scores.

[0041] Feature encoding is performed on the input of the multi-scale process parameter segmented dataset to extract normalized time-series vectors of the die temperature setpoint change curve, main motor current step feature, and material pressure plateau feature, respectively, as multi-channel inputs to the classification model. These multi-channel input vectors are loaded into the convolutional feature extraction layer of the time-series classification model. Local convolutional kernel scanning captures the local pattern distribution of various process events in the time domain, generating a multi-scale convolutional feature mapping matrix. A bidirectional gated recurrent unit (GRU) operation is performed on the multi-scale convolutional feature mapping matrix, using forward and backward state propagation to establish long-range dependencies across time slices, forming a high-dimensional time-series feature tensor containing process stage trend information. This high-dimensional time-series feature tensor is input into the fully connected output layer of the classification model to perform activation function transformation, and the classification probability is calculated using a soft maximum probability distribution, where the classification probability calculation formula is: in For class conditional probability, For category indexing, For the input feature vector, The output is a linear mapping of the corresponding category. Index all categories. Load the manually labeled rules into the rule comparison module, perform logical matching operations on each element of the attribution probability vector, confirm the category identifiers with probabilities higher than the set confidence threshold, and generate initial process stage classification results with accompanying confidence values. Through the above processing method, the segmented data from the previous step is transformed into initial classification data with probability distribution and confidence index, realizing stage pattern recognition of multi-source process features.

[0042] For example, a multi-scale dataset containing 500 segments was selected from a stable batch of the PVC extrusion process. The die temperature setpoint variation curve was normalized to the range of [-1, 1]. The maximum amplitude of the main motor current step amplitude after standardization was 0.45, and the mean value of the feature vector during the material pressure plateau period was 0. The convolutional feature extraction layer of the time-series classification model was configured with a kernel width of 3, a channel count of 16, and a stride of 1; the number of bidirectional GRU units was set to 32, and the Dropout ratio was 0.2. The output dimension of the fully connected layer was 4, corresponding to the start-up segment, steady-state segment, material change transition segment, and shutdown segment. The soft maximum probability distribution calculated that the probability of belonging to the steady-state segment was 0.82, and the probabilities for the other segments were 0.05, 0.08, and 0.05, respectively. The manual labeling rule sets the confidence threshold to 0.8. When the probability of a steady-state segment being assigned reaches 0.82, the classification result is confirmed, and the initial process stage classification label with a confidence level of 0.82 is finally output. This result then enters step S1.5 for effective segment screening, which significantly improves the classification accuracy and significantly reduces the misjudgment rate of stage switching judgment.

[0043] S1.5: Based on the initial process stage classification results, the duration distribution of consecutive occurrences of each category is statistically analyzed and a confidence filtering threshold is set to filter out effective segments that meet the definitions of start-up segment, steady-state segment, material change transition segment and shutdown segment, so as to generate a stage label sequence with confidence threshold and typical duration interval.

[0044] Step S2: Based on the stage label sequence, collect data on crust thickness, surface roughness, and PID deviation within the corresponding time period to construct a target observation matrix. Specifically, this includes: S2.1: Obtain the time interval boundary value corresponding to each label in the stage label sequence, and use the timestamp interpolation algorithm to resample the entire process parameter stream of the historical stable production batch to generate a synchronous process parameter time series that is strictly aligned with the semantic label time axis, thereby eliminating the time misalignment error caused by the difference in the frequency of multi-source data acquisition.

[0045] The input objects are the stage label sequence obtained after step S1.5 and the entire process parameter stream of historical stable production batches. The start and end boundary values ​​of the time interval corresponding to each label in the semantic label sequence are extracted to form a stage time boundary index set. The original timestamps of the multi-source sensor data in the historical process parameter stream are compared with the stage time boundary index set to locate the parameter segments corresponding to each stage. A timestamp interpolation algorithm is performed on these segments to construct a unified time reference for the sampling time series of each sensor, and a discrete time sampling point set is generated with the shortest sampling interval as the target sampling step size. Interpolation operations are performed on the parameter values ​​of missing sampling points, using the following interpolation formula: in, The interpolated parameter values ​​are the target values. The parameter value of the most recent sampling point. The parameter value of the most recent sampling point. For the target sampling point time, , These represent the most recent and consecutive sampling times. The interpolated multi-source parameter data are sorted by time index to form a synchronized process parameter time series with a unified time reference. Through the above resampling and interpolation processing, the stage label sequence from the previous step is transformed into a synchronized process parameter time series strictly aligned with the time axis, thus eliminating time misalignment errors caused by differences in the frequency of multi-source data acquisition.

[0046] For example, in PVC extrusion production, the stage time boundary index set includes the start-up stage [0s, 120s], the steady-state stage [121s, 1800s], the material changeover stage [1801s, 1950s], and the shutdown stage [1951s, 2100s]. In the historical process parameters, the die temperature setpoint sampling interval is 2s, the main motor current sampling interval is 1s, and the material pressure sampling interval is 3s. After locating the corresponding parameter segments for each stage, a complete set of sampling points is generated with a target sampling step size of 1s. In the steady-state stage, a certain die temperature sampling point is missing for 300s. The most recent sampling point is at 298s with a temperature value of 185.2℃, and the most recent sampling point is at 302s with a temperature value of 186.0℃. Substituting these values ​​into the interpolation formula, the calculation is: v = 185.2 + (186.0 + 186.0 + 185.2 + ... 185.2) / (302 298)×(300 (298) The interpolated temperature value for 300 seconds was 185.6℃. After this processing, all parameter data were arranged on a unified time series, eliminating the problems of original sampling frequency differences and time misalignment, and providing high-precision and synchronous basic data support for subsequent ternary target data extraction.

[0047] S2.2: Based on the time series of the synchronous process parameters, the laser thickness gauge data interface, the surface roughness meter scanning data interface, and the die temperature control system log interface are called respectively to extract the set of measured values ​​of the skin layer thickness, the set of arithmetic mean deviation values ​​of the surface roughness, and the set of cumulative PID deviation of the die temperature control system within the corresponding time interval, so as to form the original three-dimensional target data set.

[0048] S2.3: Perform sliding window statistical filtering on the original ternary target data set to remove abnormal noise data that exceeds three times the standard deviation, and use linear interpolation to complete the missing data segments caused by instantaneous sensor failure, so as to generate a cleaned and repaired high-completeness ternary target observation dataset.

[0049] S2.4: Based on the high-integrity ternary target observation dataset, the real-time values ​​of screw speed, melt pressure, and cooling wind speed are selected from the time series of synchronous process parameters as control variable constraints. A multi-dimensional data structure containing control variable columns and ternary target columns is constructed to form a preliminary constrained target observation matrix.

[0050] Based on the cleaned and repaired high-integrity ternary target observation dataset, the real-time sensor data interface in the synchronous process parameter time series is called to accurately extract the real-time value of the screw speed and generate a screw speed data column, so as to ensure a one-to-one correspondence between the data column and the ternary target observation dataset in terms of time index.

[0051] Perform the same time index matching operation on the real-time value of melt pressure, extract the second-level records of the pressure sensor through the data filtering function and form a melt pressure data column, ensuring that this data column shares the same time axis as the screw speed data column and the three-dimensional target column.

[0052] The real-time value of cooling airflow is filtered and extracted, and the recorded data of the airflow measurement and control unit is used to generate a cooling airflow data column, which forms a control variable set with the first two control variables.

[0053] The control variable set is concatenated with the three-dimensional target observation dataset in time index order to construct a multidimensional matrix structure containing three columns of control variables (screw speed, melt pressure, and cooling wind speed) and three columns of target data (skin layer thickness, surface finish, and PID deviation accumulation).

[0054] Perform format consistency verification on the multidimensional matrix structure, check the data column types, time index continuity and missing value imputation status, and output a preliminary target observation matrix with constraints.

[0055] By using the above column concatenation and format consistency verification methods, the high-completeness observation dataset from the previous step is transformed into multi-dimensional structured data containing three types of targets and three types of control variables, thus making the constraints explicit before causal structure learning.

[0056] For example, on a polyvinyl chloride (PVC) extrusion production line, a high-integrity ternary target observation dataset was obtained, containing measured values ​​of skin layer thickness ranging from approximately 0.25 mm to 0.35 mm, surface roughness arithmetic mean deviation values ​​ranging from approximately 0.8 μm to 1.2 μm, and PID deviation accumulation values ​​ranging from approximately 5 to 15 units of deviation integral. The time series sampling frequency of the synchronous process parameters was 1 Hz, with real-time screw speed fluctuating between 28 rpm and 35 rpm, melt pressure fluctuating between 12 MPa and 14 MPa, and cooling air velocity fluctuating between 4 m / s and 6 m / s. The three control variable columns and the ternary target column were concatenated according to time index to form a multidimensional matrix of 18,000 rows × 6 columns.

[0057] To verify the effectiveness of explicit control variable constraints, the correlation coefficients between each control variable and the crust thickness can be calculated. Substituting the time-corresponding values ​​of screw speed and thickness into the formula yields significantly correlated coefficients, thus demonstrating the constraint effect in the causal modeling stage. This matrix can be directly used as input in subsequent causal learning, significantly improving the accuracy and robustness of causal path identification.

[0058] S2.5: Perform normalization scaling on the preliminary constrained target observation matrix to map the measured value of the crust layer thickness, the arithmetic mean deviation value of the surface roughness, the cumulative deviation of the PID of the die temperature control system, and each control variable to a standard numerical space with zero mean and unit variance, so as to generate the final standardized target observation matrix for causal structure learning.

[0059] like Figure 2 As shown, step S3 involves performing causal structure learning on the observation matrix, eliminating spurious correlation paths, identifying the intensity and direction of the direct causal effect of mold temperature disturbance on the crust thickness, surface finish, and process stability, and generating a compensated causal map. Specifically, this includes: S3.1: Based on the control variable constraints in the target observation matrix, conditional independence tests are performed on the measured value of the crust layer thickness, the surface roughness meter scanning data, and the cumulative PID deviation of the die temperature control system to generate an initial undirected causal skeleton graph containing potential causal connections.

[0060] S3.2: Perform orientation operation on the initial undirected causal skeleton graph, and transform undirected edges into directed edges by identifying collision structures and non-collision structures to generate a preliminary directed acyclic graph.

[0061] Based on the node and edge relationship structure of the initial undirected causal skeleton graph, a complete set of nodes containing three types of target variables—skin layer thickness, surface finish, and process stability—as well as a die temperature disturbance variable is selected as the execution object for directional computation.

[0062] For each pair of adjacent nodes in the node set, a collision structure test is performed. The conditional independence criterion is used to determine the dependency relationship of their common neighbor nodes under the conditional constraints. Undirected edges that meet the definition of collision structure are marked as oriented edges and their inbound and outbound endpoints are recorded.

[0063] For the undirected edges marked as needing orientation, establish a direction determination rule: when a pair of variables exhibits causal dependence after removing the control variables and there exists a third node that forms a V structure with the pair of variables, the undirected edge is transformed into a directed edge pointing from the potential cause node to the potential result node.

[0064] For undirected edges that do not form a collision structure, perform non-collision structure processing. Based on Markov properties and prior control variable constraints, determine the causal flow direction of the edges and transform them into directed edges that do not form cycles to satisfy the constraints of the directed acyclic graph construction.

[0065] The algorithm performs loop detection on the oriented edge set. If a closed loop structure is detected, the algorithm calls the direction reversal and edge deletion strategy to break the loop and re-verify the topological validity to ensure the acyclicity and consistency of the entire graph with causal flow.

[0066] By using a directional operation processing method, the initial undirected causal skeleton diagram from the previous step is transformed into a preliminary directed acyclic graph that represents the causal flow between variables, thereby accurately depicting the direct interaction path direction between the mold temperature disturbance and the three types of compensation targets.

[0067] For example, in a causal structure learning task during a polyvinyl chloride (PVC) extrusion process, the node set is set to include die temperature perturbation (T), skin layer thickness (H), surface finish (S), and process stability (St). In the initial skeleton graph, there are undirected edges between H and S, both connected to T. During collision structure discrimination, it is found that T is a common neighbor of H and S, and under the conditional control variables of screw speed and melt pressure, H and S are conditionally independent, satisfying the V structure criterion. Therefore, the undirected edge H-S is oriented as H→S. For the undirected edge between T and St, through non-collision structure discrimination, it is identified that St is affected by T and is thus oriented as T→St. During loop detection, no closed loop is found, so the current direction is maintained. In the final generated preliminary directed acyclic graph, the edge set includes H→S, T→H, T→St, and S→St. Its direction information clearly depicts the causal path of the die temperature perturbation on each target, effectively improving the accuracy of subsequent causal strength coefficient calculation and the interpretability of topological association analysis.

[0068] S3.3: Based on the prior knowledge constraints of historical high-quality batch data, the redundant connections and false causal paths in the preliminary directed acyclic graph are pruned to output a simplified causal structure graph that retains only the direct path of the mold temperature disturbance to the three types of compensation targets.

[0069] Based on the preliminary topology of the directed acyclic graph and prior knowledge constraints from historical high-quality batch data, the target association rule set and the list of influence ranges of die temperature disturbances stored in the prior knowledge base are invoked to perform validity checks on each directed edge in the graph, filtering out redundant connections that do not conform to the actual process principles. For spurious causal paths caused by data noise or parameter collinearity, the control variable intervention rules in the prior constraints are used to perform inverse verification operations, identifying and marking invalid false correlations between path nodes through conditional probability difference calculations. For edges with multi-level intermediate nodes but not in the direct chain of die temperature disturbances, their influence coefficients are calculated based on the path length judgment threshold and the causal effect reduction model. When the coefficient is lower than the threshold, such edges are pruned to ensure the simplicity of the path chain. For the retained edges directly related to the die temperature disturbances, topology consistency checks are performed to ensure that the direction of the causal relationship is consistent with the actual process influence chain. If an incorrect direction is found, the edge is reversed to conform to physical laws. Through the above pruning and correction processes, the initial directed acyclic graph is transformed into a simplified causal structure graph that only contains the direct action paths of three types of compensation targets: mold temperature disturbance to skin layer thickness, surface finish, and process stability, thus achieving high accuracy and interpretability of the causal structure.

[0070] For example, in a steady-state batch data environment with a screw speed of 380 Nm, a melt pressure of 2.5 MPa, and a cooling fan speed of 2200 rpm, the initial directed acyclic graph contains 28 edges, of which 9 are related to the die temperature disturbance. The effective direct action path rule defined by the prior knowledge base requires a path length ≤ 2 and a direction pointing towards the target node, calculated using the formula for conditional probability difference: ,in For conditional probability differences, For the mold opening temperature disturbance variable, For the target variable, To control the variable set, paths with a calculation result less than 0.05 are considered spurious causal relationships and are pruned. After this determination, 5 redundant connection paths and 2 spurious causal paths are pruned. The remaining simplified causal structure diagram contains only one direct path from the die temperature disturbance to the crust layer thickness, surface finish, and process stability. Furthermore, the direction of each path is verified by topology consistency to conform to the physical laws of historical batches. The final simplified causal structure diagram accurately describes the direct effect of the die temperature disturbance, significantly improving the reliability of subsequent causal strength coefficient calculations.

[0071] S3.4: Based on the effective causal paths in the simplified causal structure diagram, the linear regression residual analysis method is used to calculate the standardized causal intensity coefficient of the mold temperature disturbance on each compensation target, so as to quantify the influence weight of each causal edge and generate a weighted set of causal association edges.

[0072] Based on the effective causal paths in the simplified causal structure graph, the set of direct causal connections between the mold temperature disturbance variable and the corresponding compensation target variable is locked, and a directed edge index table is established to facilitate subsequent strength calculation and processing.

[0073] For each valid causal edge, the mold temperature disturbance sequence in the historical stable batch data is used as the input of the independent variable, and the measured sequence of the compensation target is used as the input of the dependent variable. A multiple linear regression model containing control variable constraints is constructed, and the regression residual extraction operation is performed.

[0074] The variance and covariance indices are calculated for the regression residual sequence. Residual analysis is used to remove the random fluctuations that the control variables could not explain, so as to separate the net effect component of the die temperature disturbance on the compensation target.

[0075] The net effect components are mapped to a standardized numerical space, and the weight quantization is calculated using the following standardized causality strength coefficient formula: in, This is the sequence of net effect components of the die opening temperature disturbance. To compensate for the target net response sequence, This represents the covariance calculation operator. This represents the standard deviation calculation operator.

[0076] Perform the above on all valid causal edges. Coefficient calculation, and in If the absolute value is lower than a preset intensity threshold, an intensity rejection operation is performed to retain causal connections that have a significant impact on the compensation target.

[0077] The retained causal connections and their corresponding standardized causal strength coefficients are written into the weighted causal association edge set, and the edge set index and topology consistency check flag are updated to ensure the accuracy and interpretability of subsequent causal graph construction.

[0078] By using linear regression residual analysis and standardized intensity calculation, the simplified causal structure diagram from the previous step is transformed into a weighted causal association edge set that can be directly used for dynamic weight allocation, thereby achieving quantitative characterization and improved comparability of the die temperature disturbance effect.

[0079] For example, in the steady-state process data of PVC extrusion, the standard deviation of the net effect sequence of die temperature disturbance is 0.12, the standard deviation of the net response sequence of skin layer thickness is 0.08, and their covariance is 0.0076. Plugging the parameters into the formula to calculate the β coefficient, the result is β≈0.79, which reaches the preset strength threshold of 0.5 or higher. Therefore, this causal edge is considered a significant connection and added to the weighted causal association edge set. In the same batch of data, the β coefficients of the surface finish net response sequence and the die temperature disturbance net effect sequence are approximately 0.42, which is below the threshold, so the corresponding causal edges are removed. After repeated calculations and strength screening, the significant causal strength coefficient for the skin layer thickness target is 0.79, and the process stability target is 0.65. A weighted causal association edge set is generated, which can significantly improve the accuracy and priority allocation of control strategies when constructing the causal graph later.

[0080] S3.5: Integrate the weighted causal association edge set and the stage label sequence to construct a compensated causal graph with topological association characteristics, so as to form the compensated causal graph that supports online weight migration and dynamic switching.

[0081] Based on the weighted causal relationship edge set and stage label sequence set output by the preceding steps, the causal relationship storage interface is called to extract the weighted edge set of the corresponding stage by semantic label index, so as to realize the initial binding mapping between nodes and stage labels.

[0082] Structural feature scanning is performed on the target node set in the binding mapping to identify the co-occurrence relationships of nodes across stages in terms of skin layer thickness, surface finish, and process stability, and cross-stage reference pointers between nodes are established to form a node relationship chain with topological association characteristics.

[0083] Based on the relationship chain, a graph construction operation is performed on the causal edge set in each stage to generate a directed graph instance with the target node as the core and the edge weight as the standardized causal strength coefficient, and the cross-stage reference pointer is transformed into an inter-graph topological association index.

[0084] For each generated graph instance at each stage, the topology consistency verification module is called to verify the legality of the node's in-degree, out-degree, and causal path direction, eliminate illegal connections, and correct the directional parameters of the weight edges to ensure that the graph structure satisfies the directed acyclic condition.

[0085] The set of stage graph instances that pass the verification is written into the causal graph cache area and marked as four independent causal graphs that support online weight migration and dynamic switching, so that the subsequent dynamic weight allocation module can directly call them.

[0086] Through the above integration and construction process, the weighted edge set of the previous step is transformed into a multi-graph data structure with stage association, topological consistency and indexable calling, so as to realize online switching and migration support of multi-target die temperature compensation weights under different process stages.

[0087] For example, in a stable PVC extrusion production line, the stage semantic label set includes four categories: startup stage, steady-state stage, material changeover stage, and shutdown stage. The weighted causal association edge set records the standardized causal strength coefficients (0.62, 0.48, and 0.35) of the die temperature disturbance on three target nodes: skin thickness, surface finish, and process stability. For the steady-state stage, the system extracts the weighted edge set according to the label index and constructs a directed graph instance with the three target nodes as its core. The total in-degree weight of the skin thickness node is 0.62, the total in-degree weight of the surface finish node is 0.48, and the total in-degree weight of the process stability node is 0.35. Cross-stage reference pointers establish weighted associations between the skin thickness and surface finish nodes between the steady-state stage and the material changeover stage. The weighted edge directions undergo consistency checks to ensure the graph satisfies the directed acyclic condition. This steady-state stage causal graph is stored in a cache and forms a set with the other three stage graphs, possessing a topological association index structure. The set is input into the dynamic weight allocation module, and the new stage map is quickly loaded through the index during stage switching. The smooth migration of the in-degree normalized weight vector is achieved without rebuilding the map structure, which significantly improves the weight switching response speed and the stability of multi-objective compensation.

[0088] like Figure 3 As shown, step S4 involves: acquiring the real-time process parameter stream and inputting it into a time series classifier, outputting the current stage label and drift probability, and determining whether to trigger a stage switching signal based on the confidence level crossing a threshold. Specifically, this includes: S4.1: Perform sliding window truncation and standardization on the real-time second-level process parameter stream to eliminate dimensional differences and construct a fixed-length time-series input vector, generating a standardized multivariate time-series data matrix as the basic input object for subsequent feature extraction.

[0089] It should be noted that the time-series classifier can adopt architectures such as pure convolutional networks and recurrent neural networks. Taking the convolutional-recurrent neural network hybrid architecture as an example, the input is a five-dimensional real-time process parameter stream consisting of the die temperature setpoint, measured die temperature, main motor current, material pressure, and PID output value from the most recent 60 sampling periods. After sliding window truncation and standardization, it is sequentially input into two layers of time-series convolutional networks (the first layer has 3×5 kernels and outputs 32 channels, and the second layer has 5×1 kernels and outputs 64 channels) to extract local spatiotemporal correlation features. Then, it learns the time dependency relationship through a gated recurrent unit containing 64 hidden units. Finally, the fully connected classification head (64→32→4) outputs the probability distribution of the four stage labels of the current equipment startup stage, steady-state extrusion stage, material change transition stage, and shutdown cleaning stage through a soft maximum function. The maximum probability value is taken as the current stage label, and 1 minus the maximum probability value is taken as the drift probability. The classifier is trained offline using the stage label sequence generated by S1 as a supervision signal. After training convergence, it is fixedly deployed for online stage recognition.

[0090] S4.2: Based on the standardized multivariate time-series data matrix, perform hierarchical feature mapping operation of convolutional neural network, use local receptive field to capture the spatiotemporal correlation features between the sudden change point of the mold temperature setpoint and the step change of the main motor current, and generate a deep feature tensor containing high-order abstract information.

[0091] Based on the standardized multivariate time-series data matrix input, the convolutional layer of the convolutional neural network performs multi-channel sliding convolution operations. The convolutional kernel size is set to 3×3, and sampling is performed with a time axis step size of 1. The numerical gradient pattern of the die temperature setpoint near the abrupt change point and the amplitude response pattern of the main motor current at the step change point are captured through the local receptive field. Multiple sets of feature maps output from the convolutional layer are fed into a batch normalization layer, where normalization is adjusted according to the mean and variance of each feature channel to ensure the consistency of feature value distribution across different channels and reduce the interference of input data fluctuations on model recognition. A nonlinear activation function mapping, such as a hyperbolic tangent function mapping, is applied to the normalized feature maps. This curve transformation enhances the significant difference between the die temperature abrupt change and the motor step change, highlighting the spatiotemporal correlation features in a higher-order space. The activated feature maps are then fed into a secondary convolutional layer with a kernel size of 1×3 to extract fine feature changes along the time axis. This layer integrates feature patterns from different time scales within the receptive field to enhance sensitivity at process stage switching boundaries. The features of the second convolutional layer are superimposed to form a deep feature tensor. This tensor contains the joint spatiotemporal feature embedding of the abrupt change in the mold temperature setpoint and the step change in the main motor current at each time step. Through the feature hierarchy mapping processing method, the standardized multivariate time series data matrix of the previous step is transformed into a high-dimensional structured deep feature tensor, so as to realize the accurate extraction of the high-order abstract information required for the determination of the stage label sequence.

[0092] For example, a standardized multivariate temporal data matrix is ​​input into a convolutional neural network containing two convolutional layers. The first convolutional layer has 32 kernels, a size of 3×3, and a stride of 1; the second convolutional layer has 64 kernels, a size of 1×3, and a stride of 1. The batch normalization layer has a momentum coefficient of 0.9, and the hyperbolic tangent activation function is set to the range [-1, 1]. In actual execution, the gradient change at the abrupt change in the mold temperature setpoint significantly increases the feature response value in the first convolutional layer, while the amplitude response of the step change in the main motor current shows a peak in the feature response curve of the second convolutional layer. After normalization, the numerical fluctuation range of the convolutional output is stabilized between 0.05 and 0.95, ensuring the separability of the input data for the subsequent classification head. The final output dimension of the deep feature tensor is [time step × 64 channels]. In the time steps before and after the process stage switching, the response mode of the tensor channels undergoes significant heterogeneity, which provides a significant improvement in the stage judgment and discrimination performance for the nonlinear transformation and soft maximum probability distribution calculation of S4.3.

[0093] S4.3: Utilize the deep feature tensor input to the fully connected classification head to perform nonlinear transformation and soft maximum probability distribution calculation, so as to quantify the probability of belonging to the start-up segment, steady-state segment, material change transition segment or shutdown segment at the current moment, and generate the current stage label sequence with confidence value and the corresponding label drift probability.

[0094] Based on the deep feature tensor containing high-order abstract information generated in the previous sub-step, this deep feature tensor is used as the input object of the fully connected classification head, and matrix multiplication of the weight matrix and the feature tensor is performed to realize the mapping of high-dimensional features to low-dimensional classification probability space.

[0095] A bias vector is added to the result of the matrix multiplication operation to form a linear transformation output matrix, ensuring compensation for the prior bias of the model during the mapping process.

[0096] Nonlinear activation operations are performed on each element in the linear transformation output matrix, using modified linear units (ReLU) or hyperbolic tangent functions (Tanh) to enhance the classification head's ability to distinguish complex feature patterns.

[0097] The output vector after nonlinear activation is input into the softmax function to calculate the probability distribution normalization of the four types of process stage labels. The specific formula is as follows: in, For the first Linear transformation output value of the process stage This represents the probability of belonging to a particular stage.

[0098] The probability values ​​of the above four types of process stages are used as confidence scores, and the absolute difference of the change in the current tag probability within adjacent sampling periods is calculated to form a tag drift probability index.

[0099] Through the above processing method, the deep feature tensor of the previous step is transformed into a structured classification output containing the confidence and drift probability of the start-up segment, steady-state segment, material change transition segment and shutdown segment, so as to realize the quantitative identification of the label sequence of the current stage.

[0100] For example, in the operation scenario of a PVC extrusion production line, the deep feature tensor dimension is (1×128), the fully connected classifier head weight matrix is ​​set to (128×4), the bias vector length is 4, the nonlinear activation uses the ReLU function, and the output probability values ​​after Softmax normalization are 0.05 for the start-up segment, 0.82 for the steady-state segment, 0.10 for the material change transition segment, and 0.03 for the shutdown segment, with a confidence level of 0.82 for the corresponding steady-state segment. The label drift probability is calculated by comparing it with the confidence level of 0.65 for the steady-state segment of the previous cycle, with a difference of 0.17, indicating improved stability of the characterization stage. Applying this result to the subsequent stage switching threshold detection can significantly improve the accuracy of switching judgment and improve the smoothness of the weight transfer process.

[0101] S4.4: Based on the current stage label sequence and label drift probability, construct a time series stack of continuous sampling periods, execute the confidence threshold crossing detection algorithm within the sliding window, identify transient behavior where the built-in confidence crosses the confidence threshold for five consecutive sampling periods, and generate preliminary stage switching candidate event markers.

[0102] S4.5: Perform jitter verification and status determination based on the preliminary stage switching candidate event markers to confirm that a substantial migration of the process stage has occurred and eliminate sensor noise interference, and generate the stage switching signal.

[0103] The candidate event markers for the initial stage switching are verified by multi-source synchronous timestamps. The original sampling records of the die temperature, main motor current and material pressure sensors are called up, and the numerical change trends of each sensor in the sampling period before and after the candidate event are compared to screen out erroneous event markers with isolated fluctuations.

[0104] De-jitter verification is performed on the candidate events after review. A fixed-length time window is used to count the number of consecutive changes in the attribution confidence. A threshold filtering algorithm is used to remove false handover signals introduced by abnormal fluctuations in a single cycle.

[0105] For candidate events that have been validated for jitter removal, construct the transition matrix between the current state and the target state of the state machine. Combine the stage definition rules to determine the legality of the state transition. Events that do not meet the process stage transition path constraints are directly marked as canceled.

[0106] The rate of change of drift probability is calculated for legitimate migration events. The stable interval of the drift probability is determined using an exponentially weighted ratio calculation method. The drift rate is determined using the following formula: in, This indicates the confidence level of the current sampling period. This represents the rate of change of the drift probability.

[0107] The rate of change of the drift probability is compared with a preset stability threshold. If the rate of change is less than the threshold, the drift state is confirmed to be stable, and a final stage switching trigger signal is generated.

[0108] By performing de-jitter verification and state determination, the preliminary candidate events formed in the previous step are transformed into final stage switching trigger signals that have undergone noise filtering and legality verification, thereby achieving precise activation of the subsequent weight migration protocol.

[0109] For example, in a polyvinyl chloride (PVC) extrusion production line, the die temperature sampling frequency is 1Hz, the main motor current sampling frequency is 2Hz, and the material pressure sampling frequency is 1Hz. A fixed-length time window of 3 seconds is used to count the number of confidence level changes. The threshold is set to consider two consecutive crossovers as valid. The drift probability stability threshold is set to 0.05, and the candidate event P(t) = 0.82. 1) = 0.28. The drift probability change rate is calculated using the formula, and the result is 1.928, which is greater than the stability threshold. Therefore, it is determined that the drift probability needs to stabilize further. After the drift probability change rate drops to 0.01 for two consecutive cycles, a trigger signal is generated, and the weight transfer process begins. In this embodiment, the accurate generation of the trigger signal significantly improves the reliability of stage switching identification and avoids erroneous weight switching caused by isolated sensor disturbances.

[0110] Step S5: Based on the compensation causal graph corresponding to the current stage label activation, extract the in-degree normalized value of the node as the initial weight, and perform weight migration when a switching signal is detected. Specifically, this includes: S5.1: Based on the current time-stage label sequence and drift probability output by the previous steps, perform index matching on the four pre-stored compensation causal graphs with topological associations to lock and load the compensation causal graph instance that uniquely corresponds to the current time-stage label sequence.

[0111] It should be noted that the in-degree refers to the sum of the standardized causal strength coefficients of all directed edges pointing to a certain target node in the compensated causal graph; the normalization refers to dividing the in-degree by the sum of the in-degrees of all target nodes in the current graph to obtain the initial weight of the target.

[0112] Based on the current stage label sequence and drift probability output in the previous steps, the stage index parsing module is called to perform uniqueness verification on the semantic label, ensuring that the current label has no possibility of duplicate matching in the preset four-class process stage classification system, and obtaining a verified single label identifier.

[0113] Using tags as search keys, four pre-stored index tables of compensated causal graphs with topological associations are accessed. Hash mapping operations are performed to locate the physical address of the corresponding graph in the storage structure, thus achieving fast index matching.

[0114] The matched graph is subjected to version consistency verification. The stored version number is compared with the version number currently maintained by the running version control module. Cache instances with inconsistent versions are removed to ensure that the causal strength coefficient and node topology of the loaded graph are up-to-date.

[0115] The corresponding graph node and standardized causal strength coefficient set are read through the topology loading interface, stored in a temporary data buffer, and the current process stage label is attached as metadata to achieve a one-to-one correspondence between data and stage status.

[0116] During the loading process, the integrity of the causal path structure is checked to verify that the matching relationship between the number of nodes and the edge set conforms to the preset structural constraints, so as to eliminate structural errors caused by storage corruption or abnormal writes.

[0117] Through the above index matching and loading process, the stage label results identified in the previous step are transformed into corresponding compensation causal graph instances, thus providing a precise data foundation for subsequent multi-objective in-degree statistics and initial weight calculation.

[0118] For example, in the steady-state section of a PVC extrusion production line, the current stage tag sequence detection output is "steady-state section," with a drift probability of 0.12. The stage index parsing module performs a uniqueness check on this tag, confirming that the tag exists in a unique position within the four tag sets. Using the tag as a retrieval key, the index table is accessed, and the storage address 0x7A3F of the steady-state section graph is located through hash mapping. Version consistency verification compares the stored version number v3.2 with the running version number v3.2. The loading interface reads the node set of the steady-state section graph, which includes three types of target nodes and edge sets. The standardized causal strength coefficients are 0.65 for the skin layer thickness node, 0.55 for the surface finish node, and 0.48 for the process stability node. Integrity checks confirm that the number of nodes and edges is 3, conforming to topological constraints. A temporary buffer stores the graph and tag metadata, providing input data for subsequent in-degree statistical calculations. In subsequent S5.2, the in-degree can be effectively extracted and an initial weight allocation vector can be generated, realizing the weight benchmark for multi-objective collaborative optimization in the steady-state section.

[0119] S5.2: Perform in-degree statistical calculations on the three target nodes representing the crust layer thickness, surface finish, and process stability in the locked compensated causal graph instance to obtain the sum of the standardized causal intensity coefficients received by each target node as the original in-degree data set.

[0120] The input condition is a locked instance of the compensated causal graph, which includes three target nodes: skin layer thickness, surface finish, and process stability, as well as their causal edge weights associated with the die temperature disturbance.

[0121] For each target node in the graph instance, the standardized causal intensity coefficients of all incident edges are read, and the coefficient values ​​are retrieved one by one according to the edge source node to ensure complete coverage.

[0122] The retrieved set of incident edge coefficients is numerically accumulated using floating-point precision to avoid precision loss due to the superposition of multiple edges, thus forming the cumulative in-degree value of the node.

[0123] Repeat the above accumulation process for all target nodes to construct an original in-degree data set containing three nodes, where each element corresponds to the sum of causal strength coefficients received by the target node.

[0124] To ensure consistency in weight calculation, duplicate coefficients of parallel paths appearing during the accumulation process are removed to eliminate the cumulative error caused by repeated counting of multiple paths.

[0125] By using the above-mentioned in-degree statistical processing method, the graph nodes indexed and matched in the previous step are transformed into a set of raw in-degree data that can quantitatively represent the relative priorities of multiple objectives, thus realizing the basic data preparation for multi-objective weight allocation.

[0126] S5.3: Perform a normalization mapping operation based on the original in-degree data set to convert the original in-degree data of each target node into an initial weight allocation vector with values ​​ranging from zero to one, so as to establish the baseline priority distribution of multi-objective collaborative optimization under the current process stage.

[0127] S5.4: Generate a stage switching trigger signal based on the confidence level crossing the threshold within the continuous sampling period. When the stage switching trigger signal is determined to be valid, perform a freeze operation on the currently effective original spectrum weights and generate a freeze status flag to prevent unexpected fluctuations in the weights during the transition period.

[0128] S5.5: Activate the new graph weight enablement protocol based on the frozen state flag bit, load the initial weight allocation vector into the dynamic weight register, and complete the transfer of control from the original graph weight in the frozen state to the new graph weight, so as to generate the intermediate state data stream of weight migration to be smoothed.

[0129] The frozen state flag is input as a trigger condition to the control and permission management module to perform state variable verification and protocol consistency check operations, ensuring that the weight migration protocol complies with the legality constraints of process stage switching within the current time slice.

[0130] The new graph weight loading subroutine is called to extract the pre-calculated initial weight allocation vector from the weight buffer and write it into the dynamic weight register, thus completing the formatted mapping of the data structure to maintain the consistency between the weight components and the index of the target node.

[0131] During the writing process of the dynamic weight register, a standardized verification operation is performed on the in-degree normalized value. The numerical offset of each weight component is detected and corrected in real time using a fixed tolerance range to avoid distortion in subsequent smoothing processing due to floating-point calculation errors.

[0132] A bidirectional control interlock logic is established between the original and new map weight registers. The write permission of the original register is closed by the control handover signal, while the write and read permissions of the new register are activated, thereby realizing the physical and logical switching of the weight source.

[0133] Generate an intermediate weight migration data stream containing the original weight freeze value, the new weight initial value, and the target node index. Encapsulate this data stream into a time-series weight switching record in a unified format to provide a continuous data source for subsequent smooth transition calculations.

[0134] By coordinating the frozen state flag and the new weight activation protocol, the index matching result of the previous step is transformed into intermediate weight migration data that can be directly used in smooth calculation, thus achieving seamless transition and controllability of the weight switching process.

[0135] For example, when switching from the steady-state section to the material change transition section of a PVC extrusion production line, the system detects a stage switching trigger signal and generates a freeze status flag. At this time, the initial weight allocation vector of the new spectrum is [0.42, 0.33, 0.25], corresponding to the three targets: skin layer thickness, surface smoothness, and process stability, respectively. This vector is loaded into a dynamic weight register, configured with a word width of 64-bit floating point, and the index table is arranged in order of target node ID. During the loading process, a standardization check is performed using a tolerance range of ±0.0001, and the detected offset values ​​are corrected in real time to ensure that the output weight components are accurate. The timing delay of the control authority handover logic is set to 2 sampling periods, and the control authority is transferred by disabling the original register write permission and enabling the new register permission. The final generated intermediate weight transition data stream contains timestamp 1678453200, node index [1,2,3], frozen weights [0.50,0.30,0.20], and initial new weight values ​​[0.42,0.33,0.25]. This data serves as input for the S6 exponential smoothing calculation, verifying that the smoothness and accuracy of weight switching in the next 30-second transition window are significantly improved.

[0136] Step S6: Perform exponential smoothing calculation on the initial weights within the transition window to complete the smooth weight switching and generate a weight allocation vector. Specifically, this includes: S6.1: Obtain the weight migration protocol trigger signal for freezing the original spectrum weights and enabling the new spectrum weights, extract the standardized causal intensity coefficient vector of the original spectrum and the standardized causal intensity coefficient vector of the new spectrum at the current time, and calculate the discrete time step sequence based on the 30-second transition window duration to generate a weight transition control benchmark containing time index.

[0137] S6.2: Based on the discrete time step sequence in the weight transition control benchmark, perform element-wise exponential decay operation on the original spectrum normalized causal intensity coefficient vector using a preset decay factor to generate the original weight decay component sequence that decreases with time.

[0138] Based on the discrete time step sequence in the weighted transition control benchmark, the normalized causal intensity coefficient vector of the original spectrum in the frozen state is used as the input object for the decay operation.

[0139] For each time step index, a preset decay factor parameter is invoked to construct the baseline formula for exponential decay calculation: in, These are the initial values ​​of the original weights at the time of freezing. As the attenuation factor, This is the index for the discrete time step.

[0140] Apply the formula to each element of the vector and perform an element-wise exponential decay calculation to ensure that each element decreases smoothly as the time step increases.

[0141] During the decreasing process, the product of the decay factor and the time step is processed by floating-point precision to avoid the curve shape deviating from the expected shape due to accumulated errors in long sequence calculations.

[0142] Arrange all the decayed element values ​​at all time steps in a time sequence to form the original weight decay component sequence, ensuring that the sequence structure corresponds one-to-one with the time step index and has a monotonically decreasing characteristic.

[0143] Through the above decay operation, the original weight vector from the previous step is transformed into weight component data that decreases over time, thus achieving a smooth fading effect of the original weight contribution during the weight migration process.

[0144] S6.3: Based on the discrete time step sequence in the weight transition control benchmark, perform element-wise exponential growth operation on the new spectrum standardized causal intensity coefficient vector using complementary growth factors to generate a new weight growth component sequence that increases over time.

[0145] S6.4: Perform periodic vector superposition processing on the original weight decay component sequence and the new weight growth component sequence to eliminate the influence of a single weight source mutation and generate an intermediate transition weight allocation vector sequence.

[0146] A periodic vector superposition operation is performed on the original weight decay component sequence (decreasing over time) and the new weight growth component sequence (increasing over time). The operation targets two weight component sequences indexed by a 30-second transition window discrete time step sequence. For the same time step index, a parallel vector operation unit is invoked to perform arithmetic summation of the corresponding elements' decay and growth components in the numerical space, generating a set of intermediate weight allocation vector elements containing the superposition result. To avoid priority imbalance caused by abrupt changes in a single weight source, the time synchronization of the superposition operation is ensured by relying on a consistent step sequence, resulting in a smooth transition between the decreasing trend of the decay component and the increasing trend of the growth component within each period. A structured formula is used to describe the superposition operation process: in This represents the intermediate transition weight allocation vector. This represents the original weight decay component vector. This represents the new weight growth component vector. This serves as the discrete-time step index. Within each cycle, the above superposition results are written to the corresponding positions in the intermediate transition weight allocation vector sequence, forming a complete sequence structure for subsequent normalization and verification processing. By using a cycle-by-cycle vector superposition process, the decay and growth components of the previous step are mapped to smooth intermediate weight allocation data, achieving a seamless and coordinated transition effect between the source weights and the target weights.

[0147] For example, during the transition from the material changeover stage to the steady-state stage in PVC extrusion production, the 30-second transition window is divided into 300 sampling periods, each period lasting 0.1 seconds. The initial value of the original weight decay component sequence is [0.45, 0.35, 0.20], with a decay factor set to 0.98; the initial value of the new weight growth component sequence is [0.30, 0.40, 0.30], with a growth factor set to 0.02. In the 15th period, the original weight decay component vector is [0.36, 0.28, 0.16], and the new weight growth component vector is [0.36, 0.43, 0.33]. The superposition formula is called, resulting in [0.72, 0.71, 0.49], which is then written into the 15th period position of the intermediate transition weight allocation vector sequence. After all three hundred cycles of superposition processing, the change curve of the intermediate transition weight allocation vector on each target category is continuous and smooth, which significantly improves the stability of the weight transfer process. During the verification process, the control error fluctuation of each target is significantly reduced, and the smooth switching of weight priority between the material change transition section and the steady state section is realized.

[0148] S6.5: Perform normalization verification processing based on the intermediate transition weight allocation vector sequence to correct the cumulative deviation caused by floating-point operations, so as to output the final weight allocation vector without sudden change jitter for multi-objective collaborative optimization.

[0149] Step S7: Based on the weight allocation vector, monitor the target achievement status. If the target fails to reach the preset threshold and its causality strength coefficient is higher than the limit value, temporarily increase the target weight and trigger the target conflict buffer mechanism. Specifically, this includes: S7.1: Obtain the weight allocation vector of the current sampling period and the measured feedback data of the three types of targets, and perform deviation threshold comparison processing on the measured feedback data of the three types of targets to generate a target achievement judgment sequence that includes the skin layer thickness deviation state, surface smoothness deviation state and process stability deviation state.

[0150] It should be noted that when the thickness of the crust layer fails to meet the standard for several consecutive cycles and its causal strength coefficient is higher than the limit value, the system temporarily increases the weight of the target while proportionally reducing the weight of other targets, and automatically restores the original allocation after a certain period. This is the target conflict buffering mechanism.

[0151] S7.2: Based on the non-compliance target identifiers within ten consecutive sampling periods in the target achievement determination sequence, the standardized causal intensity coefficients of the corresponding nodes are extracted by calling the pre-stored compensation causal graph, and the standardized causal intensity coefficients are subjected to limit value logic verification processing to generate a candidate set of targets to be intervened with high causal correlation characteristics.

[0152] Based on the identifiers of non-compliant targets appearing in multiple consecutive sampling periods (e.g., ten) of the target achievement determination sequence, a pre-stored compensation causal graph instance is invoked to perform a numbered index retrieval for the non-compliant target nodes to locate the corresponding causal path information. After node location is completed, the standardized causal intensity coefficient vector associated with each node is extracted, and a temporary mapping table is established with the target type as the index key and the causal intensity coefficient as the value. A limit value logic verification operation is performed on the standardized causal intensity coefficients, using a mathematical inequality comparison method to determine whether each coefficient is greater than a preset limit value threshold. The threshold is set based on process stability requirements and historical deviation tolerance. In the limit value logic verification, when the judgment condition is met, the corresponding target type is added to the candidate set of targets to be intervened; when the condition is not met, the target type is not included in the candidate set. Through this step-by-step verification and screening method, the output candidate set of targets to be intervened contains only target types with high causal correlation strength that may lead to a decrease in control performance under the current process state. Logical rigor ensures the targetedness and effectiveness of the intervention behavior. This process transforms the deviation determination results from the previous step into an intervention target dataset with high causal correlation characteristics, enabling precise target targeting during subsequent dynamic weight reconstruction.

[0153] S7.3: Based on the target type and its corresponding standardized causal intensity coefficient in the candidate set of targets to be intervened, perform dynamic weight reconstruction calculation based on causal gain factor to generate temporary weight increment parameters for correcting the original weight allocation vector.

[0154] S7.4: The weight components of the corresponding target in the original weight allocation vector are superimposed and updated using the temporary weight increment parameter to generate a corrected weight allocation vector that suppresses target conflict fluctuations and simultaneously activates the target conflict buffer mechanism.

[0155] Based on the temporary weight increment parameters corresponding to the candidate set of targets to be intervened, the original weight allocation vector of the current sampling period in the dynamic weight register is called as the basic data object for the superposition operation.

[0156] The weight components corresponding to the target to be intervened in the original weight allocation vector are added element by element, and the precise floating-point operation mode is used to avoid the accumulation of rounding errors.

[0157] Normalization verification is performed on the superimposed weight allocation vector, and causal consistency verification is performed on the normalized modified weight allocation vector to ensure that the modified weights are consistent with the causal intensity coefficients in the pre-stored target-temperature compensated causal spectrum.

[0158] The verified corrected weight allocation vector is written to the dynamic weight register, and the target conflict buffer mechanism flag is activated in the control logic, so that the corrected weight allocation vector can be applied in real time in subsequent PID regulation.

[0159] By using the processing method of superimposed update and normalization verification, the temporary weight increment parameter of the previous step is transformed into a corrected weight allocation vector that suppresses target conflict fluctuations, thereby realizing the adaptive priority adjustment of the compensation strategy in the case of multi-target mismatch.

[0160] Based on the generated events of the corrected weight allocation vector, the corresponding time index information and causal path node parameters of the candidate set of targets to be intervened are extracted to establish an event triggering baseline data structure containing timestamps.

[0161] For each target type in the candidate set of intervention targets, perform causal path traversal, call the causal graph storage module to retrieve the direct causal edges associated with the target node and their standardized causal strength coefficients, and pair and map the node type with the causal coefficients one by one to form a set of target parameters with associated causal paths.

[0162] The target type, timestamp, and corresponding causal strength coefficient are written into the structured data encapsulation buffer in field order. Field label verification is used to ensure the integrity and consistency of each data field, and a basic data package body that can be called by the interface is generated in the buffer.

[0163] The basic data packet body is encoded, and millisecond-level precise identification of the timestamp field is introduced. Unit and dimension information is added to the causal strength coefficient field to ensure the consistency of numerical interpretation when the interface is highlighted.

[0164] The highlighting and rendering module of the human-computer interaction interface is invoked to bind the above-mentioned encoded structured warning data packet to the highlighting event identifier and output to the display queue at the front end of the interface to realize the real-time visualization of the warning signal.

[0165] By encapsulating structured data and binding it with visualization, the corrected weight allocation vector results from the previous step are transformed into early warning data containing time markers, target categories, and causal strength, enabling rapid identification of mismatched targets and providing decision support for engineers.

[0166] For example, during the transition from the steady-state stage to the material replacement stage of polyvinyl chloride extrusion, the modified weight allocation vector generates an event identifier at the 120-second mark. The candidate set of targets to be intervened includes two targets: skin layer thickness and surface smoothness, with standardized causality strength coefficients of 0.82 and 0.76, respectively. The timestamp is set to 2024-06-01 14:35:20.125ms, and the target type fields are labeled as Thickness and Gloss, respectively. The causality coefficient field is appended with the unit "causality strength coefficient". The structured early warning data packet is established as {timestamp: 2024-06-01 14:35:20.125ms, target type: Thickness, causality coefficient: 0.82; timestamp: 2024-06-01 14:35:20.125ms, target type: Gloss, causality coefficient: 0.76}. The highlight rendering module was invoked in the human-computer interaction interface, and the prompt boxes corresponding to the two targets were marked in red. The front-end display queue was updated in real time. Engineers observed a risk of mismatch between the thickness and surface finish indicators and intervened within 3 seconds. After applying this processing method, the intervention response speed of the compensation strategy in this transition section was significantly improved, avoiding the delay in mold temperature compensation caused by target mismatch.

[0167] Step S8: Generate a switching record based on the stage switching signal, generate a migration log based on the weight migration process, and store the weight allocation vector, switching record, and migration log together to form a weight allocation execution file. Specifically, this includes: S8.1: Obtain the weight allocation vector, stage switching trigger signal and weight migration intermediate state data stream of the current sampling period, and perform structured encapsulation processing on the multi-source heterogeneous control data to generate a standard evidence data package containing timestamp index, process stage identifier and weight value matrix.

[0168] Perform multi-source heterogeneous data synchronization calls on the weight allocation vector, stage switching trigger signal, and weight migration intermediate state data stream of the current sampling period to ensure that the timestamp accuracy of each data object is unified to the millisecond level.

[0169] The weight allocation vector is mapped to the stage switching trigger signal to establish a two-way binding relationship between the weight vector index and the stage identifier index, so as to achieve fast retrieval in the evidence storage data.

[0170] The intermediate data stream of weight transition is parsed and structured modeled, and the weight change component sequence is divided into time series matrix form and stored in association with the aforementioned binding index.

[0171] The synchronized multi-source heterogeneous control data is sorted and grouped by field, and arranged in a fixed order according to timestamp index, process stage identifier and weight value matrix to form a standard data structure template with stable field positions.

[0172] The above standard data structure template is serialized and encoded to generate a standard evidence storage data packet that meets the requirements of evidence storage data payload, and the original timestamp index is retained to support subsequent time sequence verification.

[0173] By using a structured encapsulation process, the results of the previous step are transformed into a data format with a fixed field order, a unified time base, and direct storage capability, thus achieving seamless integration with the evidence storage module.

[0174] For example, the weight allocation vector collected on the PVC extrusion production line is [0.42, 0.35, 0.23], the stage switching trigger signal value is 1, and the intermediate data stream of weight migration contains decreasing and increasing components per second within a 30-second transition window. When the three types of data are called synchronously, all timestamps are uniformly calibrated to 14:32:05 on May 15, 2024 as the reference point. The weight vector index uses integer sequence mapping, the stage identifier index value is "steady-state segment", and the changing component sequence is parsed to form a 30-row, 3-column matrix. A standard data structure template is established in the order of [timestamp index | process stage identifier | weight value matrix], and JSON serialization is used to generate a data package for evidence storage. In this embodiment, the field positions of the encapsulated data package are stable, the timestamp accuracy reaches the millisecond level, and it is successfully identified and confirmed by the evidence storage system. The generated evidence storage data can be directly archived in the evidence storage area. During subsequent audit backtracking, the corresponding weight adjustment record can be quickly located according to the time index, which significantly improves the efficiency of data retrieval and verification.

[0175] S8.2: Extract key feature fields based on the standard evidence storage data package, calculate data fingerprint information using a hash digest algorithm, generate a unique digital digest value for verifying data integrity and use it as the input payload of the evidence storage data.

[0176] S8.3: Construct a hash chain node structure based on the unique digital digest value, and perform an integrity verification operation in combination with the previous evidence hash value to generate a legitimate new block header information confirmed by the network nodes and establish the temporal logical relationship of data evidence storage.

[0177] S8.4: Using the legal new block header information, write the standard evidence storage data packet into the evidence storage area and execute the automatic archiving logic to generate the original record of the mold temperature dynamic weight allocation execution archive with tamper-proof characteristics.

[0178] S8.5: Based on the original records of the execution archive with dynamic weight allocation of the die temperature, establish a causal path retrieval index and associate it with the prior knowledge constraints of historical high-quality batch data to generate a structured execution archive library that supports audit backtracking analysis and process knowledge accumulation.

[0179] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0180] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0181] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A die temperature feedforward compensation method based on the prediction of PVC skin layer thickness, specifically including: S1: Obtain historical process parameters of the polyvinyl chloride extrusion process, align and segment them, and generate a stage tag sequence. S2: Based on the stage label sequence, collect the crust thickness, surface roughness and PID deviation data within the corresponding time period to construct the target observation matrix. S3: Perform causal structure learning on the observation matrix, remove spurious correlation paths, identify the intensity and direction of the direct causal effect of the mold temperature disturbance on the skin layer thickness, surface finish and process stability, and generate a compensated causal spectrum. S4: Obtain the real-time process parameter stream and input it into the timing classifier, output the current stage label and drift probability, and determine whether to trigger the stage switching signal based on the confidence level crossing the threshold. S5: Activate the corresponding compensation causal graph based on the current stage label, extract the in-degree normalized value of the node as the initial weight, and perform weight migration when a switching signal is detected. S6: Perform exponential smoothing calculation on the initial weights within the transition window to complete the smooth weight switching and generate a weight allocation vector. S7: Based on the weight allocation vector, monitor the target achievement status. If the target does not reach the preset threshold and its causal strength coefficient is higher than the limit value, temporarily increase the target weight and trigger the target conflict buffer mechanism. S8: Generate a switching record based on the stage switching signal, generate a migration log based on the weight migration process, and store the weight allocation vector, switching record and migration log together to form a weight allocation execution file.

2. The die temperature feedforward compensation method based on polyvinyl chloride skin layer thickness prediction according to claim 1, characterized in that, Step S3 specifically includes: Based on the control variable constraints in the target observation matrix, conditional independence tests are performed on the measured value of the crust layer thickness, the surface roughness meter scanning data, and the cumulative PID deviation of the die temperature control system to generate an initial undirected causal skeleton graph containing potential causal connections. An orientation operation is performed on the initial undirected causal skeleton graph. By identifying collision structures and non-collision structures, the undirected edges are transformed into directed edges, generating a preliminary directed acyclic graph. Based on prior knowledge constraints from historical high-quality batch data, redundant connections and false causal paths in the preliminary directed acyclic graph are pruned to output a simplified causal structure graph that retains only the direct path from the mold temperature disturbance to the three types of compensation targets. Based on the effective causal paths in the simplified causal structure graph, the linear regression residual analysis method is used to calculate the standardized causal intensity coefficient of the mold temperature disturbance on each compensation target, so as to quantify the influence weight of each causal edge and generate a weighted set of causal association edges. By integrating the weighted causal edge set and the stage label sequence, a compensated causal graph with topological correlation characteristics is constructed to form the compensated causal graph that supports online weight migration and dynamic switching.

3. The die temperature feedforward compensation method based on polyvinyl chloride skin layer thickness prediction according to claim 1, characterized in that, Step S4 specifically includes: The real-time second-level process parameter stream is subjected to sliding window truncation and standardization normalization to eliminate dimensional differences and construct a fixed-length time-series input vector, generating a standardized multivariate time-series data matrix as the basic input object for subsequent feature extraction. Based on the standardized multivariate time-series data matrix, perform hierarchical feature mapping operation of convolutional neural network, use local receptive field to capture the spatiotemporal correlation features between the sudden change point of the mold temperature setpoint and the step change of the main motor current, and generate a deep feature tensor containing high-order abstract information. The deep feature tensor is used to input a fully connected classification head to perform nonlinear transformation and soft maximum probability distribution calculation, so as to quantify the probability of belonging to the start-up segment, steady-state segment, material change transition segment or shutdown segment at the current moment, and generate a current stage label sequence with confidence value and corresponding label drift probability. Based on the current stage label sequence and label drift probability, a time series stack of continuous sampling periods is constructed. A confidence threshold crossing detection algorithm within a sliding window is executed to identify transient behavior where the built-in confidence crosses from below the confidence crossing threshold to above the confidence crossing threshold in five consecutive sampling periods, generating preliminary stage switching candidate event markers. Based on the preliminary stage switching candidate event markers, perform jitter verification and status determination to confirm that a substantial migration of the process stage has occurred and eliminate sensor noise interference, thereby generating the stage switching signal.

4. The die temperature feedforward compensation method based on polyvinyl chloride skin layer thickness prediction according to claim 1, characterized in that: In the alignment and segmentation process, sudden changes in die temperature, step changes in main motor current, and material pressure plateaus are detected; before the alignment and segmentation process, the historical process parameters are sequentially subjected to bandpass filtering, outlier removal, missing value completion, and standardization.

5. The die temperature feedforward compensation method based on polyvinyl chloride skin layer thickness prediction according to claim 1, characterized in that: The temporal classifier adopts a hybrid structure of convolutional-recurrent neural networks and outputs the probability of the current time belonging to each stage label. When the maximum belonging probability exceeds the preset confidence threshold, it is determined to be a valid classification; otherwise, a stage switching pending confirmation state is triggered.

6. The die temperature feedforward compensation method based on polyvinyl chloride skin layer thickness prediction according to claim 1, characterized in that: When pruning and identifying false paths, the causal structure learning uses a conditional probability difference less than a preset threshold as the basis for invalidation. The causal strength coefficient is obtained by regression residual analysis to obtain a standardized coefficient. Only when the absolute value of this coefficient exceeds the strong correlation threshold is it judged as a strongly correlated path, which is then used for weight allocation.

7. The die temperature feedforward compensation method based on polyvinyl chloride skin layer thickness prediction according to claim 1, characterized in that: The transition window has an adjustable length, and is combined with an exponential smoothing decay factor and a limit on the step size of the new weight growth to avoid actuator shocks caused by weight jumps.

8. The die temperature feedforward compensation method based on the prediction of polyvinyl chloride skin layer thickness according to claim 1, characterized in that: The monitoring of the achievement of the target is performed statistically according to a fixed number of sampling periods. If a target fails to meet the target for several consecutive periods and its causal intensity coefficient exceeds the warning threshold, the weight of the target is temporarily increased and all weights are normalized. At the same time, a structured warning data package containing timestamp, target type and causal intensity value is generated for human-computer interaction prompts.

9. The die temperature feedforward compensation method based on polyvinyl chloride skin layer thickness prediction according to claim 1, characterized in that: The weight allocation vector is output to the temperature compensation priority controller in real time as the basis for allocating multi-target die temperature setpoints; the controller calculates the comprehensive deviation in a weighted sum form and outputs the die temperature feedforward compensation amount.

10. The die temperature feedforward compensation method based on polyvinyl chloride skin layer thickness prediction according to claim 1, characterized in that: The evidence storage encapsulates each stage switch record, weight allocation vector, and weight migration log into a structured data format and saves it by calculating a hash digest; at the same time, it establishes a causal path retrieval index with timestamps and stage tags as index keys to support traceable auditing of the entire process of weight migration and stage switch operations.

Images