Automatic control method and system for plastic processing production line

Abstract

The invention relates to the technical field of production line control, and discloses an automatic control method and system for a plastic processing production line. The method comprises the steps of collecting technological parameters of a plastic processing production line and transmitting the technological parameters to a central control system to generate a database; multi-dimensional parameter correlation analysis is executed, a parameter and quality mapping relation is established through a CNN-LSTM hybrid network, and an optimization model is formed; calculating an optimal control parameter, generating a control strategy and issuing the control strategy to an execution unit; and monitoring a response result, updating the model in real time, and forming closed-loop adaptive control. According to the method, the mapping relation between the process parameters and the product quality is accurately established through the deep learning model, the optimal control parameters are automatically calculated, and closed-loop adaptive control based on production feedback is realized.

Description

Technical Field

[0001] This application relates to the technical field of production line control, and particularly to an automatic control method and system for a plastic processing production line. Background Art

[0002] In the modern plastic processing industry, the injection molding process is a key technology for producing plastic products. Traditional control of plastic processing production lines mainly relies on manual experience to set process parameters and realizes basic automation control through PLCs and industrial control computers. Some advanced plastic processing enterprises have started to apply industrial Internet of Things technology to network monitor the equipment on the production line, collect process parameters such as temperature, pressure, and speed, and use a database system to record production data. Some enterprises also try to use single control algorithms, such as PID control or simple fuzzy control, to independently control equipment such as injection molding machines, dryers, and mold temperature controllers. In terms of quality control, product quality evaluation is mainly carried out through off-line sampling inspection, and then process engineers manually adjust process parameters according to the quality inspection results in order to improve product quality.

[0003] There is a lack of in-depth coordinated control among the devices of traditional plastic processing production lines. The control of each link such as raw material management, drying treatment, and injection molding is fragmented, and it is impossible to achieve optimized control based on the full-process data. The existing control methods mainly rely on manual experience and simple algorithms, and cannot effectively handle the complex non-linear relationship between process parameters and product quality. Especially in the face of multi-variety and small-batch production scenarios, the parameter optimization cycle is long and the debugging cost is high. The traditional control system lacks the ability of adaptive learning and cannot automatically optimize the control strategy according to the data accumulated during the production process, resulting in poor product quality stability in the case of raw material batch changes, environmental condition fluctuations, etc.

[0004] With the exploration of the application of artificial intelligence technology in the industrial field, some research institutions and enterprises have begun to try to apply neural networks to plastic processing process control. These attempts mainly focus on using a single type of neural network model, such as BP neural network or simple convolutional neural network, to predict or optimize certain specific parameters. However, this type of single model structure faces the technical problem of difficulty in simultaneously processing the spatial characteristics and temporal characteristics of process parameters. In the plastic processing process, parameters such as raw material temperature and pressure not only have spatial distribution correlation, but also have dynamic characteristics that change over time. A single model cannot simultaneously process the spatial characteristics and temporal characteristics of process parameters, and capture the complex correlation patterns between raw material characteristics, process parameters and product quality, resulting in low accuracy of single model analysis. In addition, when applying deep learning models to process industrial production data, problems such as uneven data distribution, large differences in feature scales, and unstable training are often faced. The neural network models in the prior art often lack optimization processing mechanisms specifically for the characteristics of industrial data. For example, the application of technologies such as feature standardization and batch normalization in industrial models is insufficient, and advanced structures such as residual connections that can enhance the stability of deep network training are rarely considered. This makes the model prone to gradient vanishing and overfitting during training, making it difficult to fully leverage the advantages of deep learning in complex pattern recognition and to build a stable and efficient parameter-quality mapping model. Parameter settings in existing control systems often use fixed values, lack an intelligent adjustment mechanism based on real-time production conditions, and are unable to cope with dynamic changes in the production process. Traditional control systems have failed to fully leverage the advantages of artificial intelligence technologies such as deep learning, and are unable to mine the complex patterns hidden in production data, resulting in insufficient understanding of the relationship between process parameters and product quality, and limited parameter optimization effects. Summary of the invention

[0005] The present application provides an automatic control method and system for a plastic processing production line, which is used to integrate the whole process production data, accurately establish the mapping relationship between process parameters and product quality through a deep learning model, automatically calculate the optimal control parameters, and realize closed-loop adaptive control based on production feedback. At the same time, according to the deep learning model structure that can simultaneously process the spatial characteristics and temporal characteristics of process parameters, it can effectively capture the complex correlation patterns between raw material characteristics, process parameters and product quality, and improve the accuracy of the parameter-quality mapping model.

[0006] In a first aspect, the present application provides an automatic control method for a plastic processing production line. The automatic control method for the plastic processing production line includes: collecting process parameters and equipment status data of each station of the plastic processing production line, transmitting the collected data to a central control system through a distributed sensing network to generate a production status database; based on the production status database, performing multi-dimensional parameter correlation analysis, establishing a mapping relationship between process parameters and product quality through a deep learning model that combines a convolutional neural network and a long short-term memory network. The deep learning model includes three convolutional layers for feature extraction, two LSTM layers for time series modeling, and two fully connected layers for parameter mapping to form a parameter optimization model; according to the parameter optimization model, calculating the optimal control parameters for each station, generating a control strategy including an execution instruction sequence, and sending the control strategy to each execution unit; monitoring the response results and control effects of the execution unit, and performing real-time update on the parameter optimization model to form a closed-loop adaptive control system.

[0007] In the first implementation manner of the first aspect, the collecting process parameters and equipment status data of each station of the plastic processing production line, transmitting the collected data to a central control system through a distributed sensing network to generate a production status database includes: performing RFID tag identification on plastic raw materials, collecting raw material types, batch numbers, and supplier information, and performing preliminary processing on the collected information through an edge computing node to obtain raw material basic data; setting a multi-point temperature and humidity sensor and a weighing sensor array inside the raw material bin to monitor the raw material storage environment parameters and inventory in real time to obtain raw material status data; monitoring the operating status of the injection molding machine, collecting parameters such as melting temperature, injection pressure, holding pressure time, mold temperature, cooling time, injection speed, clamping force, screw rotation speed, and back pressure value, and transmitting them to a data acquisition unit through an industrial bus to obtain injection molding process data; collecting the energy consumption data of production equipment, measuring current, voltage, power factor, active power, reactive power, and harmonic content through power parameter sensors installed on each equipment to obtain energy consumption data; aligning the raw material basic data, raw material status data, injection molding process data, and energy consumption data through time stamps, establishing a multi-time scale data synchronization matrix to obtain a correlation data set; performing data cleaning and outlier detection on the correlation data set, removing noise and outliers through wavelet transform and statistical filtering algorithms, and storing the processed data in a structured database to form the production status database.

[0008] In the second implementation of the first aspect, based on the production status database, perform multi-dimensional parameter correlation analysis, and establish a mapping relationship between process parameters and product quality through a deep learning model that combines a convolutional neural network and a long short-term memory network. The deep learning model includes three convolutional layers for feature extraction, two LSTM layers for time series modeling, and two fully connected layers for parameter mapping, forming a parameter optimization model, including: extracting historical production data from the production status database, classifying it according to product models and raw material batches, reducing the data dimension through the principal component analysis algorithm to obtain a feature data set; performing a correlation analysis on the feature data set, calculating the correlation strength between parameters through the Pearson correlation coefficient and mutual information calculation methods, screening out the key process parameters that affect product quality, and forming a key parameter set; constructing a convolutional neural network module based on the key parameter set, setting three convolutional layers, each layer using 64, 128, and 256 3×3 convolutional kernels respectively to extract features from the process parameter sequence, introducing non-linear transformation through the ReLU activation function to obtain a feature mapping matrix; inputting the feature mapping matrix into the long short-term memory network module, passing through two LSTM layers each containing 128 neurons to model the parameter time series relationship, and combining the forget gate, input gate, and output gate mechanisms to handle long-term dependencies to obtain a time series feature vector; processing the time series feature vector through two fully connected layers, the first layer contains 256 neurons, the second layer contains 128 neurons, using the Dropout technique to prevent overfitting, and improving the training stability through batch normalization to obtain an optimized deep learning model; performing a sensitivity analysis on the optimized deep learning model, calculating the output change rate by perturbing the input parameters, constructing a parameter importance ranking matrix, and combining the expert rule system to form the parameter optimization model.

[0009] In the third implementation of the first aspect, the temporal feature vector is processed through two fully-connected layers. The first layer contains 256 neurons, and the second layer contains 128 neurons. The Dropout technique is used to prevent overfitting, and batch normalization is applied to improve training stability, resulting in an optimized deep learning model, including: performing data standardization processing on the temporal feature vector, calculating the mean and standard deviation of each dimension feature through the Z-score standardization method, converting the feature data into a standard normal distribution with a mean of 0 and a standard deviation of 1 to obtain a standardized feature vector; inputting the standardized feature vector into the first fully-connected layer, performing a linear transformation on the input features through 256 neurons, calculating the weighted sum by applying a weight matrix and the input vector to obtain the original output value of the first layer; performing batch normalization processing on the original output value of the first layer, adjusting the feature distribution by calculating the mean and variance of the samples within the mini-batch, and introducing learnable scaling parameter γ and translation parameter β to obtain a normalized output value; applying the LeakyReLU activation function to the normalized output value, keeping the positive values unchanged, multiplying the negative values by a slope coefficient of 0.01, introducing a non-linear transformation, and applying Dropout processing with a rate of 0.5 to randomly turn off 50% of the neurons to obtain the final output of the first layer; using the final output of the first layer as the input of the second fully-connected layer, performing a linear transformation through 128 neurons, performing batch normalization and LeakyReLU activation, and applying Dropout processing with a rate of 0.3 to obtain the final feature representation; adding a residual connection to the final feature representation, performing element-wise addition of the temporal feature vector after dimensional transformation and the final feature representation, and calculating the probability distribution of each category through the Softmax function to obtain the optimized deep learning model.

[0010] In the fourth implementation of the first aspect, the method of optimizing the model according to the parameters, calculating the optimal control parameters for each station, generating a control strategy including an execution instruction sequence, and sending the control strategy to each execution unit includes: constructing an objective function based on the parameter optimization model, assigning different weight values to product quality indicators, production efficiency indicators, and energy consumption indicators, and establishing a multi-objective optimization mathematical model through linear weighted combination to obtain an optimization problem expression; setting constraint conditions for the optimization problem expression, including upper and lower limits of injection temperature, injection pressure range, mold temperature range, holding pressure time range, and cooling time range, and describing the process boundary through linear inequalities to obtain a set of constraint conditions; inputting the optimization problem expression and the set of constraint conditions into a solution module, searching the parameter space through an iterative calculation method, selecting multiple candidate parameter combinations in each iteration, calculating the objective function value, and retaining the parameter combination with a smaller objective function value to enter the next iteration. After multiple iterations, the initial control parameters for each station are obtained; performing a sensitivity test on the initial control parameters for each station, calculating the product quality change rate by changing the parameter values positively and negatively, marking the parameter as a highly sensitive parameter when the change rate exceeds the set threshold, setting a finer control range for the highly sensitive parameter, and generating a control parameter table considering stability; compiling an instruction sequence based on the control parameter table, dividing the control parameters into a startup section, a production section, and a shutdown section according to the process stage, each stage including parameter setting values, execution timing, and logical judgment conditions, and organizing them in the order of execution sequence to form the control strategy; distributing the control strategy to each execution unit, including a raw material supply unit, a drying unit, an injection molding unit, a mold temperature unit, and a post-treatment unit, through an industrial communication bus, sending a policy data packet and receiving an acknowledgment signal to confirm that the control strategy has been successfully sent to each execution unit.

[0011] In the fifth implementation of the first aspect, when compiling the instruction sequence based on the control parameter table, the control parameters are divided into a startup segment, a production segment, and a shutdown segment according to the process stage. Each stage includes parameter setting values, execution time sequences, and logical judgment conditions, which are organized in the order of execution to form the control strategy, including: performing segmented processing on the control parameter table, dividing the parameters into equipment preheating parameters, raw material drying parameter groups, injection molding parameter groups, and cooling and molding parameter groups according to the process flow, establishing a parameter linked list structure according to the production process to obtain a segmented parameter set; constructing a decision tree model for the segmented parameter set, setting equipment status judgment nodes, material status judgment nodes, and quality inspection judgment nodes, determining the parameter adjustment direction and adjustment amplitude according to the judgment results, and generating a multi-branch execution path; converting the multi-branch execution path into a state machine description, defining the initial state, conversion conditions, target state, and conversion actions, and establishing a complete process flow chart using the finite state machine representation method to obtain a state transition matrix; establishing a timing constraint relationship based on the state transition matrix, setting the time interval for parameter change, the waiting time for parameter response, and the timeout time for status confirmation, and generating an execution sequence list including timestamps; performing security verification on the execution sequence list, checking the parameter change rate, state jump amplitude, and equipment response ability, identifying potential risk points and adding protection logic to obtain a set of instructions after security verification; translating the set of instructions after security verification into a device control language, encapsulating it according to the communication protocol requirements of different execution units, and adding a checksum and a response mechanism to form the control strategy.

[0012] In the sixth implementation manner of the first aspect, monitoring the response results and control effects of the execution units and performing real-time updates on the parameter optimization model to form a closed-loop adaptive control system, including: collecting device operation data and quality inspection data from each of the execution units, recording the deviation between the actual execution value and the set value of the control parameter through an online monitoring system, and simultaneously obtaining product size, appearance, strength, and weight indicators to obtain an execution feedback data set; performing online analysis on the execution feedback data set, calculating the parameter fluctuation situation and the quality fluctuation trend through a sliding window technique, establishing the correlation between the parameter deviation and the quality change, and generating a process status evaluation report; constructing an incremental learning sample based on the process status evaluation report, packaging the parameter set value, actual execution value, and quality result of the current production batch as training samples, adding them to the model training data pool to form an updated training set; inputting the updated training set into a model update engine, performing parameter fine-tuning on the deep learning model through a transfer learning method, keeping the model skeleton structure unchanged and only updating the weight parameters to obtain an updated deep learning model; performing a verification test on the updated deep learning model, evaluating the model performance through a mixed validation set of historical data and current data, calculating the prediction accuracy and generalization ability indicators, judging the effectiveness of the model update, and generating a model performance evaluation result; dynamically adjusting the learning strategy according to the model performance evaluation result, reducing the learning rate or reverting to the previous version when the model performance deteriorates, and confirming the update and deploying the updated deep learning model as the latest parameter optimization model when the model performance improves to form the closed-loop adaptive control system.

[0013] In a second aspect, the present application provides an automatic control system for a plastic processing production line. The automatic control system for a plastic processing production line includes: A transmission module, configured to collect process parameters and equipment status data of each station of the plastic processing production line, transmit the collected data to a central control system through a distributed sensing network, and generate a production status database; An analysis module, configured to perform multi-dimensional parameter correlation analysis based on the production status database, and establish a mapping relationship between process parameters and product quality through a deep learning model that combines a convolutional neural network and a long short-term memory network. The deep learning model includes three convolutional layers for feature extraction, two LSTM layers for time series modeling, and two fully connected layers for parameter mapping to form a parameter optimization model; A generation module, configured to calculate the optimal control parameters of each station according to the parameter optimization model, generate a control strategy including an execution instruction sequence, and send the control strategy to each execution unit; An update module, configured to monitor the response results and control effects of the execution units, and perform real-time updates on the parameter optimization model to form a closed-loop adaptive control system.

[0014] In a third aspect, an automatic control device for a plastic processing production line is provided, including: a memory and at least one processor, wherein instructions are stored in the memory; the at least one processor invokes the instructions in the memory so that the automatic control device for the plastic processing production line executes the above-mentioned automatic control method for the plastic processing production line.

[0015] In a fourth aspect, a computer-readable storage medium is provided, in which instructions are stored, and when it runs on a computer, it causes the computer to execute the above-mentioned automatic control method for the plastic processing production line.

[0016] In the technical solution provided by this application, by collecting the process parameters and equipment status data of each work station, and combining with the distributed sensing network technology, the integration and transmission of the whole production process data are realized, effectively solving the technical problems of the lack of in-depth collaborative control between various devices and the fragmentation of control in each link in the traditional plastic processing production line, and providing a data basis for global optimization decision-making. Based on the production status database, multi-dimensional parameter correlation analysis is performed, and a deep learning model combining convolutional neural network and long short-term memory network is applied to realize the simultaneous extraction and processing of the spatial features and temporal features of the process parameters, breaking through the technical bottleneck that the traditional single model cannot comprehensively capture the complex correlations of the parameters, and establishing a more accurate mapping relationship between the process parameters and the product quality. The three-layer convolutional layer structure design adopted by the present invention enables the model to gradually extract the spatial correlation features between the process parameters from low dimension to high dimension, adapting to the characteristics of multi-parameter interaction in plastic processing; the introduction of two LSTM layers effectively captures the long-term and short-term impacts of parameter changes over time on the product quality, solving the problem that traditional control methods are difficult to handle temporal dependencies; the optimized design of the two fully connected layers significantly improves the stability and generalization ability of the model in industrial data processing by applying technical means such as batch normalization and Dropout. According to the parameter optimization model, the optimal control parameters of each work station are calculated and an execution instruction sequence is generated, replacing the subjectivity and uncertainty of the traditional control method relying on manual experience, and realizing data-driven scientific decision-making. The present invention monitors the response results and control effects of the execution unit to update the parameter optimization model in real time, forming a closed-loop adaptive control system, overcoming the limitation of the traditional control system lacking self-learning ability, and endowing the control system with the ability of continuous optimization. Especially in the specific application field of plastic processing, the deep learning algorithm of the present invention fully considers the industry characteristics, and specifically designs a network structure suitable for processing the interaction relationships of parameters such as temperature, pressure, and time, accurately modeling the complex non-linear relationships between the material characteristics, process conditions and product quality. The contribution of the algorithm features to the solution is prominently reflected in its ability to automatically extract the key patterns affecting the quality of plastic products from the massive production data, and optimize the parameters based on these patterns, avoiding the blindness and inefficiency of parameter adjustment in the traditional methods, significantly improving the automation level and product quality stability of plastic processing, while reducing energy consumption and raw material waste. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0018] Figure 1Schematic diagram of an embodiment of the automatic control method for a plastic processing production line in an embodiment of the present application; Figure 2 Schematic diagram of an embodiment of the automatic control system for a plastic processing production line in an embodiment of the present application; Figure 3 It is a structural schematic block diagram of the automatic control device for a plastic processing production line in an embodiment of the present invention. Specific embodiments

[0019] The embodiments of the present application provide an automatic control method and system for a plastic processing production line. Terms such as "first", "second", "third", "fourth", etc. (if any) in the specification, claims and above-mentioned drawings of the present application are used to distinguish similar objects and do not necessarily describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "comprising" or "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0020] For ease of understanding, the specific process of the embodiments of the present application will be described below. Please refer to Figure 1 , an embodiment of the automatic control method for a plastic processing production line in an embodiment of the present application includes: Step S101: Collect the process parameters and equipment status data of each station of the plastic processing production line, and transmit the collected data to the central control system through a distributed sensing network to generate a production status database; Step S102: Based on the production status database, perform multi-dimensional parameter correlation analysis, and establish a mapping relationship between process parameters and product quality through a deep learning model combining a convolutional neural network and a long short-term memory network. The deep learning model includes three convolutional layers for feature extraction, two LSTM layers for time series modeling, and two fully connected layers for parameter mapping to form a parameter optimization model; Step S103: According to the parameter optimization model, calculate the optimal control parameters for each station, generate a control strategy including an execution instruction sequence, and send the control strategy to each execution unit; Step S104: Monitor the response results and control effects of the execution units, and update the parameter optimization model in real time to form a closed-loop adaptive control system.

[0021] It is understandable that the execution entity of this application can be an automatic control system for a plastic processing production line, or it can also be a terminal or a server. Specifically, it is not limited here. The server is taken as the execution entity in the embodiments of this application for illustration.

[0022] Specifically, the process parameters and equipment status data of each station on the plastic processing production line are collected, and the collected data is transmitted to the central control system through a distributed sensing network to generate a production status database. The distributed sensing network is a decentralized monitoring system composed of multiple sensing nodes, which can collect various parameters at different stations and perform collaborative processing. In specific implementation, an RFID reader is installed in the raw material warehouse to identify the RFID tags of each bag of plastic raw materials, and information such as raw material type, batch number, and supplier information is obtained. RFID is the abbreviation of radio frequency identification technology, which can identify specific targets through radio signals and read and write relevant data. A multi-point temperature and humidity sensor and a weighing sensor array are set inside the raw material warehouse to monitor the raw material storage environment and inventory in real time. The temperature and humidity sensor uses a digital sensor with an accuracy of ±0.5°C and ±2%RH, and the weighing sensor uses a strain gauge type with an accuracy of ±0.05%. A hot air temperature sensor and a raw material moisture content detector are installed at the drying station to monitor the drying parameters. Pressure sensors, displacement sensors, temperature sensors, etc. are installed at the injection molding station to collect parameters such as the melting temperature, injection pressure, holding pressure time, mold temperature, cooling time, injection speed, clamping force, screw speed, and back pressure value of the injection molding machine. These sensors are connected to the edge computing node through a fieldbus (such as PROFINET, EtherCAT). The edge computing node is a small computing device deployed near the data source, which can preprocess the data and reduce the data transmission volume. The edge node is connected to the central control system through an industrial Ethernet, and the OPC UA protocol is used to realize data exchange. After receiving all the data, the central control system first aligns the timestamps to solve the problem of asynchronous data collection time at different stations. Then, data denoising is performed through wavelet transform. Wavelet transform is a time-frequency localization analysis method that can effectively filter out high-frequency noise. Next, outlier detection is performed, and a statistical-based method is used to identify outliers. For example, data points exceeding the mean ±3 times the standard deviation will be marked as outliers. The processed data is finally stored in a structured database to form a production status database. This solves the problem of mutual fragmentation in the control of each link of the traditional plastic processing production line and realizes the integration of the whole process data.

[0023] Based on the production status database, perform multi-dimensional parameter correlation analysis, and establish the mapping relationship between process parameters and product quality through a deep learning model that combines a convolutional neural network and a long short-term memory network. First, extract historical production data from the production status database and classify it according to product models and raw material batches. Then, reduce the data dimension through the principal component analysis algorithm. Principal component analysis is a statistical method that transforms a set of potentially related variables into a set of linearly uncorrelated variables through orthogonal transformation, retaining the main information of the original data. Specifically, calculate the data covariance matrix, solve the eigenvalues and eigenvectors, sort them by the size of the eigenvalues, and select the first few principal components whose cumulative contribution rate reaches a specified threshold. Next, perform a correlation analysis on the reduced-dimensional feature dataset, calculate the correlation strength between different parameters, and screen out the key process parameters that affect product quality. Then, construct a deep learning model that combines a convolutional neural network and a long short-term memory network. This hybrid model can simultaneously process the spatial features and temporal features of process parameters, solving the problem that a single model structure cannot comprehensively capture the complex correlations of parameters. The convolutional neural network part contains three convolutional layers, each using a different number of convolutional kernels to extract features from the process parameter sequence, and introducing a non-linear transformation through the ReLU activation function. Convolutional operations can extract local correlation features between parameters, and convolutional kernels in different layers extract different levels of features. The data after feature extraction is input into the long short-term memory network module, and the temporal relationship of the parameters is modeled through two LSTM layers. LSTM is a special recurrent neural network with three gating mechanisms: an input gate, a forget gate, and an output gate, which can effectively handle long-term dependence problems. The two-layer LSTM structure can hierarchically capture dependence relationships at different time scales, with the first layer handling short-term dependencies and the second layer handling long-term dependencies. The temporal feature vector output by the LSTM is finally mapped through two fully connected layers. The first layer contains more neurons, and the number of neurons in the second layer decreases, forming a funnel-shaped structure. Optimization means such as batch normalization and Dropout technology are adopted in the processing of the fully connected layer, effectively solving problems such as uneven industrial data distribution and unstable training. The finally trained model realizes the accurate mapping between process parameters and product quality, forming a parameter optimization model.

[0024] Optimize the model according to the parameters and calculate the optimal control parameters for each station. Construct an objective function, linearly weight and combine the product quality index, production efficiency index, and energy consumption index to establish a multi-objective optimization mathematical model. Then set the constraint conditions, including equipment performance limitations, process safety boundaries, etc. Taking the injection molding process as an example, parameters such as injection temperature, injection pressure, mold temperature, and holding time have clear upper and lower limit ranges. Input the objective function and constraint conditions into the solution module, and use the iterative optimization algorithm to search the parameter space to find the optimal solution. During specific implementation, first initialize multiple groups of candidate parameter combinations, input these parameters into the parameter optimization model to calculate the objective function value, retain the parameter combinations with smaller objective function values, slightly perturb these parameters to generate new candidate combinations, and repeat the above process multiple times until finally converging to the optimal solution. Conduct a sensitivity test on the determined control parameters to identify the highly sensitive parameters that have a significant impact on product quality, and set control intervals for these parameters. Then divide the optimized control parameters into a startup section, a production section, and a shutdown section according to the process stages to form a complete control instruction sequence. Finally, send the control strategy to each execution unit through the industrial communication bus, including the raw material supply unit, drying unit, injection molding unit, etc. This solves the problem that traditional control methods rely on manual experience and cannot effectively handle the complex non-linear relationship between process parameters and product quality.

[0025] Monitor the response results and control effects of the execution unit, and update the parameter optimization model in real time. First, collect the actual operation data from each execution unit, including equipment operation parameters and product quality inspection data. Calculate the parameter deviation by comparing the set value and the actual execution value of the control parameter. At the same time, obtain the product quality indicators, such as dimensional accuracy, surface quality, mechanical properties, etc. Analyze these feedback data online to establish the correlation between parameter deviation and quality change. According to the analysis results, construct the data of the current production batch into an incremental learning sample and add it to the model training data pool. Then use the transfer learning method to fine-tune the parameters of the deep learning model, keeping the model skeleton structure unchanged and only updating the weight parameters. Conduct a verification test on the updated model to evaluate the prediction accuracy and generalization ability. According to the model performance evaluation results, dynamically adjust the learning strategy to form a closed-loop adaptive control system. This solves the technical problem that traditional control systems lack adaptive learning ability and realizes the automatic optimization of control strategies.

[0026] In the embodiment of the present application, In a specific embodiment, the process of executing step S101 may specifically include the following steps: Perform RFID tag identification on the plastic raw material, collect the raw material type, batch number, and supplier information, and preliminarily process the collected information through the edge computing node to obtain the raw material basic data; A multi-point temperature and humidity sensor and a weighing sensor array are set inside the raw material bin to monitor the raw material storage environment parameters and inventory in real time, and raw material status data is obtained; Monitor the operating status of the injection molding machine, collect parameters such as melting temperature, injection pressure, holding pressure time, mold temperature, cooling time, injection speed, clamping force, screw speed, and back pressure value, and transmit them to the data acquisition unit through the industrial bus to obtain injection molding process data; Collect the energy consumption data of production equipment, measure current, voltage, power factor, active power, reactive power, and harmonic content through power parameter sensors installed on each equipment to obtain energy consumption data; Align the raw material basic data, raw material status data, injection molding process data, and energy consumption data through timestamps, establish a multi-time scale data synchronization matrix, and obtain a correlation data set; Perform data cleaning and outlier detection on the correlation data set, remove noise and outliers through wavelet transform and statistical filtering algorithms, and store the processed data in a structured database to form a production status database.

[0027] Specifically, RFID tag identification is performed on plastic raw materials to collect raw material-related information. RFID (Radio Frequency Identification) is a radio frequency identification technology that automatically identifies targets and obtains relevant data through radio signals. In this solution, each batch of plastic raw material bags is attached with an RFID tag, which contains data such as raw material types (such as PP, PE, ABS, etc.), batch numbers, supplier information, production dates, and material specifications. When the raw materials are put into storage, the RFID reader at the warehouse entrance scans and identifies the tags to obtain raw material information. The RFID reader sends the identified information to the edge computing node. The edge computing node is a small computing unit deployed near the data source and is responsible for preliminary data processing. The edge computing node performs format conversion on the raw material data, converts unstructured data into structured data; performs information screening, extracts key parameters such as melt index, density, and color; and at the same time performs redundancy check on the data to exclude duplicate records. The processed data forms standardized raw material basic data and is stored in the local cache, which not only reduces the burden on the central server but also reduces the network transmission pressure.

[0028] Meanwhile, multi-point temperature and humidity sensors and a weighing sensor array are installed inside the raw material bin to monitor the raw material storage environment parameters and inventory in real time. The temperature and humidity sensors are digital sensors and are distributed in a 3×3×3 three-dimensional grid layout at various positions in the raw material bin, which can monitor the three-dimensional distribution of temperature and humidity in the warehouse and identify potential abnormal areas. The weighing sensors are installed under the raw material storage racks and convert gravity into an electrical signal through the strain bridge principle to monitor the inventory of various raw materials in real time. The data collected by these sensors are processed by the local signal conditioning circuit, amplified and filtered, and then transmitted to the raw material bin monitoring unit through the RS485 bus or a wireless communication module (such as ZigBee, LoRa). The raw material bin monitoring unit combines the temperature and humidity data and the inventory data to form complete raw material status data, including information such as the current inventory of each raw material and the storage temperature and humidity conditions.

[0029] A variety of special sensors are installed on the injection molding machine to collect process parameters in real time. The melt temperature is measured by thermocouples, which are arranged in different areas of the barrel; the injection pressure is monitored by pressure sensors, which are installed at key positions in the injection system; the holding pressure time and cooling time are recorded by the injection molding machine controller; the injection speed is calculated by differentiating the displacement sensor and time; the clamping force is measured by a strain gauge force sensor; the screw speed is obtained through an encoder or a Hall sensor; and the back pressure value is collected by a special back pressure sensor. The data of these sensors are transmitted to the data acquisition unit through an industrial field bus (such as PROFIBUS, DeviceNet, EtherCAT, etc.). The data acquisition unit collects each parameter at a preset sampling frequency (usually 10 - 100Hz) to form time series data, constituting complete injection molding process data.

[0030] Power parameter sensors are installed on the power lines of the main equipment such as injection molding machines, dryers, and mold temperature controllers. A three-phase power parameter measurement module is used to monitor parameters such as current, voltage, power factor, active power, reactive power, and harmonic content at the same time. The power parameter acquisition frequency is usually lower than the process parameter, generally set to 1Hz, which is sufficient to reflect the change of energy consumption. These data are transmitted to the energy monitoring module through industrial Ethernet to form an equipment energy consumption database, recording the energy consumption of each equipment under different working conditions.

[0031] When integrating raw material basic data, raw material status data, injection molding process data, and energy consumption data, it is first necessary to solve the problem of time asynchronization caused by different data sources and different sampling frequencies. Through timestamp alignment processing, a multi-time scale data synchronization matrix is established. The specific operation is based on the lowest sampling frequency. Aggregate high-sampling frequency data according to time windows (take the average, maximum, or minimum value) to ensure that all data has a unified time point. For example, if energy data is collected at 1Hz and injection molding process data is collected at 50Hz, then 50 injection molding data points are aggregated into 1 point, corresponding to the energy data. In this way, a correlation dataset containing all parameters is formed, and each record contains raw material information, environmental status, process parameters, and energy consumption data at the same moment.

[0032] Data cleaning and outlier detection of the correlation dataset are key steps to ensure data quality. Data cleaning includes handling missing values, removing duplicate records, correcting format errors, etc. For missing values, select an appropriate imputation method according to the data type, such as linear interpolation, forward value filling, or mean filling. Outlier detection uses wavelet transform and statistical filtering algorithms. Wavelet transform is a time-frequency localization analysis method. By decomposing the signal into different frequency components, it can effectively identify abnormal fluctuations. In specific implementation, select an appropriate wavelet basis function (such as Daubechies wavelet), perform multi-scale decomposition on the signal, and suppress high-frequency noise components during the reconstruction process. The statistical filtering algorithm is based on the statistical characteristics of the data. For example, use the 3σ principle to mark data points that deviate from the mean by more than 3 standard deviations as outliers, and correct or delete them. The processed data is finally stored in a structured database, using a relational database or a time series database to establish a production status database.

[0033] In a specific embodiment, the process of executing step S102 may specifically include the following steps: Extract historical production data from the production status database, classify it according to product model and raw material batch, and reduce the data dimension through the principal component analysis algorithm to obtain a feature dataset; Perform a correlation analysis on the feature dataset, calculate the correlation strength between parameters through Pearson correlation coefficient and mutual information calculation methods, and screen out the key process parameters that affect product quality to form a key parameter set; Based on the key parameter set, construct a convolutional neural network module, set three convolutional layers, each layer uses 64, 128, and 256 3×3 convolutional kernels respectively, extract features from the process parameter sequence, and introduce non-linear transformation through the ReLU activation function to obtain a feature mapping matrix; Input the feature mapping matrix into the long short-term memory network module. Through two LSTM layers, each containing 128 neurons, model the parametric temporal relationship, and combine the forget gate, input gate, and output gate mechanisms to handle long-term dependencies, obtaining a temporal feature vector; Process the temporal feature vector through two fully connected layers. The first layer contains 256 neurons, and the second layer contains 128 neurons. Use the Dropout technique to prevent overfitting and improve the training stability through batch normalization, obtaining an optimized deep learning model; Conduct a sensitivity analysis on the optimized deep learning model. Calculate the output change rate by perturbing the input parameters, construct a parameter importance ranking matrix, and combine the expert rule system to form a parameter optimization model.

[0034] Specifically, extract historical production data from the production status database. These data include information on raw materials, process parameters, equipment status, and product quality in multiple dimensions. The data extraction uses SQL query statements, setting filtering conditions such as time range, product type, etc., such as "SELECT * FROM production_data WHERE production_date BETWEEN '2024-01-01' AND '2024-03-31'". The extracted data is classified according to product models (such as car bumpers, instrument panels, interior parts, etc.) and raw material batches. The hierarchical clustering method is used to group data with similar characteristics into one group. For product model classification, direct classification is carried out using product codes; for raw material batch classification, clustering is based on the physical properties of raw materials (such as melt index, density). The classified data constitutes multiple sub-datasets, and each sub-dataset represents the production situation of a specific product model and raw material batch combination.

[0035] Since plastic processing production involves numerous parameters, usually including dozens or even hundreds of dimensions such as temperature, pressure, and time, directly modeling has a large computational amount and is prone to introducing noise. Therefore, it is necessary to reduce the data dimension through the principal component analysis algorithm. Principal Component Analysis (PCA) is a statistical method that transforms potentially related variables into a set of linearly uncorrelated variables through orthogonal transformation. These new variables are called principal components. The specific operation steps include: first, calculate the covariance matrix of the dataset, then solve the eigenvalues and eigenvectors of this matrix, sort them by eigenvalue size, select the top K eigenvectors to form a projection matrix, and finally transform the original data to the new coordinate system through the projection matrix to obtain the data after dimension reduction. The selection of the K value is usually based on the cumulative contribution rate, that is, select the number of principal components that can explain at least 85% of the variance of the original data. The data processed by principal component analysis is called the feature dataset, with a significantly reduced dimension but retaining the main information of the original data.

[0036] Perform a correlation analysis on the feature dataset to identify which process parameters have the greatest impact on product quality. The correlation analysis uses two complementary methods: Pearson correlation coefficient and mutual information calculation. The Pearson correlation coefficient measures the linear correlation between two variables, with a value range of [-1, 1]. A value of 1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no correlation. When calculating, the data is first standardized, and then the covariance of the two standardized variables is calculated. Mutual information is an indicator that measures the degree of mutual dependence between two random variables and can capture non-linear correlations. Mutual information is calculated based on probability distributions and is solved by estimating the joint probability distribution and marginal probability distribution. Based on the correlation metrics calculated by the two methods, rank the association strengths between all process parameters and product quality metrics (such as dimensional accuracy, surface quality, strength, etc.), and select the parameters with higher correlation coefficients or mutual information values to form a key parameter set. This key parameter set contains process parameters that have a significant impact on product quality, usually reducing the number of parameters to 20-30% of the original, significantly reducing the complexity of subsequent modeling.

[0037] Construct a convolutional neural network module for feature extraction based on the key parameter set. A Convolutional Neural Network (CNN) is a deep learning model specifically designed to process data with a grid topology structure and is used in this method to extract the spatial features of process parameters. Set three convolutional layers. The first layer uses 64 3×3 convolutional kernels, the second layer uses 128 3×3 convolutional kernels, and the third layer uses 256 3×3 convolutional kernels. The number of convolutional kernels increases layer by layer to capture features from low-level to high-level. The convolution operation performs local feature extraction on the process parameter sequence through a sliding window method. After each layer of convolution, a batch normalization layer and a ReLU activation function are connected. The ReLU (Rectified Linear Unit) activation function has the form f(x)=max(0,x), that is, the input less than 0 outputs 0, and the input greater than 0 remains unchanged. This non-linear transformation enhances the model's expressive ability. The output of the convolutional network is a multi-dimensional feature mapping matrix that contains the feature representations extracted from the original parameters.

[0038] The feature mapping matrix is input into the long short-term memory network module to model the parametric temporal relationship. The long short-term memory (LSTM) network is a special type of recurrent neural network that is good at dealing with long-term dependencies in sequential data. This method uses a two-layer LSTM structure, with each layer containing 128 neurons. The first layer of LSTM captures short-term temporal features, and the second layer of LSTM integrates long-term dependencies. The core of LSTM lies in its unique gating mechanism, including the forget gate, input gate, and output gate. The forget gate determines what information to discard, the input gate determines what information to update, and the output gate determines what information to output. Through these three gating mechanisms, LSTM can effectively handle the information transmission problem in long sequence data, avoiding the problem of gradient vanishing or explosion in traditional recurrent neural networks. The output after LSTM processing is a vector containing temporal features, which integrates the variation characteristics of process parameters in the time dimension.

[0039] The temporal feature vector is processed through two fully connected layers to complete the final feature mapping. The fully connected layer is also called the dense layer, which is the most basic layer type in neural networks, and each neuron is connected to all neurons in the previous layer. The first fully connected layer contains 256 neurons, and the second layer contains 128 neurons, presenting a funnel-shaped structure, gradually reducing the feature dimension and extracting more abstract feature representations. Multiple optimization techniques are adopted in the processing of the fully connected layer: Dropout is a regularization technique that randomly shuts down a certain proportion of neurons during training to prevent the model from overfitting; Batch Normalization accelerates the training process by normalizing the input of each layer, improves training stability, and alleviates the problem of internal covariate shift. The output of the fully connected layer directly corresponds to the product quality index, forming a complete mapping relationship to obtain the optimized deep learning model.

[0040] Sensitivity analysis is performed on the optimized deep learning model to evaluate the influence degree of different parameters on the model output. Sensitivity analysis is achieved by perturbing the input parameters and observing the output changes. The specific operation is to add a small perturbation (such as increasing or decreasing by 5%) to each input parameter, record the output change rate, and the larger the change rate, the more significant the parameter influence. The calculation formula is: change rate = (perturbed output - original output) / original output / perturbation amplitude. Based on the sensitivity analysis results, a parameter importance ranking matrix is constructed to identify the process parameters that are most critical to product quality. Finally, combined with the expert rule system, the data-driven analysis results are combined with the experience of industry experts to form a parameter optimization model. The expert rule system includes rules such as process safety boundaries and material property constraints to ensure that the parameters generated by the model are within the actual operable range.

[0041] In a specific embodiment, the process of executing step S103 may specifically include the following steps: Perform data normalization on the time series feature vector. Calculate the mean and standard deviation of each dimension feature through the Z-score normalization method, convert the feature data into a standard normal distribution with a mean of 0 and a standard deviation of 1, and obtain the normalized feature vector; Input the normalized feature vector into the first fully connected layer, perform a linear transformation on the input features through 256 neurons, calculate the weighted sum by applying the weight matrix and the input vector, and obtain the original output value of the first layer; Perform batch normalization on the original output value of the first layer. Adjust the feature distribution by calculating the mean and variance of the samples within the mini-batch, and introduce the learnable scaling parameter γ and translation parameter β to obtain the normalized output value; Apply the LeakyReLU activation function to the normalized output value, keep the positive values unchanged, multiply the negative values by the slope coefficient of 0.01, introduce a non-linear transformation, and apply Dropout processing with a rate of 0.5 to randomly turn off 50% of the neurons to obtain the final output of the first layer; Use the final output of the first layer as the input of the second fully connected layer, perform a linear transformation through 128 neurons, perform batch normalization and LeakyReLU activation, and apply Dropout processing with a rate of 0.3 to obtain the final feature representation; Add a residual connection to the final feature representation, perform an element-wise addition of the time series feature vector after dimensional transformation and the final feature representation, and calculate the probability distribution of each category through the Softmax function to obtain the optimized deep learning model.

[0042] Specifically, the time series feature vector is the feature extracted from the long short-term memory network and contains the information of the process parameters changing over time. Since the dimensions and numerical ranges of different parameters vary greatly (for example, the temperature ranges from 200 to 300 °C, while the pressure may be in the range of dozens of MPa), directly inputting into the fully connected layer will lead to unstable training. The Z-score normalization method is a commonly used data normalization technique. By calculating the mean and standard deviation of each dimension feature, the data is converted into a standard normal distribution with a mean of 0 and a standard deviation of 1. The specific operation is to calculate the mean and standard deviation of each dimension of the time series feature vector respectively, and then subtract the mean and divide by the standard deviation. For example, for the injection molding temperature feature dimension, first calculate the mean (such as 250 °C) and standard deviation (such as 10 °C) of this dimension for all samples, and then subtract 250 °C from each sample's temperature value and divide by 10 °C to obtain the normalized value. After such processing, most of the temperature feature values are distributed between -3 and 3, which is in a similar numerical range as other features such as pressure and time, eliminating the training bias caused by dimensional differences and obtaining the normalized feature vector.

[0043] The standardized feature vectors are input into the first fully connected layer for processing. The fully connected layer is the most basic layer type in a deep learning network, where each neuron is connected to all neurons in the previous layer. The first fully connected layer contains 256 neurons, and each neuron receives all input features. A weighted sum is calculated through the linear combination of the weight matrix and the input vector. The specific calculation process is as follows: the input vector performs a dot product operation with the weight vector corresponding to each neuron, and then the bias term is added. If the dimension of the input feature vector is n, the dimension of the weight matrix is 256×n, and each row represents the weight vector of a neuron. The values in the weight matrix are randomly initialized before training and then continuously optimized through the backpropagation algorithm. The linear transformation operation of the fully connected layer essentially searches for hyperplanes in a high-dimensional space that can distinguish different quality levels, but the expressive power of a simple linear transformation is limited. Therefore, subsequent non-linear processing is required. The calculation results of the fully connected layer are called the original output values, and the distribution of these values is usually very uneven. Directly using them for the next step of processing will affect the training stability. Therefore, normalization processing is required.

[0044] Batch normalization is performed on the original output values of the first layer. This is an effective technique for solving the problem of internal covariate shift in deep learning training. Internal covariate shift refers to the change in the input distribution of each layer during the network training process due to parameter updates, which causes subsequent layers to continuously adapt to this change and slows down the training speed. Batch normalization solves this problem by normalizing the input of each layer. The specific operation is to calculate the mean and variance of each feature dimension within a mini-batch (a small batch of training samples, usually 32 or 64 samples), then perform normalization, and then introduce learnable scaling parameter γ and translation parameter β for adjustment. γ and β are trainable parameters and are optimized through the gradient descent algorithm, allowing the model to learn the optimal scaling and offset values for each feature. Batch normalization not only accelerates the training process but also acts as a regularization, reducing the dependence on other regularization techniques such as Dropout. The processed data is called the normalized output values, and the distribution is more uniform.

[0045] Apply the LeakyReLU activation function to the normalized output values to introduce non-linear transformation. The activation function is a key link in introducing non-linearity in neural networks. Without an activation function, a multi-layer neural network is equivalent to a single-layer linear model with limited expressive power. LeakyReLU is an improved version of the ReLU activation function, which solves the problem of neuron death caused by the derivative of the ReLU function being 0 in the negative half-axis. LeakyReLU remains unchanged for positive values and multiplies negative values by a small slope coefficient (such as 0.01), so that there is also a small gradient in the negative value interval, which helps the gradient transmission during the training process. In plastic processing technology, the relationship between parameters is often non-linear. For example, the influence of temperature and pressure on product quality is not a simple linear superposition. LeakyReLU can effectively capture this non-linear relationship. After the activation function, apply the Dropout technique to randomly turn off a certain proportion (such as 50%) of the neurons. Dropout is a powerful regularization technique that prevents the model from overfitting and improves the generalization ability by randomly deactivating some neurons during the training process. The specific implementation of Dropout is to directly set the output values of some neurons to zero, which is equivalent to these neurons not participating in the calculation in the current training batch. The final output of the first layer obtained after such processing contains the preliminary abstraction and refinement of the original features.

[0046] Use the final output of the first layer as the input for the second fully-connected layer for processing. The second fully-connected layer contains 128 neurons, and the number of neurons is less than that of the first layer, presenting a "pyramid" or "funnel" structure, which is a common design pattern in deep learning models and helps to gradually extract higher-level abstract features. The processing logic of the second fully-connected layer is similar to that of the first layer, including linear transformation, batch normalization, LeakyReLU activation, and Dropout processing. The difference is that the Dropout rate of the second layer is set to 0.3, which is lower than 0.5 of the first layer. This is because neurons closer to the output layer have a greater impact on the final result, and too high a Dropout rate may lose important information. After processing by the second fully-connected layer, the final feature representation is obtained. This feature representation has performed multi-level abstraction and refinement on the original process parameters and contains the key information about the influence on product quality.

[0047] Add a residual connection to the final feature representation, and perform element-wise addition of the temporal feature vector after dimensional transformation and the final feature representation. Residual connection is an effective technique to solve the problem of vanishing gradients in the training of deep neural networks. By constructing a "shortcut", it allows information to directly skip multiple layers from early layers to subsequent layers. In this method, to add a residual connection, the temporal feature vector needs to be dimensionally transformed first to match its dimension with the final feature representation. Dimensional transformation is usually achieved through a linear mapping layer, which maps the temporal feature vector to the same dimensional space as the final feature representation. Then, these two vectors are element-wise added, which is equivalent to retaining a part of the original feature information on the basis of the deep features, helping to improve the stability and accuracy of model training. Finally, the result is converted into a probability distribution through the Softmax function. The Softmax function is a function that converts a vector into a probability distribution. For each element of the input vector, the exponential is calculated first, and then divided by the sum of the exponentials of all elements to obtain a value between 0 and 1, and the sum of all values is 1. In plastic processing control, the output of the Softmax function can represent the probability distribution of product quality grades, such as the probabilities of high-quality products, qualified products, and defective products. Through this series of processes, an optimized deep learning model is obtained, which can accurately predict the impact of process parameters on product quality.

[0048] In a specific embodiment, the process of performing step S104 may specifically include the following steps: Based on the parameter optimization model, construct an objective function, assign different weight values to the product quality index, production efficiency index, and energy consumption index, and establish a multi-objective optimization mathematical model through linear weighted combination to obtain an optimization problem expression; Set constraint conditions for the optimization problem expression, including the upper and lower limits of injection temperature, injection pressure range, mold temperature range, holding pressure time range, and cooling time range, and describe the process boundaries through linear inequalities to obtain a set of constraint conditions; Input the optimization problem expression and the set of constraint conditions into the solution module, search the parameter space through an iterative calculation method. In each round of iteration, select multiple candidate parameter combinations, calculate the objective function value, and retain the parameter combination with a smaller objective function value to enter the next round of iteration. After multiple rounds of iteration, obtain the initial control parameters for each station; Perform a sensitivity test on the initial control parameters for each station. By making positive and negative changes to the parameter values, calculate the product quality change rate. When the change rate exceeds the set threshold, mark the parameter as a highly sensitive parameter, and set a more refined control range for the highly sensitive parameter to generate a control parameter table considering stability; Compile an instruction sequence based on the control parameter table. Divide the control parameters into a startup segment, a production segment, and a shutdown segment according to the process stages. Each stage includes parameter setting values, execution time sequences, and logical judgment conditions, and is organized in the order of execution to form a control strategy; Distribute the control strategy to each execution unit through the industrial communication bus, including the raw material supply unit, drying unit, injection molding unit, mold temperature unit, and post-treatment unit, send the policy data packet and receive the confirmation signal to confirm that the control strategy is successfully sent to each execution unit.

[0049] Specifically, calculate the optimal control parameters and generate the control strategy according to the parameter optimization model, which solves the problem that traditional control methods rely on manual experience and cannot effectively handle the complex non-linear relationship between process parameters and product quality. Construct the objective function based on the parameter optimization model. The objective function is the core of the optimization process and is used to evaluate the advantages and disadvantages of different combinations of process parameters. In plastic processing production, it is usually necessary to comprehensively consider three aspects of indicators: product quality, production efficiency, and energy consumption. Product quality indicators include dimensional accuracy, surface quality, mechanical strength, etc.; production efficiency indicators include production cycle time, production capacity, qualification rate, etc.; energy consumption indicators include power consumption per unit product, raw material utilization rate, etc. These indicators often restrict each other. For example, improving product quality may require extending the production cycle, and reducing energy consumption may affect product quality. Therefore, it is necessary to establish a multi-objective optimization mathematical model. The specific method is to assign different weight values to each indicator and form a single objective function through linear weighted combination. The setting of the weight value reflects the relative importance of different indicators and is usually determined by process experts according to product characteristics and production requirements. For example, for high-precision electronic enclosures, the product quality weight may be set to 0.6, and the production efficiency and energy consumption are 0.3 and 0.1 respectively; for ordinary packaging materials, the production efficiency weight may be higher. The objective function is usually designed to be as small as possible. Therefore, for indicators that need to be maximized (such as product strength), its negative value or reciprocal needs to be taken. Multiple indicators are transformed into dimensionless relative values through normalization processing to eliminate the influence of dimensional differences and obtain the final expression of the optimization problem.

[0050] Setting constraints on the optimization problem expression is an important step to ensure that the optimization result falls within the feasible region. In plastic processing and production, the constraints mainly come from process boundaries and equipment limitations. The upper and lower limits of the injection temperature are determined according to the raw material characteristics. For example, the injection temperature of polypropylene (PP) is usually in the range of 190 - 280 °C. Temperatures above this range will cause raw material degradation, and temperatures below this range will result in poor fluidity. The injection pressure range is limited by the performance of the injection molding machine, generally between 50 - 150 MPa. The mold temperature range is related to the cooling and solidification of the product. For example, the mold temperature of PP is usually in the range of 20 - 80 °C. The holding pressure time range affects the density and size of the product, generally between 1 - 30 seconds. The cooling time range determines the demolding state of the product, usually related to the wall thickness and material of the product. Thicker products require longer cooling times. These constraints are described by linear inequalities in the form of parameter lower limit ≤ parameter ≤ parameter upper limit. In addition to the constraints on individual parameters, sometimes the associated constraints between parameters need to be considered. For example, the difference between the injection temperature and the mold temperature needs to be within a certain range to avoid excessive product stress caused by too large a temperature difference. All the constraint conditions are combined to form a set of constraint conditions, which defines the feasible region in the parameter space. The optimization process only searches for the optimal solution within this feasible region.

[0051] Input the optimization problem expression and the set of constraint conditions into the solution module, and search the parameter space through an iterative calculation method. Iterative optimization is a commonly used method for solving complex optimization problems. By continuously improving the parameter combinations, it gradually approaches the optimal solution. In plastic processing control, heuristic optimization algorithms such as genetic algorithms and particle swarm algorithms are used to search the parameter space. Taking the genetic algorithm as an example, first, randomly generate multiple initial parameter combinations that satisfy the constraint conditions, which are called the initial population. Each parameter combination includes all the parameters to be optimized, such as injection temperature, injection pressure, mold temperature, etc. Input these parameter combinations into the parameter optimization model (i.e., the previously trained deep learning model), predict the corresponding product quality, production efficiency, and energy consumption indicators, and then substitute them into the objective function to calculate the objective function value. According to the magnitude of the objective function value, select the parameter combinations with smaller objective function values and retain them, and eliminate the combinations with larger objective function values. The retained parameter combinations generate new parameter combinations through crossover and mutation operations to form a new generation of population. The crossover operation means that two parameter combinations exchange some parameter values; the mutation operation means that the parameter values randomly vary within the constraint range. The new generation of population is input into the parameter optimization model again, calculate the objective function value, perform selection, crossover, and mutation to form an updated population. Through multiple generations of iterative evolution, the parameter combinations in the population are gradually optimized and finally converge to the optimal or approximately optimal parameter combination to obtain the initial control parameters for each station.

[0052] Perform a sensitivity test on the initial control parameters of each work station. This step is to evaluate the impact of changes in different parameters on product quality and ensure the stability and robustness of the control strategy. The basic idea of the sensitivity test is to observe the change in product quality by making small perturbations to the parameter values. The specific method is to increase or decrease a single parameter by a certain percentage (such as ±2%) while keeping other parameters unchanged, and use the parameter optimization model to predict the change in product quality. Calculate the product quality change rate, which is the ratio of the quality change amount to the parameter change amount. The larger the change rate, the more sensitive the parameter is to product quality. When the change rate exceeds a preset threshold (such as 0.5), mark this parameter as a highly sensitive parameter. For highly sensitive parameters, a more refined control range needs to be set, such as increasing the control accuracy by one order of magnitude or narrowing the parameter fluctuation range. This is because small fluctuations in highly sensitive parameters will cause significant changes in product quality and more stringent control is required. Through the sensitivity test, a control parameter table considering stability is obtained, which includes the optimal value and allowable fluctuation range of each process parameter, as well as the sensitivity level mark.

[0053] Compile an instruction sequence based on the control parameter table. Divide the control parameters according to the process stages to form a complete control strategy. Plastic processing production is generally divided into three stages: startup stage, production stage, and shutdown stage. In the startup stage, parameter settings focus on equipment preheating and system stability, such as gradually increasing the temperature and running at low speed without load; in the production stage, parameter settings focus on product quality and production efficiency, including the core process parameters during formal production; in the shutdown stage, parameter settings focus on equipment protection and maintaining raw material performance, such as cooling down and cleaning the machine. Each stage is further divided into multiple sub-stages. For example, the production stage includes sub-stages such as mold closing, injection, holding pressure, cooling, mold opening, and ejection. Set the parameter setting values, execution time sequences, and logical judgment conditions for each sub-stage respectively. The parameter setting values are determined according to the optimal parameters obtained from the previous optimization; the execution time sequence specifies the time points and order of parameter changes; the logical judgment conditions are used to handle abnormal situations or stage transition conditions, such as "if the pressure in the holding pressure stage is lower than the threshold, end the holding pressure in advance and enter the cooling stage". All parameters and logical instructions are arranged and organized in the order of execution to form a complete control strategy. This control strategy is actually a state machine that can automatically switch to the next state according to the current state and sensor feedback, realizing the automatic control of the production process.

[0054] The control strategy is distributed to each execution unit through the industrial communication bus to complete the issuance of control instructions. The industrial communication bus is a dedicated network for device - to - device communication in the field of industrial automation. Common ones include PROFIBUS, DeviceNet, EtherCAT, etc. In this method, the control strategy is split into instruction subsets suitable for each execution unit. The raw material supply unit receives raw material pretreatment parameters; the drying unit receives parameters such as drying temperature and time; the injection molding unit receives parameters such as injection molding temperature, pressure, and speed; the mold temperature unit receives mold temperature control parameters; the post - processing unit receives finished product processing parameters. The instructions received by each unit include parameter set values and logic control instructions, and these instructions are packaged into data packets conforming to the communication protocol. The data packet usually includes fields such as unit address, instruction type, parameter value, and checksum, and is transmitted to each execution unit through the industrial bus. After receiving the instructions, the execution unit performs corresponding actions and returns an acknowledgment signal indicating that the instructions have been received and executed. The central controller receives the acknowledgment signal and verifies whether the control strategy has been successfully issued to all execution units. If a certain unit does not return an acknowledgment signal, it will resend or give an alarm prompt.

[0055] In a specific embodiment, the process of performing step S105 may specifically include the following steps: Perform segmented processing on the control parameter table. According to the technological process, the parameters are divided into equipment pre - heating parameter groups, raw material drying parameter groups, injection molding parameter groups, and cooling molding parameter groups. A parameter linked - list structure is established according to the production process to obtain a segmented parameter set; Construct a decision - tree model for the segmented parameter set, set equipment - state judgment nodes, material - state judgment nodes, and quality - inspection judgment nodes. According to the judgment results, determine the parameter adjustment direction and adjustment amplitude, and generate multi - branch execution paths; Convert the multi - branch execution paths into state - machine descriptions, define the initial state, conversion conditions, target states, and conversion actions, and establish a complete technological process flowchart using the finite - state - machine representation method to obtain a state - transition matrix; Based on the state - transition matrix, establish timing constraint relationships, set the time interval for parameter changes, the waiting time for parameter responses, and the timeout time for state confirmation, and generate an execution sequence list containing timestamps; Perform safety verification on the execution sequence list, check the parameter change rate, state jump amplitude, and device response ability, identify potential risk points and add protection logic to obtain a set of instructions after safety verification; Translate the set of instructions after safety verification into device control language, package them according to the communication protocol requirements of different execution units, and add a checksum and an acknowledgment mechanism to form a control strategy.

[0056] Specifically, the control parameter table is segmented. According to the process characteristics of plastic processing technology, the parameters are divided into different functional groups. The equipment preheating parameter group includes the start-up and preheating parameters of each equipment, such as the temperature rise curve of each area of the injection molding machine, the pre-rotation speed of the screw, the temperature rise setting of the mold temperature machine, etc.; the raw material drying parameter group includes the raw material pretreatment parameters such as drying time, drying temperature, air flow rate, etc.; the injection molding parameter group is the core process parameter, including injection pressure, injection speed, holding pressure, holding time, etc.; the cooling and molding parameter group includes parameters such as cooling time, demolding temperature, ejection force, etc. This segmentation method is divided according to the time sequence and functional relevance of the production process, making the complex parameter system become well-organized. The segmented parameters are organized through a linked list structure. A linked list is a linear data structure, and each node contains a parameter value and a pointer to the next node, which is suitable for representing a sequential execution process flow. In this method, each node of the linked list contains all the parameter values of this stage, as well as the conditions and timing for transitioning to the next stage. In this way, a complete set of segmented parameters is formed, laying a foundation for the subsequent execution path planning.

[0057] A decision tree model is constructed for the set of segmented parameters to handle various situations and conditional branches in the production process. A decision tree is a judgment model in the form of a tree structure. Starting from the root node, it decides which branch to go to by judging the truth or falsehood of the conditions, and finally reaches the leaf node to obtain the result. In this method, the judgment nodes of the decision tree are divided into three categories: equipment status judgment nodes are used to evaluate whether the equipment is in a normal state, such as "whether the temperature of the injection molding machine reaches the set value"; material status judgment nodes are used to evaluate whether the raw materials meet the processing conditions, such as "whether the moisture content of the raw materials is lower than the threshold"; quality inspection judgment nodes are used to evaluate whether the product quality meets the standards, such as "whether the product size is within the allowable range". These judgment nodes are defined through logical expressions, such as "IF temperature > 230°C THEN enter the next stage ELSE continue heating". The construction of the decision tree is based on process knowledge and historical data, and the thresholds of each judgment condition are determined through the output results of the deep learning model. The judgment results are used to determine the parameter adjustment direction and adjustment range, such as "IF the size is too large THEN reduce the holding pressure by 5%". Each leaf node corresponds to a specific parameter adjustment plan, and all possible combinations of execution paths form a multi-branch execution path, covering various situations that may be encountered in the production process.

[0058] Convert the multi-branch execution path into a state machine description to achieve more standardized process flow control. A finite state machine is a computational model consisting of a finite number of states, input events, a transition function, and output actions. In this method, an initial state (such as "equipment standby"), various intermediate states (such as "heating", "injecting", "holding pressure"), and a termination state (such as "production completed") are defined. Each state corresponds to a stage in the process flow, and the transitions between states are triggered by transition conditions. The transition conditions can be time conditions (such as "preheating time reaches 30 minutes"), sensor feedback conditions (such as "temperature reaches the set value"), or operator intervention conditions. Each state transition is accompanied by a transition action, that is, performing corresponding parameter adjustments or equipment operations. The entire state machine structure is described using the finite state machine notation, including the state set, input alphabet, transition function, initial state, and accepting state. All states and transition relationships are organized into a state transition matrix, where the rows of the matrix represent the current state, the columns represent the input conditions, and the matrix elements represent the state after the transition and the actions to be performed. This representation method is intuitive and clear, easy to implement and verify, and provides a basis for subsequent timing control.

[0059] Establish timing constraint relationships based on the state transition matrix to ensure that parameter adjustments and state transitions occur in the correct time sequence. In the plastic processing process, timing control is very crucial. For example, the injection pressure needs to be applied after the mold is closed, and the holding pressure needs to start immediately after the injection ends, etc. The timing constraint relationships are achieved by setting three time parameters: the time interval of parameter change stipulates the minimum time interval between two consecutive parameter adjustments, preventing the system from becoming unstable due to overly frequent parameter adjustments; the waiting time for parameter response defines how long to wait after a parameter adjustment to evaluate the effect, considering the response delay of the system; the timeout time for state confirmation sets the maximum waiting time for state transitions, and if this time is exceeded, an exception handling is triggered. These time parameters are determined according to the equipment characteristics, raw material properties, and process requirements, and are embedded into the execution sequence in the form of timestamps. Each execution instruction carries a timestamp indicating the execution time point of the instruction or the delay time relative to the previous instruction. All instructions with timestamps are arranged in chronological order to form a complete execution sequence list, which describes the timing control strategy of the entire production process.

[0060] Perform security verification on the execution sequence list to ensure the reliability and security of the control strategy. The security verification includes three main aspects: The parameter change rate check is used to verify whether the parameter adjustment is too fast. For example, a temperature change rate exceeding 10°C per minute may cause excessive stress on the device; the state jump amplitude check is used to verify whether the state transition is reasonable. For example, directly jumping from "low-speed injection" to "high-speed cooling" is unreasonable and requires an intermediate state; the device response ability check is used to verify whether the device can perform the required actions. For example, it is not feasible to require an injection speed exceeding the maximum capacity of the device. Through these checks, potential risk points are identified, such as too fast parameter changes, unreasonable state jumps, and device overload. For the identified risk points, corresponding protection logics are added, such as gradual parameter adjustment, intermediate state transition, load limit, etc. These protection logics are embedded into the control strategy as additional conditions to form a set of instructions after security verification, ensuring that the control strategy will not cause device damage or production accidents due to design defects during actual execution.

[0061] Translate the set of instructions after security verification into specific device control languages to achieve the conversion from an abstract control strategy to actual device instructions. Different devices may use different control languages and communication protocols, and targeted instruction translation is required. Common device control languages include G-code (for numerical control devices), PLC ladder diagram language (for programmable controllers), and special device instruction sets. The translation process needs to consider the characteristics and limitations of the device and decompose the high-level control strategy into basic instructions that the device can understand. After the instruction translation, it is encapsulated according to the communication protocol requirements of different execution units. For example, the raw material supply unit, drying unit, injection molding unit, mold temperature unit, post-processing unit, etc. all have their own communication protocols. Necessary protocol headers, checksums, and response mechanisms are added during the encapsulation process to ensure the accurate transmission and execution confirmation of the instructions. The checksum is an error detection mechanism that verifies whether an error has occurred during the transmission of the instruction by calculating a specific function value of the instruction content; the response mechanism requires the execution unit to return a confirmation message after receiving and executing the instruction so that the control system knows that the instruction has been correctly executed. In this way, a complete control strategy is formed, containing all the information from high-level decision-making to low-level execution, and achieving a seamless conversion from intelligent control decision-making to actual device operation.

[0062] In a factory that produces polypropylene medical device casings, this method is applied to automatically control the injection molding production line. First, the control parameter table is segmented, and 38 process parameters are divided into four major functional groups: The equipment preheating parameter group includes the temperature set values of 6 areas of the injection molding machine (from 185°C in the feeding area to 245°C in the nozzle area) and the temperature rise curve of the mold temperature controller; the raw material drying parameter group includes parameters such as a drying temperature of 80°C, a drying time of 4 hours, and an air flow rate of 3 m³ / min; the injection molding parameter group includes core process parameters such as an injection pressure of 80 MPa, an injection speed of 70 mm / s, a holding pressure of 60 MPa, and a holding time of 5 seconds; the cooling and molding parameter group includes parameters such as a cooling time of 20 seconds and a demolding temperature of 65°C. These parameters are organized into a linked list structure in chronological order, and each node contains the parameter values of this stage and the transfer conditions. Then, a decision tree model is constructed, and multiple key judgment nodes are set, such as "whether the barrel temperature reaches the set value ±2°C", "whether the moisture content of the raw material is lower than 0.02%", "whether the product weight is within the range of 23.5 ± 0.2 g", etc. According to these judgment results, a parameter adjustment plan is determined, such as "if the product weight is too light, increase the holding pressure by 2 MPa". All possible execution path combinations form a decision tree, covering various production situations. Then, the decision tree is converted into a state machine description, defining states such as "equipment standby", "preheating", "drying", "injection preparation", "injection", "holding pressure", "cooling", "ejecting", "product inspection", etc., as well as the conversion conditions and actions between states. For example, the conversion condition from "injection" to "holding pressure" is that "the screw advances to the set position", and the conversion action is "switch to holding pressure control". All states and conversion relationships are organized into a 25×18 state transition matrix, where the rows represent the current state, the columns represent the input conditions, and the matrix elements represent the target state and actions. Based on the state transition matrix, timing constraints are established, such as the minimum interval for injection pressure adjustment is 0.5 seconds, the waiting time for temperature adjustment is 60 seconds, the timeout for mold closing state confirmation is 10 seconds, etc. All instructions are arranged in chronological order to form an execution sequence list. The safety of the execution sequence is verified, and several risk points are detected, such as too fast temperature change rate, sudden change in injection pressure, etc., and protection logics such as gradual adjustment and intermediate buffer states are added. Finally, the instruction set after safety verification is translated into the control languages of different devices. For example, the injection molding machine instructions are translated into dedicated control codes, and the mold temperature controller instructions are translated into PID control parameters, and they are encapsulated according to the communication protocols of each device and CRC checksum and response requirements are added to form a complete control strategy. This solves the technical problem that the parameter settings of traditional control systems are fixed and lack adjustment according to real-time production conditions, and realizes intelligent adaptive control based on production status.

[0063] In a specific embodiment, the process of executing step S106 may specifically include the following steps: Collect device operation data and quality inspection data from each execution unit, record the deviation between the actual execution value and the set value of the control parameter through the online monitoring system, and at the same time obtain product size, appearance, strength, and weight indicators to obtain an execution feedback data set; Perform online analysis on the execution feedback data set, calculate the parameter fluctuation situation and the quality fluctuation trend through the sliding window technology, establish the correlation relationship between the parameter deviation and the quality change, and generate a process status evaluation report; Construct an incremental learning sample based on the process status evaluation report, package the parameter set value, actual execution value, and quality result of the current production batch as training samples, and add them to the model training data pool to form an updated training set; Input the updated training set into the model update engine, and perform parameter fine-tuning on the deep learning model through the transfer learning method, keeping the model skeleton structure unchanged and only updating the weight parameters to obtain an updated deep learning model; Perform verification testing on the updated deep learning model, evaluate the model performance through a mixed validation set of historical data and current data, calculate the prediction accuracy and generalization ability indicators, judge the effectiveness of the model update, and generate a model performance evaluation result; Dynamically adjust the learning strategy according to the model performance evaluation result, reduce the learning rate or roll back to the previous version when the model performance drops, confirm the update and deploy the updated deep learning model as the latest parameter optimization model when the model performance improves, forming a closed-loop adaptive control system.

[0064] Specifically, implementing closed-loop adaptive control is a step to solve the problem that traditional control systems lack the ability of adaptive learning and cannot automatically optimize control strategies based on the data accumulated during the production process. First, collect device operation data and quality inspection data from each execution unit to establish a complete feedback closed loop. In the injection molding machine unit, collect actual execution parameters such as screw position, injection pressure, cavity pressure, and injection speed through an embedded sensor network, compare them with the set values issued by the control strategy, and calculate the execution deviation. The deviation calculation uses the point-to-point comparison method, that is, calculate the difference between the actual value and the set value at each sampling moment to form deviation data in a time series. At the same time, collect multi-dimensional quality index data of the finished product at the product quality inspection station, including key quality parameters such as product size (measured by a coordinate measuring machine or a laser scanner), appearance (detect surface defects through a machine vision system), strength (obtained through mechanical property tests such as tensile tests and bending tests), and weight (measured by a precision electronic scale). These parameters are compared with the product quality standard to calculate the quality deviation. The device operation data and quality inspection data are associated according to the time stamp and product batch to form a structured execution feedback data set, which contains a complete set value - execution value - quality result triple data.

[0065] Perform online analysis on the execution feedback dataset to identify the correlation patterns between parameter fluctuations and quality changes. First, apply the sliding window technique to process time-series data. The sliding window is a data processing method that statistically analyzes the data within a fixed-size window by moving it over the data stream. In this method, set the window size to N product cycles (e.g., 30 cycles) and the window sliding step to M cycles (e.g., 5 cycles). Calculate statistical metrics such as mean, standard deviation, and trend slope for the data within each window. Through this processing, the original high-frequency sampled data is converted into more representative eigenvalue, reducing the impact of data noise. Then calculate the parameter fluctuation conditions, including indicators such as fluctuation amplitude (measured using standard deviation), fluctuation frequency (analyzing periodicity using Fourier transform), and fluctuation trend (calculating the slope using linear regression), to quantify the equipment execution stability. At the same time, analyze the quality fluctuation trend, including indicators such as quality mean change, quality dispersion degree, and defect rate change, to reflect the overall status and change direction of product quality. Next, establish the correlation relationship between parameter deviation and quality change. Use the lag correlation analysis method to calculate the correlation coefficient between parameter deviation and quality change at different time lags, and identify the parameter-quality pairs with significant influence relationships. Integrate all the analysis results to form a process status assessment report, which contains information such as the execution status of key parameters, quality change trend, and parameter-quality correlation strength, providing a basis for model update.

[0066] Construct incremental learning samples based on the process status assessment report to achieve continuous optimization of the model. Incremental learning is a machine learning method that allows the model to learn new knowledge based on existing knowledge without retraining the entire model. In this method, construct the data of the current production batch into training samples. Each sample contains three parts of information: parameter set value (e.g., the injection molding temperature is set to 240 °C), actual execution value (e.g., the actual temperature is 238 °C), and quality result (e.g., the product size deviation is +0.05 mm). The sample structure is designed in the form of a feature vector and a label pair. The feature vector contains the combination of the parameter set value and the actual execution value, and the label is the quality result. Add the constructed samples to the model training data pool to form an updated training set together with the historical samples. The training set management adopts the first-in-first-out (FIFO) strategy to maintain a fixed-size training window (e.g., the most recent 100,000 samples). When new samples are added, the earliest samples will be removed to ensure that the model reflects the latest production status. This incremental sample construction method overcomes the limitation of traditional machine learning that requires a large amount of offline data and realizes the online continuous learning ability of the model.

[0067] The updated training set is input into the model update engine to fine-tune the parameters of the deep learning model. Transfer learning is a method of applying a pre-trained model to a new domain. By retaining the basic structure and most of the parameters of the model and only fine-tuning specific layers, it greatly reduces the training data requirements and computational volume. In this method, the skeleton structure of the original deep learning model (i.e., the model combining convolutional neural network and long short-term memory network) remains unchanged, including structural parameters such as the number of convolutional layers, the size of convolutional kernels, the number of LSTM layers, and the number of neurons. Only the weight parameters are updated, such as the weights of convolutional kernels, the weights of LSTM gate units, and the weights of fully connected layers. The update uses the mini-batch gradient descent method. Each time, a mini-batch of samples (such as 32 samples) is randomly drawn from the training set, the error between the model prediction and the actual quality result is calculated, and then the gradient is calculated through the backpropagation algorithm to update the model weights. To maintain model stability, a small learning rate (such as 0.001) is adopted to avoid drastic changes in the model caused by a single batch of samples. During the model update process, the early stopping method is used to prevent overfitting, that is, when the validation set error no longer decreases for several consecutive iterations, the update process is stopped to obtain the updated deep learning model.

[0068] Perform a validation test on the updated deep learning model to evaluate the model's performance and generalization ability. The validation test uses a mixed validation set composed of historical production data and current production data. The historical data comes from production records of different periods and batches, which can test the model's memory ability for historical patterns; the current data are the samples produced recently, which can test the model's adaptability to new trends. The validation set is strictly separated from the training set to ensure the objectivity of the evaluation. The model performance evaluation includes multiple metrics: the prediction accuracy is measured using the mean squared error (MSE) or the mean absolute error (MAE), which reflects the average deviation between the model prediction and the actual quality; the generalization ability is evaluated using the cross-validation method, and the robustness of the model is reflected by the performance consistency on different subsets; the sensitivity analysis evaluates the sensitivity of the model to parameter fluctuations by adding small perturbations to the input features and observing the output changes. All evaluation metrics are integrated to form the model performance evaluation result, which provides a basis for subsequent adjustment of the learning strategy.

[0069] Dynamically adjust the learning strategy according to the model performance evaluation results to ensure the effectiveness and stability of model updates. The learning strategy adjustment is based on the performance change trend. When it is found that the model performance deteriorates (such as the prediction error increases and the generalization ability decreases), a conservative strategy is adopted: reduce the learning rate (such as from 0.001 to 0.0001) and decrease the amplitude of a single update; or roll back to the previous version, abandon the current update, and avoid the deterioration of the model performance. When the model performance improves (such as the prediction error decreases and the generalization ability increases), an aggressive strategy is adopted: confirm the update and deploy the updated deep learning model as the latest parameter optimization model; at the same time, appropriately increase the learning rate to accelerate the subsequent learning process. The model update and deployment adopt hot-switching technology, that is, smoothly transition to the new model without interrupting production to ensure the continuity of the control process. Through this dynamic learning strategy adjustment, a true closed-loop adaptive control system is formed, which can continuously optimize the control strategy according to the production feedback and solve the technical problem that the traditional control system lacks the adaptive learning ability.

[0070] The automatic control method for a plastic processing production line in the embodiments of the present application has been described above. Next, the automatic control system for a plastic processing production line in the embodiments of the present application will be described. Please refer to Figure 2 , an embodiment of the automatic control system for a plastic processing production line in the embodiments of the present application includes: A transmission module 201, configured to collect the process parameters and equipment status data of each station of the plastic processing production line, transmit the collected data to the central control system through a distributed sensing network, and generate a production status database; An analysis module 202, configured to perform multi-dimensional parameter correlation analysis based on the production status database, establish a mapping relationship between process parameters and product quality through a deep learning model that combines a convolutional neural network and a long short-term memory network. The deep learning model includes three convolutional layers for feature extraction, two LSTM layers for time series modeling, and two fully connected layers for parameter mapping to form a parameter optimization model; A generation module 203, configured to calculate the optimal control parameters for each station according to the parameter optimization model, generate a control strategy including an execution instruction sequence, and send the control strategy to each execution unit; An update module 204, configured to monitor the response results and control effects of the execution unit, and perform real-time update on the parameter optimization model to form a closed-loop adaptive control system.

[0071] Through the collaborative cooperation of the above-mentioned various components, by collecting the process parameters and equipment status data of each work station, and combining with the distributed sensor network technology, the integration and transmission of the whole production process data are realized, effectively solving the technical problems of the lack of in-depth collaborative control between various equipment and the fragmentation of control in each link in the traditional plastic processing production line, and providing a data basis for global optimization decision-making. Based on the production status database, multi-dimensional parameter correlation analysis is performed, and a deep learning model combining convolutional neural network and long short-term memory network is applied to realize the simultaneous extraction and processing of the spatial features and temporal features of the process parameters, breaking through the technical bottleneck that the traditional single model cannot comprehensively capture the complex correlations of the parameters, and establishing a more accurate mapping relationship between the process parameters and the product quality. The three-layer convolutional layer structure design adopted in the present invention enables the model to gradually extract the spatial correlation features between the process parameters from low dimension to high dimension, adapting to the characteristics of multi-parameter interaction in plastic processing; the introduction of two LSTM layers effectively captures the long-term and short-term impacts of the parameter changes over time on the product quality, solving the problem that the traditional control methods are difficult to handle the temporal dependencies; the optimized design of the two fully connected layers significantly improves the stability and generalization ability of the model in industrial data processing by applying technical means such as batch normalization and Dropout. According to the parameter optimization model, the optimal control parameters of each work station are calculated and an execution instruction sequence is generated, replacing the subjectivity and uncertainty of the traditional control methods relying on manual experience, and realizing data-driven scientific decision-making. The present invention monitors the response results and control effects of the execution unit to update the parameter optimization model in real time, forming a closed-loop adaptive control system, overcoming the limitation of the traditional control system lacking self-learning ability and endowing the control system with the ability of continuous optimization. Especially in the specific application field of plastic processing, the deep learning algorithm of the present invention fully considers the industry characteristics, specifically designs a network structure suitable for processing the interaction relationships of parameters such as temperature, pressure, and time, accurately models the complex non-linear relationships between the material properties, process conditions and product quality. The contribution of the algorithm features to the solution is prominently reflected in its ability to automatically extract the key patterns affecting the quality of plastic products from the massive production data, and optimize the parameters based on these patterns, avoiding the blindness and inefficiency of parameter adjustment in the traditional methods, significantly improving the automation level of plastic processing and the stability of product quality, while reducing energy consumption and raw material waste.

[0072] Above Figure 2 The automatic control system for a plastic processing production line in the embodiments of the present invention is described in detail from the perspective of modular functional entities. Next, the automatic control equipment for a plastic processing production line in the embodiments of the present invention is described in detail from the perspective of hardware processing.

[0073] Figure 3FIG. 0 is a schematic structural diagram of an automatic control device for a plastic processing production line provided by an embodiment of the present invention. The automatic control device 300 for the plastic processing production line may vary greatly due to different configurations or performances, and may include one or more processors (central processing units, CPUs) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 for storing application programs 333 or data 332 (for example, one or more mass storage device terminals). Among them, the memory 320 and the storage media 330 may be transient storage or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations for the automatic control device 300 for the plastic processing production line. Further, the processor 310 may be configured to communicate with the storage media 330 and execute a series of instruction operations in the storage media 330 on the automatic control device 300 for the plastic processing production line to implement the steps of the above-mentioned automatic control method for the plastic processing production line.

[0074] The automatic control device 300 for the plastic processing production line may further include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, and so on. Those skilled in the art can understand that Figure 3 The shown structure of the automatic control device for the plastic processing production line does not limit the automatic control device for the plastic processing production line provided by the present invention, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0075] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions run on a computer, the computer is caused to execute the steps of the automatic control method for the plastic processing production line.

[0076] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described systems, systems, and units may refer to the corresponding processes in the foregoing method embodiments and will not be described herein again.

[0077] When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable an automatic control device for a plastic processing production line (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs.

[0078] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of various embodiments of the present invention.

Claims

1. An automatic control method for a plastic processing production line, characterized in that, The method includes: Collecting the process parameters and equipment status data of each station on the plastic processing production line, and transmitting the collected data to the central control system through a distributed sensing network to generate a production status database; Based on the production status database, performing multi-dimensional parameter correlation analysis, and establishing a mapping relationship between process parameters and product quality through a deep learning model that combines a convolutional neural network and a long short-term memory network. The deep learning model includes three convolutional layers for feature extraction, two LSTM layers for time series modeling, and two fully connected layers for parameter mapping, forming a parameter optimization model; According to the parameter optimization model, calculating the optimal control parameters for each station, generating a control strategy including an execution instruction sequence, and sending the control strategy to each execution unit; Monitoring the response results and control effects of the execution units, and performing real-time updates on the parameter optimization model to form a closed-loop adaptive control system.

2. The automatic control method for a plastic processing production line according to claim 1, characterized in that The step of collecting the process parameters and equipment status data of each station on the plastic processing production line, and transmitting the collected data to the central control system through a distributed sensing network to generate a production status database includes: Performing RFID tag identification on plastic raw materials, collecting raw material types, batch numbers, and supplier information, and performing preliminary processing on the collected information through edge computing nodes to obtain raw material basic data; Setting a multi-point temperature and humidity sensor and a weighing sensor array inside the raw material warehouse to monitor the raw material storage environment parameters and inventory in real time to obtain raw material status data; Monitoring the operating status of the injection molding machine, collecting parameters such as melting temperature, injection pressure, holding pressure time, mold temperature, cooling time, injection speed, clamping force, screw speed, and back pressure value, and transmitting them to the data acquisition unit through an industrial bus to obtain injection molding process data; Collecting the energy consumption data of production equipment, and measuring current, voltage, power factor, active power, reactive power, and harmonic content through power parameter sensors installed on each equipment to obtain energy consumption data; Aligning the raw material basic data, raw material status data, injection molding process data, and energy consumption data through timestamps, establishing a multi-time scale data synchronization matrix, and obtaining a correlation data set; Performing data cleaning and outlier detection on the correlation data set, removing noise and outliers through wavelet transform and statistical filtering algorithms, and storing the processed data in a structured database to form the production status database.

3. The automatic control method for a plastic processing production line according to claim 1, characterized in that, The step of based on the production status database, performing multi-dimensional parameter correlation analysis, and establishing a mapping relationship between process parameters and product quality through a deep learning model that combines a convolutional neural network and a long short-term memory network. The deep learning model includes three convolutional layers for feature extraction, two LSTM layers for time series modeling, and two fully connected layers for parameter mapping, forming a parameter optimization model includes: Extracting historical production data from the production status database, classifying it according to product models and raw material batches, and reducing the data dimension through the principal component analysis algorithm to obtain a feature data set; Perform a correlation analysis on the feature dataset, calculate the association strength between parameters through the Pearson correlation coefficient and mutual information calculation methods, screen out the key process parameters affecting product quality, and form a key parameter set; Based on the key parameter set, construct a convolutional neural network module, set three convolutional layers, each layer using 64, 128, and 256 3×3 convolutional kernels respectively, extract features from the process parameter sequence, introduce non-linear transformation through the ReLU activation function, and obtain a feature mapping matrix; Input the feature mapping matrix into a long short-term memory network module, model the parameter time series relationship through two LSTM layers each containing 128 neurons, and combine the forget gate, input gate, and output gate mechanisms to handle long-term dependencies, obtaining a time series feature vector; Process the time series feature vector through two fully connected layers. The first layer contains 256 neurons, and the second layer contains 128 neurons. Use the Dropout technique to prevent overfitting and improve training stability through batch normalization, obtaining an optimized deep learning model; Perform a sensitivity analysis on the optimized deep learning model, calculate the output change rate by perturbing the input parameters, construct a parameter importance ranking matrix, and combine with an expert rule system to form the parameter optimization model.

4. The automatic control method for a plastic processing production line according to claim 3, wherein, The processing of the time series feature vector through two fully connected layers, where the first layer contains 256 neurons and the second layer contains 128 neurons, using the Dropout technique to prevent overfitting and improving training stability through batch normalization, obtaining an optimized deep learning model, includes: Perform data standardization processing on the time series feature vector, calculate the mean and standard deviation of each dimension feature through the Z-score standardization method, and convert the feature data into a standard normal distribution with a mean of 0 and a standard deviation of 1, obtaining a standardized feature vector; Input the standardized feature vector into the first fully connected layer, perform a linear transformation on the input features through 256 neurons, calculate the weighted sum using the weight matrix and the input vector, and obtain the original output value of the first layer; Perform batch normalization processing on the original output value of the first layer, adjust the feature distribution by calculating the mean and variance of the samples within the mini-batch, and introduce learnable scaling parameter γ and translation parameter β to obtain the normalized output value; Apply the LeakyReLU activation function to the normalized output value, keep the positive values unchanged, multiply the negative values by a slope coefficient of 0.01, introduce non-linear transformation, and apply Dropout processing with a rate of 0.5 to randomly turn off 50% of the neurons, obtaining the final output of the first layer; Use the final output of the first layer as the input of the second fully connected layer, perform a linear transformation through 128 neurons, perform batch normalization and LeakyReLU activation, and apply Dropout processing with a rate of 0.3 to obtain the final feature representation; Add a residual connection to the final feature representation, perform an element-wise addition of the temporal feature vector after dimensional transformation and the final feature representation, and calculate the probability distribution of each category through the Softmax function to obtain the optimized deep learning model.

5. The automatic control method for a plastic processing production line according to claim 1, characterized in that, The optimizing the model according to the parameters, calculating the optimal control parameters for each station, and generating a control strategy including an execution instruction sequence, and sending the control strategy to each execution unit includes: Construct an objective function based on the parameter optimization model, assign different weight values to product quality indicators, production efficiency indicators, and energy consumption indicators, and establish a multi-objective optimization mathematical model through linear weighted combination to obtain an optimization problem expression; Set constraint conditions for the optimization problem expression, including the upper and lower limits of injection temperature, injection pressure range, mold temperature range, holding pressure time range, and cooling time range, and describe the process boundary through linear inequalities to obtain a set of constraint conditions; Input the optimization problem expression and the set of constraint conditions into the solution module, search the parameter space through an iterative calculation method, select multiple candidate parameter combinations in each iteration, calculate the objective function value, and retain the parameter combination with a smaller objective function value to enter the next iteration. After multiple iterations, obtain the initial control parameters for each station; Perform a sensitivity test on the initial control parameters of each station. By making positive and negative changes to the parameter values, calculate the product quality change rate. When the change rate exceeds the set threshold, mark the parameter as a highly sensitive parameter, set a finer control range for the highly sensitive parameter, and generate a control parameter table considering stability; Compile an instruction sequence based on the control parameter table, divide the control parameters into a startup segment, a production segment, and a shutdown segment according to the process stage. Each stage includes parameter setting values, execution time sequences, and logical judgment conditions, and organize them in the order of execution to form the control strategy; Distribute the control strategy to each execution unit, including the raw material supply unit, the drying unit, the injection molding unit, the mold temperature unit, and the post-treatment unit, through the industrial communication bus, send a policy data packet and receive an acknowledgment signal to confirm that the control strategy has been successfully sent to each execution unit.

6. The automatic control method for a plastic processing production line according to claim 5, characterized in that The compiling an instruction sequence based on the control parameter table, dividing the control parameters into a startup segment, a production segment, and a shutdown segment according to the process stage. Each stage includes parameter setting values, execution time sequences, and logical judgment conditions, and organize them in the order of execution to form the control strategy includes: Perform a segmentation process on the control parameter table, divide the parameters into an equipment preheating parameter group, a raw material drying parameter group, an injection molding parameter group, and a cooling molding parameter group according to the process flow, and establish a parameter linked list structure according to the production process to obtain a segmented parameter set; Construct a decision tree model for the segmented parameter set, set equipment status judgment nodes, material status judgment nodes, and quality inspection judgment nodes, and determine the parameter adjustment direction and adjustment amplitude according to the judgment results to generate a multi-branch execution path; Convert the multi-branch execution path into a state machine description, define the initial state, transition conditions, target state, and transition actions, and establish a complete process flow chart using the finite state machine representation method to obtain a state transition matrix; Based on the state transition matrix, establish timing constraint relationships, set the time interval for parameter changes, the waiting time for parameter responses, and the timeout for status confirmation, and generate an execution sequence list containing timestamps; Perform security verification on the execution sequence list, check the parameter change rate, state jump amplitude, and device response ability, identify potential risk points and add protection logic to obtain a set of instructions after security verification; Translate the set of instructions after security verification into device control language, encapsulate them according to the communication protocol requirements of different execution units, and add a checksum and response mechanism to form the control strategy.

7. The automatic control method for a plastic processing production line according to claim 1, characterized in that Monitor the response results and control effects of the execution units, and update the parameter optimization model in real time to form a closed-loop adaptive control system, including: Collect device operation data and quality inspection data from each execution unit, record the deviation between the actual execution value and the set value of the control parameter through an online monitoring system, and at the same time obtain product size, appearance, strength, and weight indicators to obtain an execution feedback data set; Perform online analysis on the execution feedback data set, calculate the parameter fluctuation situation and quality fluctuation trend through the sliding window technology, establish the correlation relationship between parameter deviation and quality change, and generate a process status evaluation report; Construct incremental learning samples based on the process status evaluation report, package the parameter set value, actual execution value, and quality result of the current production batch as training samples, and add them to the model training data pool to form an updated training set; Input the updated training set into the model update engine, perform parameter fine-tuning on the deep learning model through the transfer learning method, keep the model skeleton structure unchanged, and only update the weight parameters to obtain an updated deep learning model; Perform verification testing on the updated deep learning model, evaluate the model performance through a mixed validation set of historical data and current data, calculate the prediction accuracy and generalization ability indicators, judge the effectiveness of the model update, and generate a model performance evaluation result; Dynamically adjust the learning strategy according to the model performance evaluation result, reduce the learning rate or roll back to the previous version when the model performance decreases, confirm the update and deploy the updated deep learning model as the latest parameter optimization model when the model performance improves, to form the closed-loop adaptive control system.

8. An automatic control system for a plastic processing production line, characterized in that, For implementing the automatic control method for a plastic processing production line as described in any one of claims 1-7, the automatic control system for a plastic processing production line includes: A transmission module, configured to collect process parameters and device status data of each station of the plastic processing production line, transmit the collected data to the central control system through a distributed sensor network, and generate a production status database; An analysis module, configured to perform multi-dimensional parameter correlation analysis based on the production status database, and establish a mapping relationship between process parameters and product quality through a deep learning model that combines a convolutional neural network and a long short-term memory network. The deep learning model includes three convolutional layers for feature extraction, two LSTM layers for timing modeling, and two fully connected layers for parameter mapping, to form a parameter optimization model; A generation module, configured to optimize the parameter optimization model according to the parameters, calculate the optimal control parameters of each station, generate a control strategy including an execution instruction sequence, and send the control strategy to each execution unit; An update module, configured to monitor the response results and control effects of the execution units, and perform real-time update on the parameter optimization model to form a closed-loop adaptive control system.

9. An automatic control device for a plastic processing production line, characterized in that, It includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements the automatic control method for a plastic processing production line according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program runs on the processor, it causes the processor to execute the automatic control method for a plastic processing production line according to any one of claims 1 to 7.

Images