Production debugging control method and system for plastic container

Abstract

The invention relates to the technical field of production debugging control, and discloses a production debugging control method and system for a plastic container. The method comprises the following steps: arranging a plurality of different sensors in plastic container blow molding equipment, and simultaneously collecting blow molding process parameter data; performing wavelet threshold denoising and anomaly identification on the blow molding process parameter data to obtain a process parameter anomaly identification result; establishing a transfer function matrix between the process parameters and quality indexes according to the process parameter anomaly identification result, and obtaining a target process parameter set through multi-target optimization; and inputting the target process parameter set into a double-integral enhanced recurrent neural network for parameter regulation and control calculation to obtain a process parameter adjustment amount. According to the method, early detection and accurate positioning of process abnormity in the blow molding process are realized, the quality fluctuation risk and the defective product rate are greatly reduced, and the technical problem that different response characteristic parameters are difficult to coordinate is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of production debugging control, and particularly relates to a production debugging control method and system for plastic containers. Background Art

[0002] With the growth in the demand for ultra-clean containers and plastic containers for containing dangerous goods, the requirements for the consistency, sealing performance, and cleanliness of product quality are continuously increasing. However, the detection, debugging, and process control of traditional blow molding equipment mainly rely on the setting of empirical parameters and simple PID control. This method exposes many deficiencies in the production process of high-end products such as HDPE ultra-clean barrels. In the prior art, the setting of blow molding process parameters often only considers the ideal state, lacking effective responses to interference factors such as temperature fluctuations, pressure changes, and material batch differences in the actual production environment, resulting in large fluctuations in product quality, low qualification rates, and poor consistency.

[0003] In terms of the data analysis of the detection and debugging of blow molding equipment, there are currently common problems in the industry such as few data acquisition points, low acquisition frequencies, and limited sensor accuracies, which cannot comprehensively reflect the dynamic change characteristics during the blow molding process. Especially for multi-layer forming technology and high-precision filtration blow molding processes, due to the complex coupling relationships and non-linear characteristics between parameters, traditional single-parameter analysis methods are difficult to accurately identify the root causes of abnormal working conditions and quality fluctuations, resulting in long debugging cycles and low efficiency. Summary of the Invention

[0004] The main object of the present invention is to provide a production debugging control method and system for plastic containers. The present invention realizes the early detection and accurate positioning of process anomalies during blow molding, greatly reduces the risk of quality fluctuations and the defective product rate, and solves the technical problem of difficult coordination of different response characteristic parameters.

[0005] To achieve the above object, the present invention provides a production debugging control method for plastic containers, including the following steps: Arrange a plurality of different sensors in the plastic container blow molding equipment to simultaneously collect blow molding process parameter data; Perform wavelet threshold denoising and anomaly identification on the blow molding process parameter data to obtain a process parameter anomaly identification result; Establish a transfer function matrix between process parameters and quality indicators according to the process parameter anomaly identification result, and obtain a target process parameter set through multi-objective optimization; Input the target process parameter set into a double-integral enhanced recursive neural network for parameter regulation calculation to obtain a process parameter adjustment amount.

[0006] The present invention also provides a production debugging control system for plastic containers, including: A collection module, which is used to arrange a plurality of different sensors in a plastic container blow molding device and collect blow molding process parameter data simultaneously; An anomaly recognition module, which is used to perform wavelet threshold denoising and anomaly recognition on the blow molding process parameter data to obtain a process parameter anomaly recognition result; A multi-objective optimization module, which is used to establish a transfer function matrix between process parameters and quality indicators according to the process parameter anomaly recognition result, and obtain a set of target process parameters through multi-objective optimization; A parameter regulation calculation module, which is used to input the set of target process parameters into a double-integral enhanced recurrent neural network for parameter regulation calculation to obtain a process parameter adjustment amount.

[0007] In summary, the technical solution provided by the present invention is based on a multi-point distributed sensor network and wavelet threshold denoising technology. The system realizes the acquisition and preprocessing of all parameters, high-precision, and multi-dimensional data in the blow molding process, effectively improving the quality of the original data and laying a foundation for subsequent accurate analysis. In particular, the pressure sensor array arranged on the inner wall of the mold cavity realizes the real-time high-precision monitoring of the plastic flow and filling process for the first time. The anomaly recognition method combined with the K-means clustering technology breaks through the limitation of traditional single-parameter analysis. Through multi-dimensional feature comprehensive evaluation, the mapping relationship between process parameters and product quality is established, realizing the early detection and accurate positioning of process anomalies in the blow molding process, and greatly reducing the quality fluctuation risk and the defective product rate. The introduced double-integral enhanced recurrent neural network control model simultaneously considers the immediate deviation, cumulative deviation, and second-order cumulative deviation through the double-integral error function, significantly enhancing the system's ability to suppress continuous deviation and slow-varying interference, and solving the problem of insufficient adaptability of traditional PID control in the face of interference factors such as temperature fluctuations, pressure changes, and material batch differences. The multi-time domain cooperative control strategy realizes the cooperative optimization of the temperature control group, pressure control group, time control group, and mechanical control group through parameter grouping and decentralized coordination algorithms, and solves the technical problem of difficult coordination of parameters with different response characteristics. Description of the Drawings

[0008] Figure 1 is a schematic diagram of the steps of a production debugging control method for a plastic container in an embodiment of the present invention; Figure 2 is a structural block diagram of a production debugging control system for a plastic container in an embodiment of the present invention.

[0009] The realization, functional characteristics, and advantages of the object of the present invention will be further described with reference to the embodiments and the accompanying drawings. Detailed Embodiments

[0010] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0011] Referring to Figure 1 , this embodiment provides a production debugging control method for plastic containers, including the following steps: S1. Arrange a plurality of different sensors in the plastic container blow molding equipment to simultaneously collect the blow molding process parameter data; Among them, install platinum resistance temperature sensors in the barrel area of the blow molding equipment. These sensors are respectively arranged in the feeding section, compression section, homogenization section, metering section and head section. By using their high-precision temperature measurement ability, the temperature changes of the blank in different processing stages are monitored in real time to obtain the blank temperature distribution data, which characterizes the temperature uniformity and melting state of the raw material during the heating process. Arrange high-precision pressure sensors on the inner wall of the mold cavity. These sensors are respectively distributed at the top, middle and bottom of the cavity to continuously monitor the pressure distribution in the cavity and generate the cavity pressure distribution data, which reflects the flow behavior and filling state of the molten plastic in the cavity and provides data basis for optimizing the mold design and molding process. At the same time, install high-precision pressure sensors on the air path of the blowing system. These sensors are used to monitor the pressure changes during the blowing process and record the blowing pressure change curve in real time, which characterizes the expansion effect of the gas on the molten plastic during the blow molding process and reflects the stability of the blowing pressure during the molding process, so as to ensure the uniformity and consistency of the blow molding. At the same time, install high-resolution displacement sensors on the mold clamping mechanism to monitor the mold clamping accuracy and the dynamic deformation of the mold. These sensors measure the deformation degree of the mold under high pressure conditions and record the mold dynamic deformation data, so as to ensure that the mold maintains good accuracy and stability during long-term use. All these sensors are synchronously collected through a data acquisition module, and a high-precision time synchronization technology is adopted to ensure the time stamp consistency of all data, forming a blow molding process data set. Perform data standardization processing on the collected blank temperature distribution data, cavity pressure distribution data, blowing pressure change curve and mold dynamic deformation data to eliminate the dimension difference and measurement error between different sensors. At the same time, use a multi-dimensional data fusion method to integrate these data to form blow molding process parameter data with complete time series and spatial distribution characteristics.

[0012] S2. Perform wavelet threshold denoising and anomaly identification on the blow molding process parameter data to obtain the process parameter anomaly identification result; Specifically, wavelet transform processing is performed on the blow molding process parameter data, and wavelet threshold denoising technology is used to eliminate measurement noise to improve the signal quality. During the denoising process, the db4 wavelet basis function is used, and a 5-layer decomposition structure is selected to ensure the multi-scale decomposition accuracy of the data. At the same time, the soft threshold method is used to determine the denoising threshold, so as to effectively remove high-frequency noise and retain the key information in the data. The parameter data after wavelet denoising is smoother and more real, and can more accurately reflect the dynamic changes of temperature, pressure and mold deformation during the blow molding process. For the denoised blank temperature distribution data, in order to ensure the consistency of data in different batches, time domain synchronization processing is performed on the data, and normalization is carried out using the forming cycle, standardizing the temperature curves in all preforming stages to the same time interval, and eliminating the time offset problem caused by differences in equipment startup times. After obtaining the time domain synchronized temperature curve, piecewise linear fitting is performed on it, and the temperature gradient and temperature uniformity index are calculated. The former reflects the temperature change rate of the blank at different stages, and the latter measures the uniformity of the overall heating of the blank, so as to extract the key indicators that can characterize the thermal distribution characteristics of the preforming stage. In order to analyze the flow characteristics of the molten plastic in the mold cavity, for the denoised cavity pressure distribution data, the principal component analysis method is used for dimensionality reduction processing, and the most representative pressure distribution characteristic data are extracted. These characteristics can describe the change trend of the pressure at different positions in the cavity and reflect the uniformity and pressure fluctuation of the melt during the filling process. At the same time, for the denoised blowing pressure change curve, a four-parameter equation is used for feature extraction, including features such as pressure rise rate, steady pressure time, pressure release rate, and residual pressure. These parameters effectively describe the change mode of the air pressure during the blow molding process and evaluate the impact of the blowing process on the molding quality. For the mold dynamic deformation data, since the mold is subjected to the dual action of temperature and pressure during the blow molding process, time-frequency analysis is performed on this data to extract its dynamic characteristics. Short-time Fourier transform or wavelet transform is used to perform time-frequency decomposition on the signal, and frequency domain characteristics such as frequency band energy ratio, main frequency point, and spectral peak value are extracted from it. These characteristics characterize the dynamic changes of the mold deformation and help to identify abnormal deformation modes. The temperature gradient and temperature uniformity index characteristics, pressure distribution characteristic data, blowing pressure characteristic data, and mold dynamic characteristic data are combined to construct an objective feature set, which fully characterizes the change trends of key process parameters such as temperature, pressure, and structural deformation during the blow molding process. In order to identify abnormal process parameter patterns, based on the objective feature set, the K-means clustering algorithm is used to classify the feature data, and the number of clusters is set to 5 to distinguish different process state categories. Subsequently, the density peak clustering algorithm is used to automatically detect abnormal patterns, and the mapping relationship between abnormal patterns and defect types is constructed in combination with historical data. Combining the analysis of abnormal patterns of process parameters, the abnormal identification results of process parameters are obtained.

[0013] Statistical processing is performed on each characteristic parameter in the target feature set to calculate its mean and standard deviation under normal operating conditions, and a process parameter standard space model is established. This model is used to define the normal fluctuation range of each characteristic parameter, with μ as the mean and σ as the standard deviation. Accordingly, the normal interval of the characteristic parameter is determined as [μ−3σ, μ+3σ]. This interval covers most normal production samples statistically, thus providing a quantitative benchmark for anomaly identification. Based on the standard space model, the K-means clustering algorithm is used to perform clustering analysis on the target feature set. The feature space is partitioned by minimizing the Euclidean distance from the sample points to the cluster centers, and the number of clusters is determined to be 5 according to evaluation indicators such as the silhouette coefficient, representing five different quality grade regions of super excellent, excellent, qualified, critical, and unqualified. The clustering results form a process parameter distribution structure, enabling samples in different quality states to obtain clear attribution in the feature space. After completing the quality region partitioning, an elliptical influence function is designed for each characteristic parameter, and its mathematical form is ER(d)=exp(−β·d² / R²), where d represents the Euclidean distance between the eigenvalue and the belonging cluster center, R is the influence radius of this characteristic parameter on the target quality index, and β is the influence attenuation coefficient. This function form can perform a non-linear measurement of the degree of feature deviation. The farther the distance, the greater the influence, and the lower the function value. By calculating all characteristic parameters through this influence function, the specific influence degree value of each feature on quality fluctuation in its belonging quality region is obtained. To integrate the relationship between multi-dimensional characteristic parameters and their influence degrees, principal component analysis is performed on the target feature set and its corresponding influence degree values to extract the main feature direction, and at the same time, the feature weight matrix is calculated. This matrix represents the contribution ratio of each characteristic parameter in different quality dimensions. Through principal component dimensionality reduction, redundant information is effectively eliminated, the main influencing factors are highlighted, and the feature weights are further corrected through optimization methods such as genetic algorithms or covariance analysis to ensure the scientificity and robustness of weight allocation. According to the obtained feature weight matrix and influence degree values, a weighted summation method is used for comprehensive quality evaluation, and the comprehensive quality index value QI is calculated. The formula form is QI=∑(wᵢ·ERᵢ), where wᵢ is the weight of the i-th characteristic parameter in the weight matrix, and ERᵢ is the output value of the influence function corresponding to this parameter. This comprehensive quality index value is in the interval [0,1]. The closer the value is to 1, the more stable the overall process and the better the product quality. Otherwise, there are process deviations or anomalies. The comprehensive quality index values of each sample in the target feature set are compared with the corresponding normal quality range in the standard space model to identify samples whose comprehensive quality index values deviate significantly from the normal interval. Combining their specific characteristic parameter combinations and referring to their belonging categories in the clustering region, it is further determined whether the sample belongs to an abnormal operating condition. If certain feature combinations cause a significant decrease in the quality index and deviate from the center in multiple quality dimensions, it is confirmed as a process parameter anomaly. Output the process parameter anomaly identification result.

[0014] S3. Establish a transfer function matrix between process parameters and quality indicators according to the recognition results of process parameter anomalies, and obtain a set of target process parameters through multi-objective optimization; It should be noted that based on the identified abnormal parameter combinations and the corresponding quality fluctuations, a systematic classification and evaluation of the five key quality indicators of the product are carried out. The five quality indicators are wall thickness uniformity, sealing performance, dimensional accuracy, surface quality, and cleanliness, which together constitute a multi-dimensional quality indicator system, and describe the product performance from the perspectives of structural stability, functional reliability, manufacturing accuracy, appearance consistency, and clean manufacturing. By comparing and analyzing the parameter deviation data in a large number of abnormal identification results with the actual product quality evaluation results, the preliminary influence relationships of various process parameters on different quality indicators are statistically obtained, forming a preliminary influence relationship table, which reveals the positive and negative effect trends of each parameter on the quality attributes of the product within different value ranges. Based on the preliminary influence relationship table, an orthogonal experimental scheme with good coverage is designed to combine and match the key process parameters, organize multiple batches of blow molding production tests, and collect the corresponding quality inspection data after each batch is completed to form a parameter-quality experimental data set. This experimental data set includes the set values of process parameters, the corresponding production batch information, and quality inspection results, and has been normalized for efficient training of subsequent modeling algorithms. After having a complete and high-quality data basis, the Gaussian process regression algorithm is introduced, and a mathematical mapping model between each process parameter and each quality indicator is established based on the Bayesian framework. This algorithm automatically constructs a prediction function containing non-linear relationships and uncertainty expressions through learning the experimental data, and outputs a mean function and a confidence interval for each parameter-index combination, constituting a transfer function matrix between process parameters and quality indicators. Each element in this matrix is a trained regression model, reflecting the contribution mechanism and sensitivity of a certain process parameter to a specific quality attribute. Based on the transfer function matrix, when optimizing the process parameter configuration, a multi-objective optimization model with the five quality indicators as the optimization objectives is constructed, and the five quality objectives are weighted and allocated through the analytic hierarchy process. The determination of the weights takes into account comprehensive factors such as product performance requirements, customer application environment, and quality cost ratio, and ensures the rationality of the judgment matrix through consistency testing. The multiple objective functions are weighted and fused using these weight coefficients to obtain a weighted comprehensive objective function, and the maximization of the function value corresponds to the optimal state of the comprehensive quality performance. To solve this objective function, the particle swarm optimization algorithm is used for global search. Utilizing its characteristics of population collaborative evolution, the positions and velocities of the particles are continuously updated in successive generations of iterations, making them tend towards the direction of better solutions. At the same time, a local search strategy is combined to improve the convergence speed and accuracy. During the iteration process, the algorithm evaluates the fitness value of the current particle in real time, determines whether it meets the precision tolerance or reaches the maximum number of iterations, and outputs the optimal solution, that is, the target process parameter set, after meeting the termination conditions. The target process parameter set includes key set values such as preheating temperature, blowing pressure, holding pressure time, cooling time, and screw speed.

[0015] S4. Input the target process parameter set into the dual-integral enhanced recurrent neural network for parameter regulation calculation to obtain the process parameter adjustment amount.

[0016] Specifically, the measured values of process parameters collected in real time during the current production process are used to calculate the differences from the target process parameter set, obtaining the deviation values in each parameter dimension, which constitute the process parameter deviation data and reflect the gap between the current working condition and the optimal setting. The process parameter deviation data is input into the input layer of the neural network. The number of nodes in this input layer exactly matches the dimension of the target process parameters. That is, if the target parameter set contains fifteen key process variables, the input layer should contain fifteen nodes, so as to ensure that the deviation data can be completely expressed in the network and form the target input vector. The target input vector is input into the first hidden layer and the second hidden layer of the network for non-linear transformation processing respectively. During this process, the neural network internally maps the data through the Tanh activation function, extracts the feature patterns in the deviation input, and converts them into hidden layer feature representations with high-dimensional expression capabilities. These hidden features retain the dominant information in the original deviation signal and simultaneously automatically extract the coupling relationships between key control factors during the network parameter training process. The core feature of the double-integral enhanced recurrent neural network lies in the introduction of a recurrent connection mechanism in its structure, that is, the hidden state at the current moment depends not only on the current input but also on the historical hidden states at the previous moment and the previous two moments. The network internally models these time correlations through weight matrices and bias terms to form a dynamic memory path. With the help of this recurrent mechanism, the network automatically captures the changing trend of process parameters over time, accurately identifies the cumulative effects of process disturbances, and generates a state vector reflecting the current overall state of the system. Based on the obtained state vector, the original deviation data is enhanced through the double-integral error function. This function not only considers the immediate deviation term but also integrates the first-order integral term and the second-order integral term of the error, that is, respectively reflecting the cumulative deviation and the error growth trend, enabling the network to have stronger perception and control capabilities for long-term offsets, continuous fluctuations, and trend changes. This error enhancement mechanism is adjusted by two sets of integral coefficients, and the optimal coefficients are determined through the particle swarm algorithm or other optimization methods during the training stage to ensure the positive enhancement effect of error processing on the control result. The enhanced error vector and the state vector are input into the output layer of the network, and the final adjustment value of each process parameter is generated through a linear activation function to form the process parameter adjustment amount. When performing the control operation in the output layer, the network integrates an anti-interference compensation mechanism to identify and compensate for uncontrollable external interference factors such as temperature fluctuations, pressure disturbances, and raw material batch differences in real time. This compensation mechanism operates jointly based on a state observer and a feedforward compensation path. After the state observer evaluates the disturbance intensity, the output result is adjusted through a compensation matrix to prevent problems such as over-regulation or regulation lag in the system, thereby ensuring the accuracy and stability of the control response. The finally output process parameter adjustment amount is fed back to the device execution unit to drive components such as heaters, pressure valves, screw motors, and cooling systems to perform coordinated adjustments, realizing the refined dynamic control of the blow molding process, effectively resisting the influence of external disturbances, and improving the forming consistency and production yield.

[0017] According to the process parameter adjustment amount output by the double-integral enhanced recurrent neural network, all parameters are divided into four functional control groups, namely the temperature control group, the pressure control group, the time control group, and the mechanical control group, according to the physical properties of the parameters, the control mechanism, and the type of influence on the forming process, thus forming a parameter grouping structure. The temperature control group includes the heating temperatures of each section and the mold temperature. The pressure control group covers the blowing pressure and the holding pressure. The time control group contains time-sequence related parameters such as the holding time and the cooling time. The mechanical control group involves structural motion parameters such as the screw speed and the mold-closing displacement. This grouping method helps to achieve hierarchical control and strategy coordination in the subsequent control architecture. After constructing the parameter grouping structure, clarify the coupling relationship between each control group. Taking all process parameters as the object, based on the change trend of the adjustment amount and the historical co-regulation data, calculate the correlation coefficient between the parameters, and construct a parameter coupling degree matrix accordingly. This matrix quantitatively describes the degree of mutual dependence between the parameters by analyzing the numerical linkage between any two parameters, the coupling strength of the process response, and the synergistic influence in the final product quality performance. Based on this matrix, identify which parameters have strong positive or negative correlation relationships, thereby extracting the interaction strength between different control groups, and summarizing to form a coordination requirement table between groups. This requirement table provides a quantitative basis for the synchronization setting and coordination priority determination of subsequent control strategies, ensuring that multiple groups of parameters can be synchronously adjusted or delayed according to their coupling tightness within the same control cycle. Based on the coordination requirement table between groups, construct a three-layer control architecture, which includes a bottom-layer execution control layer, a middle-layer strategy coordination layer, and a top-layer optimization decision layer respectively. The bottom-layer control layer is directly connected to the actual execution mechanism, and performs real-time control on devices such as temperature regulators, pressure valves, drive motors, and timing modules, with the characteristics of fast response speed and high adjustment accuracy. The middle-layer strategy coordination layer manages the centralized control within each parameter group and the synchronous scheduling between groups based on the parameter coupling information and the coordination requirement table, and adopts a decentralized coordination algorithm based on the event-triggering mechanism. When a certain parameter change or the control boundary is approached, the linkage adjustment of the relevant control path is immediately started. The top-layer optimization decision layer is based on the overall product quality evaluation model, periodically evaluates the satisfaction degree of the current control strategy for the target quality index, and dynamically adjusts the strategy parameters to achieve the self-learning and self-correction of the control system. To improve the adaptability of this control system to the actual blow molding process, construct a double-time-domain control scheme. In this scheme, according to the difference between the parameter change speed and the physical response time, the controlled objects are divided into slow-varying parameters that need continuous adjustment and fast-varying parameters that need fast response. Among them, the temperature control group is suitable for adopting a continuous-time control strategy due to its strong thermal inertia and slow-varying characteristics. The control system achieves precise control of the temperature field through real-time fine-tuning. The pressure control group, the time control group, and the mechanical control group are more suitable for adopting a discrete-time control strategy due to their short response time and frequent state mutations, and complete the control task by setting the optimal actions within the sampling period.To enable the collaborative operation of two types of control strategies in the same system, a time-domain conversion interface is constructed to seamlessly convert data formats, time bases, and execution mechanisms between the continuous and discrete time domains, thereby realizing cross-time-domain switching and joint scheduling of control tasks. At the same time, to ensure the safe operation of the control system under extreme or abnormal conditions, according to the historical change laws of various parameters and the equipment carrying capacity, a change rate limit ΔPmax and an absolute value range [Pmin, Pmax] are set for all process parameters to construct a safety control boundary. This boundary is checked and restricted before the execution of the control strategy to ensure that any parameter adjustment is within the safe range of the equipment, preventing molding out of control or equipment damage caused by over-regulation or mis-regulation. Integrating the parameter grouping structure, coordination requirement table, three-layer control architecture, dual-time-domain control scheme, and safety control boundary, a decentralized coordination algorithm based on event triggering is used to achieve centralized control within each control group and dynamic collaborative adjustment between groups, and a highly flexible, self-adaptive, and robust multi-time-domain collaborative control strategy is constructed.

[0018] In one example, a plurality of different sensors are arranged in a plastic container blow molding device to simultaneously collect blow molding process parameter data, including: A platinum resistance temperature sensor is installed in the barrel area of the plastic container blow molding device to monitor the temperature of the feeding section, compression section, homogenization section, metering section, and head section respectively, and obtain the blank temperature distribution data; Pressure sensors are arranged on the inner wall of the mold cavity to monitor the pressure at the top, middle, and bottom of the cavity, and obtain the cavity pressure distribution data; A pressure sensor is installed on the air path of the blowing system to monitor the pressure change during the blowing process in real time, and obtain the blowing pressure change curve; A displacement sensor is installed on the mold clamping mechanism to monitor the mold clamping accuracy and mold deformation, and obtain the mold dynamic deformation data; The blank temperature distribution data, cavity pressure distribution data, blowing pressure change curve, and mold dynamic deformation data are subjected to data standardization processing and combination to obtain the blow molding process parameter data.

[0019] In this example, temperature sensors are arranged in the barrel area of ​​the blow molding equipment. By installing high-precision platinum resistance temperature sensors in the feed section, compression section, homogenization section, metering section and head section in sequence, the thermal state information of the blank at different processing stages is captured. Platinum resistance temperature sensors are suitable for precision temperature detection tasks due to their high linearity, fast response and good stability. The sensor in the feed section is used to sense the efficiency of the initial heating of the cold material, the sensor in the compression section helps to determine the viscosity change of the material during the compression process, the sensor in the homogenization section is used to verify whether the heat distribution is uniform, the sensor in the metering section ensures that the material flows into the die head at the set temperature, and the sensor in the die head section directly affects the final blow molding quality. By synchronously collecting the temperature data of the above five areas, the blank temperature distribution data is constructed to evaluate the uniformity and response efficiency of the heating system. For example, when there is a significant temperature difference between the feed section and the homogenization section, it means that the heating efficiency is uneven. At the same time, a pressure sensor is deployed on the inner wall of the mold cavity to obtain the filling pressure distribution of the molten plastic in the cavity in real time. By arranging sensors at three different positions, namely the top, middle and bottom of the cavity, a three-dimensional pressure monitoring network is formed to effectively capture the pressure gradient distribution differences caused by melt flow, cavity structure complexity or uneven air intake direction during the blow molding process. This distributed layout can determine whether the mold is fully filled and reveal stress concentration problems caused by cavity asymmetry or insufficient local cooling. For example, if the top sensor shows that the pressure is much lower than the bottom, it means that the blow molding gas has not fully expanded to the upper area, resulting in insufficient wall thickness in the upper part or even bulging instability. By analyzing this type of pressure distribution data, the optimization design of the mold exhaust structure and air intake direction is guided. Pressure sensors are installed on the air path of the blowing system to monitor the air pressure changes during the blowing process in real time. These sensors capture the instantaneous rising pressure at the beginning of the blowing, the duration of the pressure stabilization and the pressure release speed, and then construct a blowing pressure change curve. This curve is used to analyze whether there are abnormal fluctuations in the blowing process. For example, in high-frequency production, due to fatigue of the compressed air system or valve lag, the blowing pressure rise speed slows down or the pressure stabilization time is shortened, resulting in incomplete inflation and product structural deformation. For example, when it is detected that the hysteresis time is significantly prolonged during the pressure rise stage, the system can automatically identify it as valve control hysteresis or insufficient air source, and link the control system for early warning and compensation adjustment. High-resolution displacement sensors are deployed on the mold clamping mechanism to obtain the dynamic deformation and clamping accuracy of the mold. These displacement sensors monitor in real time whether the mold is offset, open, misaligned or fatigue deformed under high-pressure blow molding and high-temperature environment. Data from different types of sensors are standardized and encoded in a unified format. Dimensional inconsistency is eliminated through normalization, and timestamps are synchronized to ensure accurate alignment of parameters on the time axis. Anomaly elimination algorithms and low-pass filtering methods are used to remove high-frequency noise or distorted data to ensure the stability and accuracy of subsequent analysis and modeling.Integrate the standardized blank temperature distribution data, cavity pressure distribution data, blowing pressure change curve, and die dynamic deformation data into a structured dataset, namely the blow molding process parameter data.

[0020] In an example, perform wavelet threshold denoising and anomaly recognition on the blow molding process parameter data to obtain the process parameter anomaly recognition results, including: Perform wavelet transform on the blow molding process parameter data to obtain the denoised parameter data; Synchronize the blank temperature distribution data in the denoised parameter data in the time domain to obtain the temperature curve in the preform stage, and perform piecewise linear fitting on the temperature curve in the preform stage to obtain the temperature gradient and temperature uniformity index features; Perform principal component analysis on the cavity pressure distribution data in the denoised parameter data to obtain the pressure distribution feature data; Extract features from the blowing pressure change curve in the denoised parameter data using a four-parameter equation to obtain the blowing pressure feature data; Perform time-frequency analysis on the die dynamic deformation data in the denoised parameter data to obtain the die dynamic feature data; Combine the temperature gradient and temperature uniformity index features, pressure distribution feature data, blowing pressure feature data, and die dynamic feature data to construct a target feature set; Perform K-means clustering and anomaly recognition based on the target feature set to obtain the process parameter anomaly recognition results.

[0021] In this example, for the collected original blow molding process parameter data, due to the influence of various external factors such as the inherent error of the sensor system, electromagnetic interference, and mechanical vibration, there is a certain degree of noise disturbance. The data is denoised by means of wavelet transform. Wavelet transform has good time-frequency localization ability. It decomposes the signal at multiple scales, splitting the original signal into sub-signals of different frequency bands, thereby effectively distinguishing high-frequency noise from low-frequency useful information. The Daubechies wavelet basis function such as db4 with good smoothness and compact support characteristics is used to perform multi-layer wavelet decomposition on the signal. A five-layer decomposition structure is selected to cover the required frequency range, and the soft threshold method is used to process the high-frequency coefficients to eliminate the sharp interference of rapid changes while retaining the main trend, obtaining the denoised parameter data. Time-domain synchronization processing is performed on the billet temperature distribution data among them. Since there are slight offsets in the sampling times of each sensor during the operation of the equipment, to ensure the comparability consistency of data from different batches, each section of the temperature curve is normalized according to a unified forming cycle, standardizing its time axis to the same interval, such as [0, 1], so as to achieve alignment in the time dimension, obtaining the temperature curve in the preforming stage. The temperature curve in the preforming stage is subjected to piecewise linear fitting. Through the piecewise least squares fitting method, the whole curve is divided into multiple continuous sections. The slope of each section is extracted as the temperature gradient, and the deviation coefficient of each section is calculated to evaluate the uniformity of temperature change, forming two core characteristics reflecting heating efficiency and thermal distribution balance, namely temperature gradient and temperature uniformity index. At the same time, principal component analysis is performed on the denoised cavity pressure distribution data. Through the principal component analysis method, its covariance matrix is decomposed, and the principal components with the maximum variance interpretation ability are extracted. These principal components compress the data dimension and effectively aggregate the most representative pressure characteristic information, forming pressure distribution characteristic data, thereby depicting the spatio-temporal distribution pattern of the internal pressure during the mold filling stage. Feature modeling is performed on the denoised blowing pressure change curve, and a four-parameter equation is used to describe it and extract key variables. The blowing process includes typical stages such as rapid rise, constant pressure maintenance, pressure release, and residual fluctuation. By calculating four parameters, namely the pressure rise rate, constant pressure duration, pressure release rate, and final residual pressure value, the dynamic characteristics of the blowing behavior are quantified. The processing of the mold dynamic deformation data relies on time-frequency analysis methods. The mold undergoes small but continuous structural deformations in a high-temperature and high-pressure environment, which are manifested as low-frequency slowly varying signals or contain periodic disturbances. Time-frequency joint analysis means such as short-time Fourier transform or continuous wavelet transform are used to convert the time-series signal into a two-dimensional spectrogram, and the proportion of band energy, the position of the main frequency point, and its corresponding amplitude change are extracted to form the mold dynamic characteristic data, revealing the trend of mold rigidity change, the degree of structural looseness, or the influence of external force interference.Integrate the temperature gradient, temperature uniformity index, principal component of pressure distribution, four-parameter characteristics of blowing pressure, and dynamic frequency domain index of the mold to construct the target feature set. Perform K-means clustering and anomaly recognition based on the target feature set. By setting an appropriate number of clusters, such as five categories, corresponding to five process states of super-optimal, excellent, qualified, critical, and unqualified, the K-means algorithm divides all samples into different clusters according to their feature distances and continuously iterates and updates the centroids until convergence. The clustering results depict the distribution boundaries of different process states in the feature space, enabling quality prediction of a newly collected sample based on its category. Combine the clustering results with the set process standard range to identify abnormal samples beyond the normal distribution boundary, which is the process parameter anomaly recognition result. This result is used to feedback to the control system, trigger the warning mechanism, or serve as the input to the optimization model.

[0022] In one example, perform K-means clustering and anomaly recognition based on the target feature set to obtain the process parameter anomaly recognition result, including: Calculate the mean and standard deviation of the target feature set to obtain the process parameter standard space model; Perform the K-means clustering algorithm on the target feature set based on the process parameter standard space model to obtain the clustering quality region division result; Based on the clustering quality region division result, design an elliptical influence function for each feature parameter, and calculate the influence degree value of each feature parameter through the elliptical influence function; Perform principal component analysis on the target feature set and the influence degree value to obtain the feature weight matrix; According to the feature weight matrix and the influence degree value, perform weighted summation calculation on the influence of each feature parameter to obtain the comprehensive quality index value; Based on the target feature set and the comprehensive quality index value, identify the process parameter combinations that exceed the normal range to obtain the process parameter anomaly recognition result.

[0023] In this example, the mean and standard deviation of the target feature set are calculated. By traversing the data samples on each feature dimension one by one, the mathematical expectation and fluctuation range of each feature parameter in the historical normal production data are calculated respectively, and a process parameter standard space model is established. This model defines the fluctuation interval of each feature dimension under normal production conditions, depicts the central state through the mean value, and depicts the allowable deviation amplitude through the standard deviation, thus establishing a criterion boundary for identifying abnormal behaviors deviating from the center. On the basis of this standard space model, in order to divide the category distribution state of feature samples in the quality dimension, K-means clustering analysis is performed on the target feature set. This algorithm uses the Euclidean distance between feature vectors as the similarity measurement standard. By setting the number of clustering centers (such as five categories), the entire feature sample space is divided into several clustering regions representing different quality states, corresponding to five quality grades of super excellent, excellent, qualified, critical and unqualified respectively. In the K-means clustering iteration process, the clustering centers are randomly initialized, then the distance from each sample point to all centers is calculated, and it is assigned to the nearest center category. Then the mean value of all points in this category is recalculated as the new centroid, and the iteration continues until the centroid no longer changes. After clustering is completed, the entire feature space is restructured, and each sample is assigned a clear quality grade label, and its spatial distribution form reveals the quality-driven relationship between process features. On the basis of the clustering quality region division result, a corresponding elliptical influence function is constructed for each feature parameter. This function takes the clustering center as the geometric focus, defines the distance d between the feature parameter value and the clustering centroid as a variable, and expresses the non-linear degree of the influence of this distance on quality through an exponential decay function within a certain influence radius R. When the feature value is near the clustering center, its influence function value is close to 1, indicating that this feature parameter has a strong supporting effect on the current quality state; while when it deviates farther from the centroid, the function value drops sharply, indicating that this parameter deviates from the ideal value and has an adverse impact on product quality. In this way, the performance of all features in different quality grade regions is quantitatively evaluated, and an influence degree matrix of each sample in each dimension is formed, thus transforming the high-dimensional feature space into an operable set of evaluation values. In order to compress the dimension and highlight the quality dominance of the main features, principal component analysis is performed on the target feature set and the corresponding influence values. Through the collaborative analysis of the feature values and their influence degrees, several principal components with high interpretability and large variance contribution rates are extracted, forming a new linear space coordinate system, and a feature weight matrix is generated accordingly. This matrix represents the projection intensity of each original feature parameter in the principal component space, that is, the influence degree of this parameter on the final comprehensive evaluation index. According to the feature weight matrix and the influence degree value, the weighted sum calculation of the influence of each feature parameter is performed to form a comprehensive quality index value, which reflects the degree of closeness of this sample to the "high-quality center" in the overall feature space. The closer the value is to 1, the closer it is to the ideal state, and the lower the value, the more likely there are process instability or quality hidden dangers.Compare the comprehensive quality index value of each sample with the quality grade threshold set in the standard space model, and combine its original feature combination in the target feature set to determine whether the sample exceeds the normal state distribution range. Once it is identified that the comprehensive quality index is significantly lower than the qualified threshold, or the distance between some key features and their corresponding cluster centers is too far, it is determined that there is a potential process anomaly in the sample. Further, combined with the deviation direction and influence degree in its feature dimension, trace which type of parameter has shifted, and finally determine the process parameter combination that causes the quality anomaly.

[0024] In one example, establish a transfer function matrix between process parameters and quality indicators according to the process parameter anomaly identification result, and obtain the target process parameter set through multi-objective optimization, including: Classify and evaluate the five quality indicators according to the process parameter anomaly identification result to obtain a multi-dimensional quality indicator system. The five quality indicators include wall thickness uniformity, sealing performance, dimensional accuracy, surface quality, and cleanliness; Conduct statistical analysis on the corresponding relationship between the parameter deviation data and product quality in the process parameter anomaly identification result to obtain a preliminary influence relationship table of each parameter on the quality indicators; Design an orthogonal experiment plan based on the preliminary influence relationship table and perform multiple batches of tests to obtain a product quality experiment data set under different parameter combinations; Use the Gaussian process regression algorithm to establish a mathematical mapping relationship between each process parameter and each quality indicator for the product quality experiment data set to obtain a transfer function matrix between process parameters and quality indicators; Construct a multi-objective optimization model with five quality indicators as optimization objectives according to the transfer function matrix, and calculate the weight coefficients of each quality objective through the analytic hierarchy process to obtain a weighted comprehensive objective function; Based on the particle swarm optimization algorithm, iteratively solve the weighted comprehensive objective function to obtain the target process parameter set.

[0025] In this example, five quality indicators are classified and evaluated according to the results of process parameter anomaly identification to obtain a multi-dimensional quality indicator system. In blow molding, the five core quality indicators jointly determine the overall performance and application reliability of the container, namely wall thickness uniformity, sealing performance, dimensional accuracy, surface quality, and cleanliness. Wall thickness uniformity reflects the consistency of material distribution during the blowing process and is the basis for evaluating mechanical strength and compressive performance; the sealing performance is the core indicator for measuring whether there is a leakage risk in the container during liquid or gas storage applications; dimensional accuracy reflects the comprehensive results of mold closing, cooling shrinkage, and overall deformation control; surface quality is related to the appearance flatness and structural integrity of the product; cleanliness is an important indicator for evaluating the level of residual particulate matter inside the product. Based on these indicators, the anomaly identification results are classified according to the scope of influence, and a multi-dimensional quality indicator system is established. All known parameter deviation samples are extracted from the anomaly identification results, and the mapping relationship between them and the product test results is statistically analyzed to construct a preliminary connection between parameter offset and quality indicator response. Through means such as clustering, correlation analysis, or regression trend comparison, the degree and direction of the impact of each process parameter on each quality indicator are obtained, and a preliminary impact relationship table is obtained. According to the preliminary impact relationship table, an orthogonal experimental scheme for the system is designed, and representative parameter combinations are selected for multi-batch product trial production and performance testing. The orthogonal experiment covers all main effects and some interaction effects with the minimum set of parameter levels, while maintaining information richness while controlling the scale of the experiment. During the experiment execution, each set of parameter settings corresponds to a set of quality inspection results, and all test samples form a complete product quality experimental data set covering various working conditions. This data set records the wall thickness, dimensions, surface, sealing performance, and cleanliness performance under different process parameter combinations. The Gaussian process regression algorithm is used to establish a mathematical mapping relationship between each process parameter and each quality indicator in the product quality experimental data set. Gaussian process regression is a non-parametric Bayesian modeling method that can capture complex non-linear relationships without explicitly specifying the function form and give the mean and confidence interval of the predicted value. This algorithm takes the covariance function as the core component, learns the association pattern between parameters, and has good generalization ability. After training with the Gaussian process model, a transfer function matrix between process parameters and five quality indicators is constructed. Each element in this matrix corresponds to the impact function of a certain parameter on a certain quality indicator, thus realizing the function mapping from the parameter space to the quality response space, with good interpretability and predictive ability. According to the transfer function matrix, a multi-objective optimization model with five quality indicators as the optimization objectives is constructed to find the optimal solution in the parameter space. Considering that the importance of different quality indicators in actual products is not the same, the analytic hierarchy process is used to assign weights to each indicator. The analytic hierarchy process calculates the relative weight coefficients of each quality indicator by constructing a pairwise comparison matrix, collecting expert judgment or historical weight data, and passing the consistency test.Based on this, each quality objective function is weighted and aggregated to form a comprehensive objective function. The weighted comprehensive objective function is iteratively solved based on the particle swarm optimization algorithm. The particle swarm optimization algorithm is a bionic intelligent algorithm. By simulating the foraging behavior of bird flocks, multiple particles search in the parameter space. Each particle represents a combination of process parameters. In each iteration, the flight direction is adjusted according to the current individual optimal position and the global optimal position. As the number of iterations increases, the particles gradually gather in the optimal solution region, thus achieving the optimization goal. During the solution process, the algorithm continuously uses the comprehensive objective function to evaluate the fitness of each particle, and updates the speed and position until convergence. The finally output particle position is the set of target process parameters. This parameter combination can ensure that all five quality indicators meet the preset range and also has good stability and anti-disturbance ability.

[0026] In an example, the set of target process parameters is input into a double-integral enhanced recurrent neural network for parameter regulation calculation to obtain the process parameter adjustment amount, including: Calculate the difference between the set of target process parameters and the measured value of the current process parameters to obtain the process parameter deviation data; Input the process parameter deviation data into the input layer of the double-integral enhanced recurrent neural network. The input layer contains input nodes matching the number of process parameters to obtain the target input vector; Input the target input vector into the first hidden layer and the second hidden layer of the double-integral enhanced recurrent neural network for non-linear transformation to obtain the hidden layer feature representation; Perform recursive calculation according to the hidden layer feature representation and the historical state information of the previous two moments to implement the memory function of the change trend of process parameters during the blow molding process, and obtain the current state vector; Apply the double-integral error function to enhance the process parameter deviation data to obtain the enhanced error vector. The double-integral error function considers three dimensions: the instantaneous deviation, the cumulative deviation, and the second-order cumulative deviation; Input the current state vector and the enhanced error vector into the output layer of the double-integral enhanced recurrent neural network, and combine with the anti-interference compensation mechanism to perform adaptive compensation for temperature fluctuations, pressure changes, and material batch differences to obtain the process parameter adjustment amount.

[0027] In this example, the difference between the target process parameter set and the measured values of the current process parameters is calculated. The difference between the two is calculated item by item to obtain the current process parameter deviation data. This difference reflects the specific deviation degree between the current working condition and the theoretical optimal setting, and constitutes the basic variable of the controller input. The deviation of each parameter represents a risk point affecting quality. For example, a low blowing pressure leads to uneven wall thickness, and a high mold temperature causes surface melt tumors and other problems. The process parameter deviation data is input into the input layer of the double integral enhanced recurrent neural network. Since the target process parameter set contains multiple dimensions, such as temperature, pressure, time, flow rate, displacement, etc., about fifteen key variables in total, the number of input layer nodes needs to strictly correspond to the parameter dimensions, that is, one-to-one mapping, to construct a complete target input vector. This input vector will be mapped and transformed layer by layer inside the network, so as to guide the neural network to learn the non-linear correspondence relationship between the input data and the control instructions. In the next non-linear mapping stage, this input vector is transmitted to the first hidden layer and the second hidden layer of the network. Each hidden layer node processes the input signal through a non-linear activation function such as Tanh, and converts it into a more abstract and expressive feature representation. This processing can effectively capture the coupling relationship, non-linear change trend and potential dynamic control mode between different deviation terms. For example, the coupling between temperature deviation and mold deformation, or the influence of pressure fluctuation on cooling time, etc., are encoded into a multi-dimensional feature map inside the hidden layer. The two-layer non-linear transformation network extracts the structured hidden layer feature representation, which will serve as the key support for the network's recursive operation and subsequent error correction. In order to enhance the network's dynamic response ability and time series modeling ability, a recursive structure is designed inside the network. The output state at each moment is not only affected by the current input, but also refers to the historical hidden state information of the previous moment and the moment before the previous moment. The double-time sequence recursive structure enables the network to have a time memory function, which can identify dynamic features such as persistent deviation, trend disturbance and periodic fluctuation, and construct a state vector representing the current system state. This state vector converges the current working condition data characteristics and inherits the patterns in the historical operation trajectory, and can judge whether the system is in a stable deviation period, a transitional disturbance period or a mutation interference period, so as to guide the network to make more reasonable adjustment decisions. At the same time, in order to improve the system's response sensitivity and regulation accuracy to abnormal changes, the original process parameter deviation data is processed, and the double integral error function is applied to enhance it. This function introduces three dimensions of instantaneous error, first integral error and second integral error, that is, it not only pays attention to the current deviation value itself, but also introduces the analysis of the time accumulation trend and acceleration change of the error, so as to effectively amplify the influence of persistent deviation, trend deviation and error growth rate on the control system. For example, if the instantaneous deviation of a certain parameter is small, but its cumulative deviation value is continuously increasing, or the error change rate is gradually increasing, it means that the system is moving away from the stable interval and immediate intervention is necessary.The enhanced error vector after processing plays a stronger guiding role in decision-making, ensuring that the controller can accurately judge anomalies at multiple scales. The current state vector and the enhanced error vector are input into the output layer of a double-integral enhanced recurrent neural network, and an anti-interference compensation mechanism is introduced in the output calculation to cope with a large number of disturbance factors existing in the actual production environment, such as temperature fluctuations, air source pressure changes, and material batch performance differences. The compensation mechanism consists of a state observer and a feedforward compensation path. The observer monitors the changes in the production environment and raw material parameters in real time. When a high coupling relationship is found between the control variable and the external disturbance, the compensation channel is triggered to automatically adjust the output result. For example, in the case where it is identified that the current blowing pressure is low and the air source supply pressure is unstable, the adjustment amplitude is dynamically increased according to the interference intensity to prevent forming failure caused by insufficient regulation. The output layer combines the error input, state evaluation, and compensation model, and finally outputs a set of targeted and robust process parameter adjustment amounts, which are transmitted as feedback signals to the equipment execution layer.

[0028] In one example, the production commissioning control method for plastic containers further includes: Dividing the process parameters into a temperature control group, a pressure control group, a time control group, and a mechanical control group according to the process parameter adjustment amounts to obtain a parameter grouping structure; Based on the parameter grouping structure, a parameter coupling degree matrix is established, and the mutual influence coefficients between the parameters are calculated according to the parameter coupling degree matrix to quantify the interaction relationships between different control groups, obtaining a coordination requirement table between groups; Based on the coordination requirement table between groups, a three-layer control architecture is constructed. The three-layer control architecture includes a bottom-layer execution control layer, a middle-layer strategy coordination layer, and a top-layer optimization decision layer; Applying a continuous-time control strategy to the temperature control group, applying a discrete-time control strategy to the pressure control group, the time control group, and the mechanical control group, and realizing seamless switching of different time-domain control methods through a time-domain conversion interface to obtain a dual-time-domain control scheme; Setting parameter change rate limits and limit range constraints according to the three-layer control architecture and the dual-time-domain control scheme to obtain a safety control boundary; According to the parameter grouping structure, the coordination requirement table between groups, the three-layer control architecture, the dual-time-domain control scheme, and the safety control boundary, realizing in-group centralized control and between-group collaborative regulation through an event-triggered decentralized coordination algorithm to obtain a multi-time-domain collaborative control strategy.

[0029] In this example, the process parameter adjustment amounts output by the neural network are functionally classified and organically divided according to the physical meanings, control methods, and objects of action corresponding to each parameter, forming four subsystems: the temperature control group, the pressure control group, the time control group, and the mechanical control group. For example, the preheating temperature, mold temperature, and temperatures of each section of the barrel are classified into the temperature control group; the blowing pressure, holding pressure, etc. are included in the pressure control group; the parameters with clear timing attributes such as holding time, cooling time, etc. are classified into the time control group; the screw speed, mold displacement, clamping stroke, etc. are classified into the mechanical control group. In this way, different functional parameters are decoupled at the structural level, facilitating the separate setting of control strategies and laying a logical foundation for subsequent in-group centralized control and inter-group collaborative control. After completing the parameter grouping, the coupling degree relationship between the parameters is established to quantify the interactive influence between each control group. By analyzing historical production data and simulation results, a parameter coupling degree matrix is constructed, and each element of this matrix represents the numerical linkage strength, adjustment coupling, and output response similarity between two parameters. For example, there is a negative correlation coupling between the mold temperature and the cooling time, while the blowing pressure and the clamping displacement have a synergistic amplification effect under certain working conditions. By calculating the correlation coefficient, information entropy contribution rate, and sensitivity coefficient of the control transfer function between the parameters, an objective coupling degree index is obtained. On this basis, the coupling strengths of the parameters in the matrix are classified, summed, and normalized according to their respective groups, forming a coordination requirement table between groups, which shows the interdependence between different control groups in matrix form. Based on the obtained parameter structure and coordination requirements, a three-layer control architecture is constructed to achieve the efficient transmission and intelligent optimization of control signals from the target parameters to the device execution. This three-layer architecture includes the bottom-layer execution control layer, the middle-layer strategy coordination layer, and the top-layer optimization decision layer. The bottom-layer execution control layer directly interfaces with the physical execution units of the device, such as temperature controllers, pneumatic pressure regulators, motor drivers, displacement actuators, etc. It has a short sampling period and a fast response speed, mainly realizing the immediate adjustment and closed-loop execution of parameters; the middle-layer strategy coordination layer performs the coordinated adjustment of multiple groups of parameters based on the grouping structure and the coordination requirement table. Its main functions include in-group centralized optimization, inter-group collaborative compensation, event-triggered scheduling, etc. The coordination layer ensures the consistency and stability of the system operation by periodically updating the adjustment strategy and real-time monitoring of the coupling changes; while the top-layer optimization decision layer makes global strategy adjustments and learning optimizations according to the quality prediction model, parameter evolution model, and working condition identification mechanism. The strategy output period is relatively long but covers more dimensional information, taking into account both short-term effects and long-term stability. Since the dynamic characteristics of different control groups are significantly different, targeted time-domain control methods are adopted.Due to its large thermal inertia and obvious response lag, the temperature control group is suitable for adopting a continuous-time control strategy. Its controller continuously outputs adjustment instructions according to the trend of error change to achieve smooth and gradual temperature tracking. The pressure, time, and mechanical control groups generally have the characteristics of rapid change and transient response and are more suitable for adopting a discrete-time control strategy, that is, making precise adjustment actions within the set time interval to prevent over-regulation or delayed control. To achieve the unified integration of these two types of time-domain control methods, a time-domain conversion interface module is introduced. This module realizes data synchronization, unit conversion, and scheduling coordination between the continuous and discrete time domains. For example, it converts the mold temperature curve into a discrete temperature adjustment sequence for the middle layer to use, or smooths the impact signal generated by the discrete controller for input to the continuous controller to ensure seamless switching and logical consistency between the control paths. The dual-time-domain control scheme improves the adaptability of the control system to the characteristics of various parameters and avoids the accumulation of adjustment errors caused by differences in response time under a single control mode. After completing the construction of the control path and time-domain logic, set the change rate limit and absolute limit range for each process parameter to construct a safety control boundary. These limit conditions prevent equipment damage or product defects caused by misoperation or algorithm oscillation and are also important boundary conditions for the self-correction and constraint calculation of the control system. During the execution of the system, all control instructions need to pass through the safety boundary check first to ensure that they are implemented within the physically controllable range, thereby improving the robustness and fault tolerance of the overall system. Based on the foregoing grouping structure, coordination requirements, three-layer control architecture, dual-time-domain scheme, and safety boundary, all control logics are embedded into the event-triggered decentralized coordination algorithm to achieve the combination of centralized control within the group and dynamic cooperative adjustment between groups. Under this algorithm architecture, each control group operates as an independent subsystem. When there is no external interference or the internal state of the group is stable, it maintains low-frequency adjustment or enters the monitoring state. Once a parameter mutation, a decrease in the quality index, or an external interference-triggered event condition is detected, the system immediately activates the control path of the relevant group and notifies other groups coupled with it to make a synchronous response according to the coordination requirement table and the set coupling strength.

[0030] Referring to Figure 2 , this embodiment provides a production debugging control system for plastic containers, including: The acquisition module 1 is used to arrange multiple different sensors in the plastic container blow molding equipment and collect the process parameter data during the blow molding process at the same time; The anomaly recognition module 2 is used to perform wavelet threshold denoising and anomaly recognition on the blow molding process parameter data to obtain the process parameter anomaly recognition result; The multi-objective optimization module 3 is used to establish the transfer function matrix between the process parameters and the quality index according to the process parameter anomaly recognition result and obtain the target process parameter set through multi-objective optimization; The parameter regulation calculation module 4 is configured to input the target process parameter set into a dual-integral enhanced recurrent neural network for parameter regulation calculation to obtain a process parameter adjustment amount.

[0031] In this embodiment, for the specific implementation of each unit in the above system embodiment, please refer to the description in the above method embodiment, and details will not be repeated here.

[0032] It should be noted that in this article, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, system, article or method including a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, system, article or method. Without further limitation, an element defined by the statement "including a..." does not exclude the presence of additional identical elements in the process, system, article or method including the element.

[0033] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A production debugging control method for a plastic container, characterized in that: include: Arrange multiple different sensors in the plastic container blow molding equipment to collect blow molding process parameter data at the same time; Performing wavelet threshold denoising and abnormality identification on the blow molding process parameter data to obtain a process parameter abnormality identification result; Establishing a transfer function matrix between process parameters and quality indicators according to the process parameter anomaly identification results, and obtaining a target process parameter set through multi-objective optimization; The target process parameter set is input into a double-integral enhanced recursive neural network to perform parameter control calculations to obtain process parameter adjustment amounts.

2. The production debugging control method of plastic container according to claim 1, characterized in that: The method of arranging a plurality of different sensors in the plastic container blow molding equipment and collecting blow molding process parameter data at the same time includes: Install platinum resistance temperature sensors in the barrel area of ​​the plastic container blow molding equipment to monitor the temperature of the feeding section, compression section, homogenization section, metering section and die section respectively to obtain the temperature distribution data of the blank; Pressure sensors are arranged on the inner wall of the mold cavity to monitor the pressure at the top, middle and bottom of the cavity to obtain cavity pressure distribution data; Install a pressure sensor on the air path of the blowing system to monitor the pressure change during the blowing process in real time and obtain a blowing pressure change curve; Install a displacement sensor on the mold clamping mechanism to monitor the clamping accuracy and mold deformation and obtain the dynamic deformation data of the mold; The blank temperature distribution data, the cavity pressure distribution data, the blowing pressure variation curve and the mold dynamic deformation data are subjected to data standardization processing and combined to obtain blow molding process parameter data.

3. The production debugging control method of plastic container according to claim 1, characterized in that: The step of performing wavelet threshold denoising and abnormality identification on the blow molding process parameter data to obtain a process parameter abnormality identification result includes: Performing wavelet transform on the blow molding process parameter data to obtain denoised parameter data; Performing time domain synchronization on the temperature distribution data of the blank in the denoised parameter data to obtain a temperature curve in the preforming stage, and performing piecewise linear fitting on the temperature curve in the preforming stage to obtain temperature gradient and temperature uniformity index characteristics; Performing principal component analysis on the cavity pressure distribution data in the denoised parameter data to obtain pressure distribution characteristic data; Using a four-parameter equation to extract features of the blowing pressure change curve in the denoised parameter data to obtain blowing pressure feature data; Performing time-frequency analysis on the dynamic deformation data of the mold in the denoised parameter data to obtain dynamic characteristic data of the mold; Combining the temperature gradient and temperature uniformity index features, the pressure distribution feature data, the blowing pressure feature data and the mold dynamic feature data to construct a target feature set; K-means clustering and anomaly recognition are performed based on the target feature set to obtain a process parameter anomaly recognition result.

4. The production debugging control method of plastic container according to claim 3, characterized in that: The performing K-means clustering and anomaly identification based on the target feature set to obtain a process parameter anomaly identification result includes: Calculating the mean and standard deviation of the target feature set to obtain a standard space model of process parameters; Executing a K-means clustering algorithm on the target feature set based on the process parameter standard space model to obtain a cluster quality area division result; Based on the cluster quality area division result, an elliptical influence function is designed for each characteristic parameter, and the influence degree value of each characteristic parameter is calculated by the elliptical influence function; Performing principal component analysis on the target feature set and the influence degree value to obtain a feature weight matrix; According to the feature weight matrix and the impact degree value, weighted sum calculation is performed on the impact of each feature parameter to obtain a comprehensive quality index value; Based on the target feature set and the comprehensive quality index value, process parameter combinations exceeding a normal range are identified to obtain process parameter abnormality identification results.

5. The production debugging control method of a plastic container according to claim 1, characterized in that: The method of establishing a transfer function matrix between process parameters and quality indicators according to the process parameter anomaly identification result, and obtaining a target process parameter set through multi-objective optimization, includes: According to the abnormal identification results of the process parameters, five quality indicators are classified and evaluated to obtain a multi-dimensional quality indicator system, wherein the five quality indicators include wall thickness uniformity, sealing performance, dimensional accuracy, surface quality and cleanliness; Performing statistical analysis on the corresponding relationship between the parameter deviation data in the process parameter abnormality identification result and the product quality, and obtaining a preliminary influence relationship table of each parameter on the quality index; Design an orthogonal experimental scheme based on the preliminary influence relationship table and perform multiple batches of tests to obtain product quality experimental data sets under different parameter combinations; A Gaussian process regression algorithm is used to establish a mathematical mapping relationship between each process parameter and each quality index for the product quality experimental data set, and a transfer function matrix between the process parameters and the quality index is obtained; A multi-objective optimization model with five quality indicators as optimization targets is constructed according to the transfer function matrix, and a weight coefficient of each quality target is calculated by the hierarchical analysis method to obtain a weighted comprehensive objective function; The weighted comprehensive objective function is iteratively solved based on a particle swarm optimization algorithm to obtain a target process parameter set.

6. The production debugging control method of a plastic container according to claim 1, characterized in that: The step of inputting the target process parameter set into a double-integral enhanced recursive neural network to perform parameter control calculation to obtain a process parameter adjustment amount includes: Calculating the difference between the target process parameter set and the actual measured value of the current process parameter to obtain process parameter deviation data; Inputting the process parameter deviation data into an input layer of a double-integral enhanced recurrent neural network, wherein the input layer includes input nodes matching the number of process parameters, to obtain a target input vector; Inputting the target input vector into the first hidden layer and the second hidden layer of the double integral enhanced recursive neural network for nonlinear transformation to obtain a hidden layer feature representation; Perform recursive calculation according to the hidden layer feature representation and the historical state information of the previous two moments to realize the memory function of the change trend of the process parameters in the blow molding process and obtain the current state vector; Applying a double integral error function to enhance the process parameter deviation data to obtain an enhanced error vector, wherein the double integral error function simultaneously considers three dimensions: instantaneous deviation, cumulative deviation, and second-order cumulative deviation; The current state vector and the enhanced error vector are input into the output layer of the dual-integral enhanced recursive neural network, and combined with an anti-interference compensation mechanism, adaptive compensation is performed on temperature fluctuations, pressure changes and material batch differences to obtain a process parameter adjustment amount.

7. The production debugging control method of a plastic container according to claim 1, characterized in that: The production debugging control method of the plastic container also includes: Dividing the process parameters into a temperature control group, a pressure control group, a time control group and a mechanical control group according to the process parameter adjustment amount to obtain a parameter grouping structure; A parameter coupling matrix is ​​established based on the parameter grouping structure, and the mutual influence coefficients between the parameters are calculated according to the parameter coupling matrix to quantify the interaction between different control groups, thereby obtaining a coordination demand table between groups; Building a three-layer control architecture based on the inter-group coordination demand table, the three-layer control architecture includes a bottom execution control layer, a middle strategy coordination layer and a top optimization decision layer; Applying a continuous time control strategy to the temperature control group, applying a discrete time control strategy to the pressure control group, the time control group and the mechanical control group, and realizing seamless switching of different time domain control methods through a time domain conversion interface to obtain a dual time domain control scheme; Setting parameter change rate limits and limit range constraints according to the three-layer control architecture and the dual-time domain control scheme to obtain a safe control boundary; According to the parameter grouping structure, the inter-group coordination requirement table, the three-layer control architecture, the dual-time domain control scheme and the safety control boundary, the centralized control within the group and the coordinated adjustment between groups are realized through an event-triggered distributed coordination algorithm to obtain a multi-time domain coordinated control strategy.

8. A production debugging control system for plastic containers, characterized in that: Steps for implementing the production debugging control method of the plastic container according to any one of claims 1 to 7, the production debugging control system of the plastic container comprising: A collection module is used to arrange multiple different sensors in the plastic container blow molding equipment and collect blow molding process parameter data at the same time; An abnormality identification module, used for performing wavelet threshold denoising and abnormality identification on the blow molding process parameter data to obtain a process parameter abnormality identification result; A multi-objective optimization module, used to establish a transfer function matrix between process parameters and quality indicators according to the process parameter anomaly identification result, and obtain a target process parameter set through multi-objective optimization; The parameter control calculation module is used to input the target process parameter set into the double-integral enhanced recursive neural network to perform parameter control calculation to obtain the process parameter adjustment amount.

Images