PVC film production system and method based on intelligent control

Abstract

A PVC film production system and method based on intelligent control belong to the field of intelligent production of high polymer materials. The method comprises the following steps: deploying a multi-point temperature sensor matrix in each roller of a calender to construct a temperature field distribution information acquisition network; constructing a fuzzy neural network adaptive PID controller to output real-time PID control parameters; establishing a deep reinforcement learning temperature optimization module adopting an actor-critic architecture, outputting a PID parameter adjustment amount as an optimization given value according to roller temperature field data, product quality indicators and historical sequences of control parameters; executing closed-loop temperature control, combining the optimization given value with real-time PID control parameters to adjust the roller heating or cooling device; and performing calendering production based on the optimized temperature parameters, and obtaining thickness uniformity and film surface flatness data through an online quality detection system, which is fed back to the deep reinforcement learning module to update the strategy. The application realizes precise adaptive regulation and control of the roller temperature of the calender.

Description

Technical Field

[0001] This application belongs to the field of intelligent manufacturing of polymer materials, specifically relating to a PVC film production system and method based on intelligent control. Background Technology

[0002] PVC film production mainly employs a calendering process. The calender uses multiple sets of rollers to continuously heat, plasticize, calender, and cool the PVC raw material, ultimately forming the film product. During the calendering process, roller temperature is a core process parameter determining product quality. The uniformity and stability of the temperature field directly affect the degree of plasticization, molecular chain orientation, and the film's thickness uniformity and surface smoothness.

[0003] Currently, temperature control in traditional PVC film production often employs fixed PID parameters or simple segmented control methods. These control strategies struggle to adapt to complex operating conditions such as fluctuations in incoming material ratios, changes in ambient temperature, and variations in equipment operating status. Furthermore, due to a lack of real-time sensing of the roller temperature field distribution, traditional methods cannot effectively address the non-uniformity and dynamic changes in the roller surface temperature field. This results in significant temperature fluctuations in the plasticizing zone, making it difficult to ensure the consistency of material molecular chain orientation, leading to problems such as poor product thickness uniformity, wavy defects on the film surface, and a high scrap rate. Therefore, achieving real-time adaptive and precise control of the calender roller temperature to ensure a uniform and stable temperature field has become a key technical challenge for improving the quality of PVC film products. Summary of the Invention

[0004] The purpose of this invention is to provide a PVC film production system and method based on intelligent control, which can effectively solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A PVC film production method based on intelligent control includes the following specific steps: By deploying a multi-point temperature sensor matrix inside each roll of the calender, temperature data at multiple locations along the axial and circumferential directions of the rolls can be acquired in real time to construct a temperature field distribution information acquisition network. A fuzzy neural network adaptive PID controller is constructed, and the real-time acquired operating condition parameter signals are input into the fuzzy neural network adaptive PID controller to output real-time PID control parameters. A deep reinforcement learning temperature optimization module is established. The deep reinforcement learning temperature optimization module adopts an actor-commentator architecture. Based on the state space consisting of roller temperature field data, product quality indicators and historical sequence of control parameters, it outputs PID parameter adjustment as optimization setpoint. Closed-loop temperature control is implemented by providing the optimized setpoint output by the deep reinforcement learning temperature optimization module to the fuzzy neural network adaptive PID controller. The fuzzy neural network adaptive PID controller combines the real-time temperature deviation to calculate the control signal, thereby adjusting the actuator of the roller heating or cooling device. The PVC film calendering production is carried out based on optimized temperature parameters, and the thickness uniformity data and film surface flatness data of the PVC film are acquired in real time through an online quality inspection system. The thickness uniformity data and film surface flatness data are fed back to the deep reinforcement learning temperature optimization module to update the strategy of the deep reinforcement learning temperature optimization module.

[0006] Furthermore, a multi-point temperature sensor matrix is ​​deployed within each roller of the calender, specifically including: Eight to 12 sensor points are arranged axially on each roller of the calender, and four to six sensor points are arranged circumferentially. A total of 32 to 72 temperature sensors are deployed on a single roller. The temperature sensors are arranged in a cross distribution along the axial and circumferential directions of the roller, and the sensor points are densely arranged in the edge effect area at both ends of the roller. The temperature sensor has a temperature measurement accuracy of ±0.1℃ to ±0.3℃ and a sampling frequency of 10 to 50 times per second. Based on sensor data that has passed the validity test, a spatial interpolation algorithm is used to reconstruct the temperature field on the roller surface, generating two-dimensional temperature distribution maps in the axial and circumferential directions, and the temperature field update frequency is synchronized with the sensor sampling frequency.

[0007] Furthermore, a fuzzy neural network adaptive PID controller is constructed, specifically including: The fuzzy neural network adaptive PID controller consists of an input layer, a fuzzification layer, a fuzzy inference layer, and an output layer. The input layer receives five input variables: temperature deviation, temperature deviation change rate, raw material viscosity, calendering speed, and inter-roll pressure. The input layer performs debiasing and normalization on the five input variables, mapping them to a uniform numerical range. A parameter boundary constraint mechanism is set after the output layer, wherein the variation range of the proportional parameter is limited to 0.5 times to 2.0 times the basic proportional parameter, the variation range of the integral parameter is limited to 0.3 times to 1.5 times the basic integral parameter, and the variation range of the derivative parameter is limited to 0.2 times to 2.5 times the basic derivative parameter.

[0008] Furthermore, the fuzzy inference layer employs a multi-layer fully connected neural network to implement fuzzy inference functionality, and performs offline pre-training and online fine-tuning of the fuzzy inference layer; Online fine-tuning employs an incremental learning method, updating network parameters each time using only the latest 8 to 32 samples, with a learning rate set to 0.001 to 0.01.

[0009] Furthermore, in the actor-critic architecture of the deep reinforcement learning temperature optimization module, both the actor network and the critic network adopt a long short-term memory network structure; The training objective of the deep reinforcement learning temperature optimization module is to maximize the expected cumulative reward. The reward function of the deep reinforcement learning temperature optimization module is composed of a weighted sum of the product quality improvement and the temperature deviation change, where the weight of product quality improvement is 0.6 to 0.8 and the weight of temperature stability is 0.2 to 0.4. The improvement in product quality is obtained by weighted summation of the improvement in thickness uniformity and the improvement in film surface smoothness.

[0010] Furthermore, during the update process, the deep reinforcement learning temperature optimization module randomly extracts batches of samples from the experience replay buffer and adopts a target network mechanism to maintain target network copies of the actor network and the critic network respectively. Every preset number of steps, the main network parameters are copied to the target network copies.

[0011] Furthermore, the temperature parameter adjustment cycle of the deep reinforcement learning temperature optimization module is set to 30 seconds to 300 seconds, and the deep reinforcement learning temperature optimization module updates the PID parameter adjustment amount as the optimization setpoint at the end of each adjustment cycle; The sampling period for closed-loop temperature control is set to 100 milliseconds to 500 milliseconds, and the control period is set to 1 second to 5 seconds.

[0012] Furthermore, implementing closed-loop temperature control also includes configuring anomaly detection and protection mechanisms: When the temperature deviation exceeds the preset safety threshold, an alarm signal is issued and the emergency response procedure is activated; When a sensor malfunction or communication interruption is detected, the fuzzy neural network adaptive PID controller automatically switches to a safe mode that operates with fixed safe PID parameters.

[0013] Furthermore, the PVC film calendering production is carried out, specifically including: The unit is configured with a multi-stage calendering unit, wherein the temperature of the feed roller is set below the material softening temperature, the temperatures of the second and third rollers are set within the material plasticizing temperature range, and the temperature of the discharge roller is set below the temperatures of the second and third rollers. The online quality inspection system measures the film thickness in real time using a laser online thickness gauge with a measurement accuracy of ±0.1 micrometer to ±0.5 micrometer. The online quality inspection system also uses a vision inspection system to acquire film surface images in real time to calculate surface roughness.

[0014] A PVC film production system based on intelligent control includes: The temperature field acquisition network consists of a matrix of multi-point temperature sensors deployed inside each roll of the calender, used to acquire temperature data at multiple locations along the axial and circumferential directions of the rolls in real time. The fuzzy neural network adaptive PID controller is used to receive operating condition parameter signals and output real-time PID control parameters. The fuzzy neural network adaptive PID controller includes an input layer, a fuzzification layer, a fuzzy inference layer, and an output layer. The deep reinforcement learning temperature optimization module adopts an actor-critic architecture to output PID parameter adjustment as the optimization setpoint based on the state space consisting of roller temperature field data, product quality indicators and historical sequences of control parameters. The closed-loop temperature control execution module is used to receive the optimized setpoint and real-time PID control parameters, and calculate the control signal by combining the real-time temperature deviation to adjust the actuator of the roller heating or cooling device. The multi-stage calendering production unit contains multiple rollers with different temperature settings for performing PVC film calendering production under optimized temperature parameters; The online quality inspection system is used to acquire PVC film thickness uniformity data and film surface flatness data in real time, and feed the thickness uniformity data and film surface flatness data back to the deep reinforcement learning temperature optimization module.

[0015] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention deploys a matrix of high-precision temperature sensors at multiple points within each roller of a calender to construct a complete temperature field distribution information acquisition network. Combined with the synergistic effect of an adaptive PID controller based on a fuzzy neural network and a deep reinforcement learning temperature optimization module, the temperature fluctuations on the roller surface are strictly controlled within a preset accuracy range. Compared with the temperature fluctuations of traditional fixed PID parameter control methods, the temperature control accuracy is significantly improved, providing a reliable temperature environment guarantee for the stable plasticization of PVC film.

[0016] 2. The precise and stable roller temperature field of this invention ensures the consistency of molecular chain orientation of PVC materials during calendering and plasticizing. Compared with traditional methods, the product thickness tolerance is significantly reduced, the film surface flatness is significantly improved, the ripple defect rate is greatly reduced, and the production scrap rate is significantly reduced, thus significantly improving the overall quality stability and pass rate of the product.

[0017] 3. The deep reinforcement learning temperature optimization module of this invention establishes a precise mapping model between temperature parameters and product quality indicators. By continuously learning historical production data and real-time quality feedback information, it autonomously discovers the optimal temperature control strategy. It can achieve dynamic optimization and adjustment of temperature parameters without manual intervention. The adaptive adjustment time of the controller after changes in operating conditions is significantly shortened compared with traditional methods, effectively coping with complex operating condition disturbances caused by fluctuations in incoming material ratios, changes in ambient temperature, and changes in equipment operating status. Attached Figure Description

[0018] Figure 1 This is an overall schematic diagram of a PVC film production method based on intelligent control. Figure 2 A schematic diagram illustrating the core principles of a deep reinforcement learning temperature optimization model; Figure 3 This is a schematic diagram of the hierarchical structure and control flow of a fuzzy neural network adaptive PID controller. Figure 4 This is a schematic diagram of the interaction and data flow in the closed-loop control process of roller temperature; Figure 5 This is a schematic diagram of a multi-stage process. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figures 1 to 5 The present invention will be further described in detail below with reference to specific embodiments.

[0020] Firstly, this invention provides a PVC film production method based on intelligent control. This method comprises five core steps: Step S1: Collecting temperature field distribution data of the calender rollers by deploying a matrix of high-precision temperature sensors at multiple points within each roller of the calender to acquire temperature data at multiple locations along the axial and circumferential directions of the rollers in real time; Step S2: Constructing an adaptive PID controller, integrating fuzzy neural network theory to achieve dynamic adaptive adjustment of PID parameters; Step S3: Establishing a deep reinforcement learning temperature optimization model to establish a mapping relationship between temperature parameters and product quality indicators; Step S4: Executing closed-loop temperature control, combining the optimization results of the deep reinforcement learning module with the adaptive PID controller; Step S5: Executing PVC film calendering production based on optimized temperature parameters to achieve precise temperature environment assurance, as detailed below.

[0021] The first step, S1, involves deploying a matrix of high-precision temperature sensors at multiple points within each roller of the calender to construct a complete temperature field distribution information acquisition network. This network forms a closed-loop information flow, from sensor physical layout and signal acquisition and transmission to data validity verification and two-dimensional temperature field reconstruction, providing a data foundation for subsequent intelligent control modules. The specific implementation includes the following sub-steps.

[0022] Step S101: Multiple temperature sensors are deployed in each roll of the calender. The sensors are arranged in a crisscross pattern along the axial and circumferential directions of the roll, forming a sensor matrix. The deployment scheme is determined according to the specific dimensions of the roll. Taking a large calender with a roll length of 2000 mm and a diameter of 600 mm as an example, 8 to 12 sensor points are arranged axially, and 4 to 6 sensor points are arranged circumferentially, for a total of 32 to 72 temperature sensors deployed on a single roll.

[0023] The axial spacing of the sensors is set according to the temperature field gradient. In edge effect areas with drastic temperature changes, such as at both ends of the roller, the sensor points are densely arranged. In the circumferential direction, the sensors are preferentially placed at key positions corresponding to heating pipes or cooling channels, as well as at transitional positions between different heating and cooling zones.

[0024] The axial position, circumferential angle, and radial depth coordinates of each sensor point are precisely recorded in the configuration file as the basic parameters for subsequent spatial interpolation calculations of the temperature field.

[0025] In step S102, a high-precision platinum resistance temperature sensor or thermocouple sensor is selected as the temperature sensor, with a temperature measurement accuracy ranging from ±0.1℃ to ±0.3℃. The sensor sampling frequency is set between 10 and 50 times per second to capture rapid temperature changes.

[0026] To withstand the high-temperature and high-pressure working environment inside the calender, the sensor probe is encapsulated with special corrosion-resistant and wear-resistant alloy materials such as Hastelloy or tantalum alloy. The sensor signal transmission wires are insulated with high-temperature resistant polytetrafluoroethylene (PTFE) material, and a shielding layer is added to the outer layer of the wires to suppress electromagnetic interference in the production environment.

[0027] In step S103, the signals from the sensor matrix are collected to the data acquisition module via a bus or star topology. The data acquisition module integrates a multi-channel 24-bit high-precision analog-to-digital converter to convert the weak analog temperature signals into digital signals.

[0028] The analog-to-digital converter (ADC) reference voltage source uses a low-temperature drift precision voltage reference to ensure that the conversion accuracy is not affected by changes in ambient temperature. Before the analog-to-digital conversion, the signal conditioning circuit preprocesses the sensor signal, with the instrumentation amplifier amplifying the weak signal and the low-pass filter suppressing high-frequency noise.

[0029] In step S104, after digitization, the temperature data is transmitted in real time to the central control unit via industrial Ethernet or fieldbus using the UDP protocol. An added verification mechanism at the application layer ensures reliable data transmission. Upon receiving the data, the central control unit performs validity checks, eliminating abnormal data points caused by sensor malfunctions or signal interference.

[0030] The inspection criteria include: whether the temperature value exceeds the physical reasonable range, which is preset to 150℃ to 220℃ according to the PVC calendering process requirements; whether the temperature gradient between adjacent sensor points on the same roller exceeds the preset gradient threshold, which is set to 5℃ / cm; and whether the temperature change rate at adjacent sampling times exceeds the preset rate threshold, which is set to 10℃ / s.

[0031] Temperature data that passes all criteria tests are timestamped, stored in a data buffer, and await calls from subsequent modules.

[0032] Step S105: Based on the sensor data that has passed the validity check at a certain sampling time, the temperature field on the roller surface is reconstructed using a spatial interpolation algorithm. Multiple sensor measurements around the grid point to be calculated are selected, and inverse distance weighted interpolation or radial basis function interpolation methods are used to calculate the estimated temperature value of the grid point.

[0033] After calculating the temperature distribution at all grid points on the entire roller surface, two-dimensional temperature distribution maps in the axial and circumferential directions are generated. The temperature field update frequency is synchronized with the sensor sampling frequency to ensure that the temperature distribution information reflects the actual temperature state of the roller in real time.

[0034] In summary, step S1 establishes a data acquisition and processing flow from physical signal sensing to a digital temperature field model. This flow ensures that the temperature field data fed into the subsequent fuzzy neural network adaptive PID controller and deep reinforcement learning temperature optimization module meets high-quality control requirements in terms of accuracy, real-time performance, and spatial integrity. This provides reliable information input for the dynamic adjustment of PID parameters in step S2 and the autonomous learning of the temperature optimization strategy in step S3.

[0035] In step S2, a fuzzy neural network adaptive PID controller is constructed, integrating the reasoning ability of fuzzy logic with the learning ability of neural networks to achieve online adaptive adjustment of PID parameters. The controller consists of four processing layers: an input layer, a fuzzification layer, a fuzzy inference layer, and an output layer. The specific implementation includes the following sub-steps.

[0036] Step S201: Determine the composition of the operating parameter signals of the input layer. The operating parameter signals received by the input layer include five input variables: temperature deviation, temperature deviation change rate, raw material viscosity, calendering speed, and roller pressure.

[0037] Temperature deviation is defined as the difference between the actual temperature of the roller and the set temperature, and the calculation formula is:

[0038] in, for Temperature deviation at any time for Real-time roller temperature for Set the roller temperature at all times.

[0039] The rate of change of temperature deviation is defined as the derivative of the temperature deviation with respect to time, and the calculation formula is:

[0040] in, This represents the temperature deviation at the previous sampling time. The sampling period.

[0041] Raw material viscosity is measured in real time using an online viscometer, covering the viscosity range of PVC materials at processing temperatures. Calendering speed is measured in real time using an encoder, reflecting the actual linear speed of the calender. Inter-roll pressure is measured in real time using a pressure sensor, reflecting the actual pressure value between adjacent rolls.

[0042] Step S202: Preprocessing is performed on the input signal. The input layer performs debiasing and normalization processing on the five operating condition parameter signals.

[0043] The debiasing process removes the DC component from each input signal, bringing its mean to zero. The normalization process maps each input signal to a uniform numerical range, with the mapping interval being... or This ensures that signals of different dimensions and magnitudes have equal importance in subsequent processing. Normalization employs a linear mapping method, with mapping parameters determined based on the historical statistical values ​​of each input signal, including the mean and standard deviation.

[0044] Taking temperature deviation as an example, first calculate the average of historical temperature deviation data. and standard deviation Then, the current temperature deviation After normalization, it transforms into:

[0045] This is the normalized temperature deviation. The remaining input variables are processed in the same way.

[0046] Step S203: In the fuzzification layer, the continuous operating condition parameters are converted into fuzzy linguistic variables. The fuzzification layer determines the fuzzy set partitioning of each input variable.

[0047] Taking temperature deviation as an example, it is divided into 7 fuzzy subsets: negative large NB, negative medium NM, negative small NS, zero ZO, positive small PS, positive medium PM, and positive large PB. The membership functions of the fuzzy subsets adopt Gaussian membership functions or triangular membership functions. The expression for the Gaussian membership function is:

[0048] in, The normalized value of the input variable. The membership function center value, For width parameter, Input variables Membership degree to fuzzy subsets. The membership function parameters are determined through offline learning based on expert experience or historical data.

[0049] The temperature deviation change rate is also divided into 7 fuzzy subsets, set in the same way as the temperature deviation. Raw material viscosity is divided into 3 fuzzy subsets based on its numerical value: low viscosity, medium viscosity, and high viscosity. Calendering speed is divided into 3 fuzzy subsets based on its magnitude relative to the rated speed: low speed, normal speed, and high speed. Roller pressure is divided into 3 fuzzy subsets based on its numerical value: low pressure, normal pressure, and high pressure.

[0050] The output of the fuzzification layer is the membership value of each input variable to each fuzzy subset, which constitutes the input vector of the fuzzy inference layer. Taking temperature deviation as an example, if the current normalized temperature deviation is 0.3, after processing by the fuzzification layer, a 7-dimensional vector is output, which represents the membership degree of the normalized temperature deviation to the 7 fuzzy subsets NB, NM, NS, ZO, PS, PM, and PB.

[0051] Step S204: Construct a multi-layer fully connected neural network in the fuzzy inference layer to realize the fuzzy inference function. The fuzzy inference layer contains multiple hidden layers, the number of which is determined according to the network capacity and computational complexity, typically 2 to 4 layers.

[0052] The number of neurons in each hidden layer is determined by the dimension of the input vector and the inference complexity, typically ranging from 32 to 128. The activation function for the hidden layer neurons is either the ReLU function or the tanh function. The ReLU function expression is:

[0053] in, The tanh function is used to weight the input values ​​for the neuron.

[0054] in, Weighted input values ​​for neurons.

[0055] The input to the fuzzy inference layer is the membership vector output by the fuzzification layer. After nonlinear transformation by the hidden layer, the output is an inference result vector. The dimension of the inference result vector matches the dimension of the output layer, corresponding to the three output quantities of the PID controller: proportional parameter adjustment, integral parameter adjustment, and derivative parameter adjustment.

[0056] Step S205: Perform offline pre-training on the fuzzy inference layer. The training data is constructed as follows: extract the operating condition parameter sequence from the historical production database as the training input sample, and extract the corresponding PID parameter adjustment sequence as the training output label.

[0057] PID parameter adjustment labels are obtained through manual tuning or system identification methods. In manual tuning, process experts label the PID parameter adjustments to be applied under each operating condition based on historical temperature profiles and product quality data. In system identification, recursive least squares or particle swarm optimization algorithms are used to calculate the PID parameter adjustments for each operating condition, with the goal of minimizing temperature control error.

[0058] The offline pre-training loss function uses the mean squared error (MSE), and its expression is:

[0059] in, The number of samples in the training batch. For the first The network prediction value of the proportional parameter adjustment for each sample For the first The percentage adjustment label value for each sample and These are the predicted value of the integral parameter adjustment and the label value, respectively. and These are the predicted value of the differential parameter adjustment and the label value, respectively.

[0060] Offline pre-training employs either backpropagation or Adam optimization to minimize the mean squared error loss function, iterating until the loss function converges. The training objective is to ensure that the fuzzy inference layer parameters can cover the PID parameter adjustment patterns under typical operating conditions. In the backpropagation algorithm, the initial learning rate is set to 0.001, and dynamically adjusted after each iteration based on the loss reduction. In the Adam optimization algorithm, the first-order moment decay factor... Set to 0.9, second-order moment decay factor Set it to 0.999.

[0061] Step S206: Perform online fine-tuning of the fuzzy inference layer. Online fine-tuning involves making small adjustments to the parameters of the fuzzy inference layer based on real-time operating data, enabling the controller to adapt to newly emerging operating characteristics.

[0062] Online fine-tuning employs an incremental learning approach, updating network parameters only with a small number of newly acquired samples each time to avoid catastrophic overwriting of historical knowledge. The number of samples typically ranges from 8 to 32. The update magnitude of online fine-tuning is controlled by the learning rate parameter, which is set to 0.001 to 0.01. Online fine-tuning also uses the mean squared error loss function, but only calculates the loss and updates parameters for the most recent batch of samples, without traversing all historical data.

[0063] Step S207: Calculate the PID parameter adjustment at the output layer and apply boundary constraints. A linear activation function is used for the output layer, directly outputting the proportional parameter adjustment. Integral parameter adjustment amount and differential parameter adjustment amount .

[0064] The real-time PID parameter equals the sum of the basic PID parameter and the adjustment amount, calculated using the following formula:

[0065]

[0066]

[0067] in, For real-time scaling parameters, For real-time integration parameters, For real-time differential parameters, Based on the basic proportional parameters, Based on the basic integration parameters, These are the basic differential parameters.

[0068] Set parameter boundary constraints after the output layer. Scale parameters. The range of variation is limited by the basic proportional parameter 0.5 to 2.0 times. Integral parameter The range of variation is limited by the basic integral parameter 0.3 to 1.5 times.

[0069] Differential parameters The range of variation is limited by the basic differential parameters The value is 0.2 to 2.5 times the baseline proportional parameter. When the calculated PID parameters exceed the boundary constraints, they are trimmed to the nearest boundary value. For example, assuming the baseline proportional parameter... If it is 10, then the real-time scaling parameter The allowable range is 5 to 20, if the calculation yields... If the value is 22, then the cutoff value is 20.

[0070] In summary, step S2 constructs a complete fuzzy neural network adaptive PID controller through input layer signal acquisition and preprocessing, fuzzification layer linguistic variable transformation, fuzzy inference layer neural network construction and training, and output layer parameter generation and constraint. The real-time PID parameters output by the controller serve as the core control law parameters for the closed-loop temperature control in step S4, working in conjunction with the optimized setpoint provided by the deep reinforcement learning temperature optimization module in step S3 to achieve precise control of the calender roll temperature.

[0071] In step S3, a deep reinforcement learning temperature optimization module is established. An actor-critic architecture is used to construct a mapping model between temperature parameters and product quality indicators. Through continuous interaction with the production environment, the module autonomously learns temperature control strategies. The specific implementation includes the following sub-steps.

[0072] Step S301 defines the state space and action space of the deep reinforcement learning temperature optimization module. The state space consists of a multi-dimensional feature vector composed of roller temperature field data, product quality indicators, and historical sequences of control parameters. Roller temperature field data includes the temperature values ​​at each sensor point output in step S1 and the temperature field interpolation results. Product quality indicators include real-time measurements of PVC film thickness uniformity and film surface flatness. The historical sequence of control parameters contains the PID parameters and temperature deviations for the past N control cycles, where N is determined based on the rate of change of the operating conditions. The dimension of the state space vector depends on the number of sensors, the length of the historical sequence, and the types of product quality indicators.

[0073] The action space is defined as a multi-dimensional PID parameter adjustment vector, including proportional parameter adjustment, integral parameter adjustment, and derivative parameter adjustment, with a dimension of 3. Each component of the action vector is first processed by the tanh activation function to limit the output to the interval [-1, 1], and then mapped to the actual action range through a linear transformation.

[0074] Step S302: Construct the actor network and the critic network. The actor network adopts a long short-term memory network structure, which includes long short-term memory layers and fully connected layers. The input is the state space vector defined in step S301, and the output is the action space vector, which is used to learn the mapping relationship from state to action.

[0075] The critic network also employs a Long Short-Term Memory (LSTM) network structure, containing LSM layers and fully connected layers. Its output dimension is 1, representing the value scalar of the state-action pair. The critic network takes the state space vector and the action vector output by the actor network as joint inputs, outputting a value function estimate to guide the actor network's policy updates.

[0076] Step S303, define the reward function. The reward function is used to evaluate the merits of the current temperature control strategy, and its expression is:

[0077] in, for Reward value at any moment Weighting for product quality improvement For temperature stability weighting, and Typical values ​​are 0.6 to 0.8 and 0.2 to 0.4, respectively. For the improvement of product quality indicators, This represents the change in temperature deviation. Its absolute value reflects the stability of temperature control.

[0078] Taking into account both thickness uniformity and film surface smoothness, the expression is:

[0079] in, and These are the weights for thickness improvement and flatness improvement, respectively, to meet the following requirements. .

[0080] This is the difference between the thickness standard deviation of the current adjustment period and the previous adjustment period. This represents the difference in surface roughness between the current adjustment cycle and the previous adjustment cycle. When product quality improves... A positive value indicates a deterioration, while a negative value indicates a worsening situation. The smaller the temperature fluctuation, the better. The smaller the value, the less the reward will be deducted.

[0081] Step S304: Set the training objective and update mechanism for the deep reinforcement learning temperature optimization module. The temperature parameter adjustment period for the deep reinforcement learning temperature optimization module is set to 30 seconds to 300 seconds. At the end of each period, the module updates the network based on the accumulated interaction experience within that period.

[0082] The update objective of the actor network is to maximize the expected cumulative reward, and the network parameters are updated through policy gradient ascent. The update objective of the critic network is to minimize the temporal difference error, and the loss function adopts the mean squared error form. The target value is obtained by weighting the immediate reward and the target network's value estimate for the next state by a discount factor, which typically ranges from 0.9 to 0.99.

[0083] Each update of the actor network and critic network involves randomly drawing a batch of samples from the experience replay buffer to break the temporal correlation between samples. The experience replay buffer stores the interaction experience at each time step in the form of quadruples, and the buffer capacity is set to match the scale of the production data. A target network mechanism is also employed, maintaining target network copies for both the actor and critic networks. Every preset number of steps, the parameters of the main network are copied to the target network to avoid policy oscillations during training.

[0084] Step S305: Execute online optimization decision-making. The deep reinforcement learning temperature optimization module continuously performs incremental updates according to the temperature parameter adjustment cycle. After completing the parameter update, the module calculates the PID parameter adjustment amount through forward propagation of the actor network based on the current state space vector, which serves as the optimization setpoint for the adaptive PID controller in step S4.

[0085] In summary, step S3 establishes a deep reinforcement learning temperature optimization module capable of online learning by defining the state-action space, constructing an actor-critic dual network structure, designing a multi-objective reward function, and deploying an experience playback and target network mechanism. The PID parameter adjustment output by the module works in conjunction with the fuzzy neural network adaptive PID controller from step S2, autonomously learning temperature control strategies through continuous interaction, driving the roller temperature towards optimization to improve product quality.

[0086] In step S4, closed-loop temperature control is performed, combining the optimization results from the deep reinforcement learning temperature optimization module in step S3 with the fuzzy neural network adaptive PID controller in step S2 to form a complete online control closed loop. The specific implementation includes the following sub-steps.

[0087] Step S401: Obtain the optimized setpoints for the deep reinforcement learning temperature optimization module. At the end of each temperature parameter adjustment cycle, the deep reinforcement learning temperature optimization module uses the proportional parameter adjustment, integral parameter adjustment, and derivative parameter adjustment output in step S305 as optimized setpoints and provides them to the fuzzy neural network adaptive PID controller.

[0088] Step S402: Real-time acquisition of roller temperature and calculation of control deviation. The temperature sensor matrix acquires the actual surface temperature of each roller in real time at a preset sampling frequency. The sampling frequency is synchronized with the clock of the central control unit. After receiving the actual temperature value, the central control unit calculates the temperature deviation at the current moment. and the rate of change of temperature deviation The calculation formula is:

[0089]

[0090] in, for The roller is set to a specific temperature. for Real-time roller temperature This represents the temperature deviation at the previous sampling time. The sampling period is [period]. The calculated [period] is [period]. and As the input signal for the fuzzy neural network adaptive PID controller.

[0091] Step S403: Execute fuzzy neural network adaptive PID control. The fuzzy neural network adaptive PID controller receives... , In addition to raw material viscosity, calendering speed, and inter-roll pressure parameters, the input layer preprocessing, fuzzification layer conversion, fuzzy inference layer inference, and output layer generation processes constructed in step S2 are used to calculate real-time PID control parameters based on the optimized setpoints provided in step S401, along with raw material viscosity, calendering speed, and inter-roll pressure parameters. , , .

[0092] Substitute the real-time PID control parameters into the discretized PID control law to calculate the control signal. The expression for the discrete PID control law is:

[0093] in, for Constantly control the output value of the signal. Real-time scaling parameters For real-time integration parameters, For real-time differential parameters, for Temperature deviation over time This is the cumulative sum of temperature deviations from the start of control to the current time. The sampling period is for Rate of change of temperature deviation over time. Control signal. The output range is 0 to 100%, corresponding to the full range adjustment range of the actuator.

[0094] Step S404: Convert the control signal into actuator action. The signal is converted into an analog voltage or analog current signal by a digital-to-analog converter and transmitted to the electrically controlled regulating valve of the roller heating or cooling device. The electrically controlled regulating valve adjusts the heating power or cooling water flow rate according to the control signal.

[0095] The heating power is adjusted by controlling the on / off ratio of the heating element, and the cooling water flow rate is adjusted by controlling the opening of the regulating valve.

[0096] The actuator changes the heating or cooling state of the roller according to the control signal, so that the surface temperature of the roller approaches the set temperature. When When the actual temperature is higher than the set temperature, the fuzzy neural network adaptive PID controller outputs a reduction in heating power or an increase in cooling water flow; when When the actual temperature is lower than the set temperature, the fuzzy neural network adaptive PID controller outputs either an increase in heating power or a decrease in cooling water flow.

[0097] Step S405: Set the control timing parameters. The sampling period for the closed-loop control process is set to 100 milliseconds to 500 milliseconds, and the control period is set to 1 second to 5 seconds. The sampling period determines the frequency of temperature data acquisition and deviation calculation, while the control period determines the frequency of control signal updates and actuator actions.

[0098] A shorter sampling period can capture rapid temperature changes, and a shorter control period can improve the controller's response speed. The combination of the two determines the speed and stability of the control response.

[0099] The temperature parameter adjustment period is set to 30 to 300 seconds in step S3. The deep reinforcement learning temperature optimization module updates the PID parameter adjustment setpoint at the end of each adjustment period, and the fuzzy neural network adaptive PID controller uses the updated setpoint for continuous closed-loop adjustment within the adjustment period.

[0100] Step S406: Execute anomaly detection and protection; configure anomaly detection and protection mechanisms for the closed-loop control process. When temperature deviation... When the preset safety threshold is exceeded, an alarm signal is issued and an emergency response procedure is initiated. The safety threshold is set according to process safety requirements, for example, ±15℃.

[0101] When a sensor malfunction or communication interruption is detected, the fuzzy neural network adaptive PID controller automatically switches to a preset safety mode to maintain safe production line operation with fixed safety PID parameters. When an actuator malfunctions, the fuzzy neural network adaptive PID controller outputs a fault indication signal and records a fault log for subsequent maintenance and analysis.

[0102] In summary, step S4 combines the optimized setpoint output by the deep reinforcement learning temperature optimization module in step S3 with the real-time deviation correction capability of the fuzzy neural network adaptive PID controller in step S2 to construct a closed-loop temperature control architecture of "upper-level optimization decision-lower-level rapid execution". Through periodic updates of the optimized setpoint, real-time feedback correction of deviations, and continuous adjustment of the actuator, the temperature fluctuation of the roller surface is controlled within a preset accuracy range, providing a precise temperature environment guarantee for the PVC film calendering production in step S5.

[0103] Finally, in step S5, based on the optimized temperature parameters generated collaboratively from steps S1 to S4, PVC film calendering production is performed. Through precise temperature environment assurance, online quality inspection, and closed-loop feedback, high-quality film forming is achieved. This process includes the following sub-steps.

[0104] Step S501: Configure the multi-stage calendering unit and set the roller temperature parameters. The PVC film calendering production line consists of multiple calenders arranged in series to form a multi-stage calendering unit. Each calender contains two or more rollers, with calendering gaps formed between the rollers. The PVC material is gradually formed into a film under compression and stretching in the calendering gaps. Taking a four-roll calender as an example, the four rollers are arranged in the following order according to the material flow direction: feed roller, second roller, third roller, and discharge roller.

[0105] The feed roller is set to a temperature below the material's softening temperature for initial compaction. The second and third rollers are set within the material's plasticizing temperature range to ensure thorough plasticization and uniform mixing. The discharge roller is set to a temperature below the second and third rollers to control the final film shape and facilitate cooling and setting. The temperature settings for each roller are determined based on the PVC formulation and product specifications, typically ranging from 150°C to 220°C.

[0106] Step S502: Perform closed-loop temperature control and online product quality detection. During the calendering process, the fuzzy neural network adaptive PID controller (step S2) continuously monitors the actual temperature of each roller. Step S4, through closed-loop temperature control, keeps the roller surface temperature fluctuation within a preset accuracy range, set to ±1℃ to ±2℃. Step S3: The deep reinforcement learning temperature optimization module evaluates the effect of the current temperature parameters according to the temperature parameter adjustment cycle and adjusts the PID parameter settings based on product quality feedback.

[0107] Product quality is monitored in real time through an online testing system, which includes two dimensions: thickness uniformity and film surface flatness.

[0108] Thickness uniformity is measured in real time using an online laser thickness gauge. The laser thickness gauge has multiple measurement points arranged along the width of the film, simultaneously measuring the thickness at different locations. The measurement accuracy reaches ±0.1 μm to ±0.5 μm, and the measurement frequency exceeds 1000 times per second. Thickness standard deviation... The control target was obtained through statistical calculations; it was the standard deviation of thickness. Less than the preset thickness threshold, which is set to 1 micrometer to 3 micrometers.

[0109] The flatness of the membrane surface is assessed by a vision inspection system that acquires images of the membrane surface in real time and calculates surface roughness. This system comprises a high-resolution industrial camera, an LED light source, and an image processing unit. The industrial camera has a resolution of 5 to 20 megapixels and a shooting speed exceeding 100 frames per second. The LED light source provides uniform illumination, ensuring consistent image quality. The image processing unit performs preprocessing on the acquired membrane surface images, including noise suppression and geometric correction, before calculating surface roughness characteristic parameters to obtain the root mean square roughness. and average roughness The control target is surface roughness. and All are less than the preset roughness threshold, which is set to 0.1 micrometers to 0.5 micrometers.

[0110] Step S503: Feedback the product quality data to the deep reinforcement learning temperature optimization module. Step S502: Real-time detection of the thickness standard deviation. Root mean square roughness average roughness As the input data for product quality indicators, the deep reinforcement learning temperature optimization module in step S3 participates in the reward signal calculation according to the reward function defined in step S303, driving the strategy update of the deep reinforcement learning temperature optimization module.

[0111] Step S504 assesses the impact of temperature stability on product quality. The surface temperature stability of the rollers directly affects the crystal orientation consistency of the PVC film. When the roller temperature fluctuation is within ±1℃, the molecular chain orientation of the PVC material remains stable during plasticization, resulting in good film thickness uniformity and consistent mechanical properties. When the roller temperature fluctuation exceeds ±2℃, changes in material viscosity lead to inconsistent molecular chain orientation, causing film thickness fluctuations and wavy defects on the film surface.

[0112] The correlation between temperature stability and product quality is gradually established through continuous trial and error learning in the temperature optimization module using deep reinforcement learning. As production data accumulates, the accuracy of the mapping model between temperature parameters and product quality indicators continues to improve.

[0113] Step S5 applies the intelligent control system built in the previous steps to the actual rolling production process. Through the closed-loop interaction between precise temperature control and real-time quality detection, product quality data continuously feeds back into the strategy updates of the deep reinforcement learning temperature optimization module. The fuzzy neural network adaptive PID controller and the deep reinforcement learning temperature optimization module work together to maintain consistent temperature control accuracy and product quality over long-term operation, reducing the scrap rate compared to traditional fixed PID parameter control methods.

[0114] On the other hand, the present invention provides a PVC film production system based on intelligent control, comprising: The temperature field acquisition network consists of a matrix of multi-point temperature sensors deployed inside each roll of the calender, used to acquire temperature data at multiple locations along the axial and circumferential directions of the rolls in real time. The fuzzy neural network adaptive PID controller is used to receive operating condition parameter signals and output real-time PID control parameters. The fuzzy neural network adaptive PID controller includes an input layer, a fuzzification layer, a fuzzy inference layer, and an output layer. The deep reinforcement learning temperature optimization module adopts an actor-critic architecture to output PID parameter adjustment as the optimization setpoint based on the state space consisting of roller temperature field data, product quality indicators and historical sequences of control parameters. The closed-loop temperature control execution module is used to receive the optimized setpoint and real-time PID control parameters, and calculate the control signal by combining the real-time temperature deviation to adjust the actuator of the roller heating or cooling device. The multi-stage calendering production unit contains multiple rollers with different temperature settings for performing PVC film calendering production under optimized temperature parameters; The online quality inspection system is used to acquire PVC film thickness uniformity data and film surface flatness data in real time, and feed the thickness uniformity data and film surface flatness data back to the deep reinforcement learning temperature optimization module.

[0115] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0116] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for producing PVC film based on intelligent control, characterized in that, include: By deploying a multi-point temperature sensor matrix inside each roll of the calender, temperature data at multiple locations along the axial and circumferential directions of the rolls can be acquired in real time to construct a temperature field distribution information acquisition network. A fuzzy neural network adaptive PID controller is constructed, and the real-time acquired operating condition parameter signals are input into the fuzzy neural network adaptive PID controller to output real-time PID control parameters. A deep reinforcement learning temperature optimization module is established. The deep reinforcement learning temperature optimization module adopts an actor-commentator architecture. Based on the state space consisting of roller temperature field data, product quality indicators and historical sequence of control parameters, it outputs PID parameter adjustment as optimization setpoint. Closed-loop temperature control is implemented by providing the optimized setpoint output by the deep reinforcement learning temperature optimization module to the fuzzy neural network adaptive PID controller. The fuzzy neural network adaptive PID controller combines the real-time temperature deviation to calculate the control signal, thereby adjusting the actuator of the roller heating or cooling device. The PVC film calendering production is carried out based on optimized temperature parameters, and the thickness uniformity data and film surface flatness data of the PVC film are acquired in real time through an online quality inspection system. The thickness uniformity data and film surface flatness data are fed back to the deep reinforcement learning temperature optimization module to update the strategy of the deep reinforcement learning temperature optimization module.

2. The PVC film production method based on intelligent control according to claim 1, characterized in that, A multi-point temperature sensor matrix is ​​deployed inside each roller of the calender, specifically including: Eight to 12 sensor points are arranged axially on each roller of the calender, and four to six sensor points are arranged circumferentially. A total of 32 to 72 temperature sensors are deployed on a single roller. The temperature sensors are arranged in a cross distribution along the axial and circumferential directions of the roller, and the sensor points are densely arranged in the edge effect area at both ends of the roller. The temperature sensor has a temperature measurement accuracy of ±0.1℃ to ±0.3℃ and a sampling frequency of 10 to 50 times per second. Based on sensor data that has passed the validity test, a spatial interpolation algorithm is used to reconstruct the temperature field on the roller surface, generating two-dimensional temperature distribution maps in the axial and circumferential directions, and the temperature field update frequency is synchronized with the sensor sampling frequency.

3. The PVC film production method based on intelligent control according to claim 1, characterized in that, Constructing a fuzzy neural network adaptive PID controller specifically includes: The fuzzy neural network adaptive PID controller consists of an input layer, a fuzzification layer, a fuzzy inference layer, and an output layer. The input layer receives five input variables: temperature deviation, temperature deviation change rate, raw material viscosity, calendering speed, and inter-roll pressure. The input layer performs debiasing and normalization on the five input variables, mapping them to a uniform numerical range. A parameter boundary constraint mechanism is set after the output layer, wherein the variation range of the proportional parameter is limited to 0.5 times to 2.0 times the basic proportional parameter, the variation range of the integral parameter is limited to 0.3 times to 1.5 times the basic integral parameter, and the variation range of the derivative parameter is limited to 0.2 times to 2.5 times the basic derivative parameter.

4. The PVC film production method based on intelligent control according to claim 3, characterized in that, The fuzzy inference layer uses a multi-layer fully connected neural network to implement fuzzy inference functionality, and performs offline pre-training and online fine-tuning of the fuzzy inference layer; Online fine-tuning employs an incremental learning method, updating network parameters each time using only the latest 8 to 32 samples, with a learning rate set to 0.001 to 0.

01.

5. The PVC film production method based on intelligent control according to claim 1, characterized in that, In the actor-critic architecture of the deep reinforcement learning temperature optimization module, both the actor network and the critic network adopt a long short-term memory network structure. The training objective of the deep reinforcement learning temperature optimization module is to maximize the expected cumulative reward. The reward function of the deep reinforcement learning temperature optimization module is composed of a weighted sum of the product quality improvement and the temperature deviation change, where the weight of product quality improvement is 0.6 to 0.8 and the weight of temperature stability is 0.2 to 0.

4. The improvement in product quality is obtained by weighted summation of the improvement in thickness uniformity and the improvement in film surface smoothness.

6. The PVC film production method based on intelligent control according to claim 1 or 5, characterized in that, During the update process, the deep reinforcement learning temperature optimization module randomly selects batches of samples from the experience replay buffer and adopts a target network mechanism to maintain target network copies of the actor network and the critic network respectively. Every preset number of steps, the parameters of the main network are copied to the target network copies.

7. The PVC film production method based on intelligent control according to claim 1, characterized in that, The temperature parameter adjustment cycle of the deep reinforcement learning temperature optimization module is set to 30 seconds to 300 seconds. At the end of each adjustment cycle, the deep reinforcement learning temperature optimization module updates the PID parameter adjustment amount as the optimization setpoint. The sampling period for closed-loop temperature control is set to 100 milliseconds to 500 milliseconds, and the control period is set to 1 second to 5 seconds.

8. The PVC film production method based on intelligent control according to claim 1, characterized in that, Implementing closed-loop temperature control also includes configuring anomaly detection and protection mechanisms: When the temperature deviation exceeds the preset safety threshold, an alarm signal is issued and the emergency response procedure is activated; When a sensor malfunction or communication interruption is detected, the fuzzy neural network adaptive PID controller automatically switches to a safe mode that operates with fixed safe PID parameters.

9. The PVC film production method based on intelligent control according to claim 1, characterized in that, The production of PVC film calendering includes: The unit is configured with a multi-stage calendering unit, wherein the temperature of the feed roller is set below the material softening temperature, the temperatures of the second and third rollers are set within the material plasticizing temperature range, and the temperature of the discharge roller is set below the temperatures of the second and third rollers. The online quality inspection system measures the film thickness in real time using a laser online thickness gauge with a measurement accuracy of ±0.1 micrometer to ±0.5 micrometer. The online quality inspection system also uses a vision inspection system to acquire film surface images in real time to calculate surface roughness.

10. A PVC film production system based on intelligent control, applied to the method described in any one of claims 1-9, characterized in that, include: The temperature field acquisition network consists of a matrix of multi-point temperature sensors deployed inside each roll of the calender, used to acquire temperature data at multiple locations along the axial and circumferential directions of the rolls in real time. The fuzzy neural network adaptive PID controller is used to receive operating condition parameter signals and output real-time PID control parameters. The fuzzy neural network adaptive PID controller includes an input layer, a fuzzification layer, a fuzzy inference layer, and an output layer. The deep reinforcement learning temperature optimization module adopts an actor-critic architecture to output PID parameter adjustment as the optimization setpoint based on the state space consisting of roller temperature field data, product quality indicators and historical sequences of control parameters. The closed-loop temperature control execution module is used to receive the optimized setpoint and real-time PID control parameters, and calculate the control signal by combining the real-time temperature deviation to adjust the actuator of the roller heating or cooling device. The multi-stage calendering production unit contains multiple rollers with different temperature settings for performing PVC film calendering production under optimized temperature parameters; The online quality inspection system is used to acquire PVC film thickness uniformity data and film surface flatness data in real time, and feed the thickness uniformity data and film surface flatness data back to the deep reinforcement learning temperature optimization module.

Images