Isoprene rubber polymerization reaction control method and system

Abstract

The present invention relates to the technical field of data processing, and discloses a method and system for controlling an isoprene rubber polymerization reaction. The method and system comprise the following steps: collecting real-time reaction data and real-time image data of the isoprene rubber polymerization reaction; extracting features of the real-time reaction data and the real-time image data based on a pre-trained deep learning model to obtain time series features and image features; fusing the time series features and the image features based on a cross-modal attention mechanism to obtain a multi-modal joint representation; processing the multi-modal joint representation through a lightweight pre-trained model embedded with Arrhenius equation constraints to output a reaction prediction result; generating and executing a reaction control strategy based on the reaction prediction result through a fuzzy control method and an anti-saturation integral PID control algorithm to achieve control of the isoprene rubber polymerization reaction; and achieving effective control of the isoprene rubber polymerization reaction by deeply fusing multi-source data with a variety of advanced algorithms.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, in particular to an isoprene rubber polymerization reaction control method and system. Background Art

[0002] Traditional industrial production of isoprene rubber relies primarily on solution polymerization. However, as the viscosity of the reaction materials gradually increases during the polymerization process, the resulting high-viscosity intermediate or final products aggregate, affecting the utilization rate of the raw materials and reducing reaction efficiency. Furthermore, high-viscosity intermediates or final products tend to adhere to the inner wall of the reaction vessel. Furthermore, the polymerization reaction is exothermic, which can cause the temperature inside the reaction vessel to be too high, affecting the quality of the final product and the utilization rate of the raw materials. In severe cases, it can cause the reaction vessel temperature to overheat, leading to dangerous accidents.

[0003] To address this phenomenon, industrial cameras are currently commonly used to capture real-time reaction images of isoprene rubber polymerization reactions. After simple processing, these images are input into a machine learning model for recognition and processing to determine the degree of reaction. The control strategy is then generated based on the experience of professionals, and the reaction process is adjusted in real time. However, simple image processing can easily lead to low accuracy in subsequent model recognition, resulting in the generated control strategy being unable to improve the reaction efficiency of the isoprene rubber polymerization reaction.

[0004] Therefore, how to effectively control the polymerization reaction of isoprene rubber and thereby improve its reaction efficiency and safety has become a technical problem to be solved urgently by those skilled in the art. Summary of the Invention

[0005] The present invention provides an isoprene rubber polymerization reaction control method and system, which solve the problem of how to effectively control the isoprene rubber polymerization reaction and thus improve the reaction efficiency and safety.

[0006] In order to solve the above technical problems, the first aspect of the present invention provides an isoprene rubber polymerization reaction control method, comprising:

[0007] Collecting real-time reaction data and real-time image data of isoprene rubber polymerization reaction;

[0008] Based on the pre-trained deep learning model, feature extraction is performed on the real-time reaction data and the real-time image data to obtain time series features and image features;

[0009] Based on the cross-modal attention mechanism, the temporal features and the image features are fused to obtain a multimodal joint representation;

[0010] Processing the multimodal joint representation through a lightweight pre-trained model embedded with Arrhenius equation constraints to output a response prediction result;

[0011] According to the reaction prediction result, a reaction control strategy is generated and executed through a fuzzy control method and an anti-saturation integral PID control algorithm to achieve control of the isoprene rubber polymerization reaction.

[0012] As one preferred solution, the pre-trained deep learning model extracts features from the real-time reaction data and the real-time image data respectively to obtain time series features and image features, including:

[0013] extracting the temperature change rate and pressure fluctuation amplitude from the real-time reaction data using a pre-trained long short-term memory network to generate time series features;

[0014] The real-time image data is processed by a pre-trained convolutional neural network and image segmentation technology to obtain image features.

[0015] As one preferred solution, the real-time image data is processed by a pre-trained convolutional neural network and image segmentation technology to obtain image features, including:

[0016] Performing grayscale and binarization processing on the real-time image data to obtain a real-time processed image;

[0017] Using a shallow convolution kernel of the convolutional neural network to extract color distribution features of the real-time processed image to obtain a color histogram;

[0018] Extracting texture roughness of the real-time processed image through a middle convolutional layer of the convolutional neural network, and processing the texture roughness based on a deep convolutional layer of the convolutional neural network to obtain texture features;

[0019] Calculating the colloid displacement field between adjacent frames in the real-time processed image using an optical flow method to generate a dynamic viscosity change feature vector;

[0020] Detecting bubbles in the real-time processed image using a pre-trained segmentation model to obtain bubble features;

[0021] The color histogram, the texture feature, the dynamic viscosity change feature vector and the bubble feature are integrated to obtain image features.

[0022] As one preferred solution, the bubbles in the real-time processed image are detected using a pre-trained segmentation model to obtain bubble features, including:

[0023] Using a pre-trained MobileNetV3 model to process the real-time processing image to obtain a bubble segmentation mask;

[0024] Performing hole filling, noise filtering, and edge refinement processing on the bubble segmentation mask to obtain a bubble connected domain;

[0025] Calculating the area, circularity and centroid coordinates of each bubble connected domain to construct the static characteristics of the bubble;

[0026] Based on the centroid coordinates, a SORT algorithm is used to associate bubbles in adjacent frames to obtain continuous coordinates of each bubble to quantify the kinematic parameters of each bubble;

[0027] Stokes' law is used to verify each of the bubbles to mark abnormal bubbles;

[0028] The area, the circularity, the kinematic parameters and the abnormal bubbles are integrated to obtain bubble characteristics.

[0029] As one of the preferred solutions, the cross-modal attention mechanism is used to fuse the temporal features and the image features to obtain a multimodal joint representation, including:

[0030] The temporal self-attention layer of the pre-trained Transformer is used to encode the temporal features to obtain a temporal feature vector;

[0031] Processing the image features using a local block of a pre-trained Vision Transformer to generate an image feature vector;

[0032] Based on a cross-modal attention mechanism, a bidirectional interaction is performed between the time series feature vector and the image feature vector, and a learning parameter matrix is ​​introduced in the interaction process to adjust the interaction strength to generate a first enhanced representation and a second enhanced representation;

[0033] The first enhanced representation and the second enhanced representation are aggregated through a gating mechanism, and a conditional routing mechanism is introduced in the aggregation process to dynamically select a fusion path based on the modal correlation between the first enhanced representation and the second enhanced representation to obtain a multimodal joint representation.

[0034] As one preferred solution, the multimodal joint representation is processed by a lightweight pre-trained model embedded with Arrhenius equation constraints to output a reaction prediction result, including:

[0035] An Arrhenius equation is constructed based on the isoprene rubber polymerization reaction, and temperature data in the real-time reaction data is input as an explicit feature into the Arrhenius equation to obtain an Arrhenius auxiliary feature;

[0036] The multimodal joint representation, the temperature data, and the Arrhenius auxiliary features are spliced ​​into an input tensor, and the input tensor is input into the lightweight pre-training model for processing, and a reaction conversion rate prediction value and a viscosity change rate prediction value for a future preset time period are output as the reaction prediction result.

[0037] As one preferred solution, based on the reaction prediction result, a reaction control strategy is generated and executed by a fuzzy control method and an anti-saturation integral PID control algorithm to achieve control of the isoprene rubber polymerization reaction, including:

[0038] Obtaining a reaction error corresponding to the reaction prediction result according to a transfer learning method, and decomposing the reaction error into a short-term error and a long-term error using a sliding window averaging method;

[0039] Constructing a dynamic weight function through the long-term error and the short-term error, and using the dynamic weight function to quantify the first contribution ratio and the second contribution ratio corresponding to the fuzzy control method and the anti-saturation integral PID control algorithm respectively;

[0040] Based on the reaction error, generating a first control signal and a second control signal by a fuzzy control method and an anti-windup integral PID control algorithm, and weighting them by the first contribution ratio and the second contribution ratio to obtain a total control signal;

[0041] According to the total control signal and the reaction error, the pulse width, pulse interval and pulse amplitude are adjusted by a gradient optimization algorithm, and the reaction control strategy is generated and executed to achieve control of the isoprene rubber polymerization reaction.

[0042] As one preferred solution, after processing the multimodal joint representation by a lightweight pre-trained model embedded with Arrhenius equation constraints and outputting a reaction prediction result, the method further includes:

[0043] According to the reaction prediction results, a pre-trained hierarchical architecture model is used to generate and execute a reaction control strategy to achieve control of the isoprene rubber polymerization reaction; wherein the hierarchical architecture model includes a bottom control layer, a middle control layer and a high-level control layer.

[0044] As one preferred solution, based on the reaction prediction results, a pre-trained hierarchical architecture model is used to generate and execute a reaction control strategy to achieve control of the isoprene rubber polymerization reaction, including:

[0045] A temperature controller is designed based on a PID algorithm to construct the bottom control layer, and a preset temperature upper limit is used as a first constraint. The deviation between the current temperature and the preset temperature is input into the bottom control layer for processing, and a control adjustment sequence is output;

[0046] The middle control layer is constructed based on a pre-trained LSTM neural network, the reaction prediction results are input into the middle control layer for processing, the predicted energy consumption is output, and a multi-objective optimization model is established with maximizing conversion rate, minimizing viscosity fluctuation, and minimizing energy consumption as the ultimate goals. The multi-objective optimization model is solved by a preset optimization algorithm with the dynamic balance of preset upper and lower limits of multiple reaction parameters and raw material feeding and discharging rates as the second constraint to obtain a prediction control sequence;

[0047] In the high-level control layer, a reward function is constructed based on the first constraint, the second constraint and the final goal, and a deep reinforcement learning algorithm is used to adjust the control adjustment sequence and the predictive control sequence, and a reaction control strategy is generated and executed to achieve control of the isoprene rubber polymerization reaction.

[0048] A second aspect of the present invention provides an isoprene rubber polymerization reaction control system, comprising:

[0049] A data acquisition module is used to collect real-time reaction data and real-time image data of the isoprene rubber polymerization reaction;

[0050] A feature extraction module is used to extract features from the real-time reaction data and the real-time image data based on a pre-trained deep learning model to obtain time series features and image features;

[0051] A feature fusion module, configured to fuse the temporal features and the image features based on a cross-modal attention mechanism to obtain a multimodal joint representation;

[0052] a reaction prediction module, configured to process the multimodal joint representation using a lightweight pre-trained model embedded with Arrhenius equation constraints and output a reaction prediction result;

[0053] The strategy execution module is used to generate and execute a reaction control strategy based on the reaction prediction result by using a fuzzy control method and an anti-saturation integral PID control algorithm to achieve control of the isoprene rubber polymerization reaction.

[0054] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0055] (1) Through the synergistic effect of real-time data and image data, the physical phenomena and dynamic changes of the isoprene rubber polymerization reaction are revealed. The deep learning model is used to extract and fuse the features of the reaction process to improve the comprehensiveness of the reaction characterization.

[0056] (2) Embedding the Arrhenius equation into the pre-trained model forces the model to follow physical laws during the prediction process, which improves the physical rationality of the prediction while maintaining high efficiency;

[0057] (3) Intelligent control strategies are used to improve the stability and efficiency of reaction control. Deep learning models and intelligent control algorithms are combined to achieve intelligent monitoring and effective control of the reaction process, reduce manual intervention, improve production efficiency and consistency, and further improve the safety and economy of production. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the implementation. Obviously, the drawings described below are only some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0059] Figure 1 This is a flow chart of a method for controlling an isoprene rubber polymerization reaction provided by one embodiment of the present invention;

[0060] Figure 2 This is a structural diagram of an isoprene rubber polymerization reaction control system provided by one embodiment of the present invention;

[0061] Reference numerals:

[0062] Among them, 10, data acquisition module; 20, feature extraction module; 30, feature fusion module; 40, reaction prediction module; 50, strategy execution module. DETAILED DESCRIPTION

[0063] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings and embodiments. Obviously, the embodiments described are only some embodiments of the present invention, rather than all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0064] In the description of this application, the terms "first," "second," "third," etc. are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first," "second," "third," etc. may explicitly or implicitly include one or more of the features. In the description of this application, unless otherwise specified, "plurality" means two or more.

[0065] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or an indirect connection through an intermediate medium, or it can be a communication between the two components. The terms "vertical", "horizontal", "left", "right", "up", "down" and similar expressions used herein are for illustrative purposes only, and do not indicate or imply that the system or component referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood according to the specific circumstances.

[0066] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this application have the same meanings as those commonly understood by those skilled in the art. The terms used in this specification are only for describing specific embodiments and are not intended to limit the present invention. Those skilled in the art will understand the specific meanings of the above terms in this application according to specific circumstances.

[0067] In one embodiment, if Figure 1 As shown, the first aspect of the present invention provides an isoprene rubber polymerization reaction control method, comprising:

[0068] S1, collecting real-time reaction data and real-time image data of isoprene rubber polymerization reaction;

[0069] Specifically, multiple sensors are installed in the reactor or kettle of the isoprene rubber polymerization reaction to collect real-time reaction data, such as time series data such as temperature, pressure, viscosity, stirring power, and reaction time. Real-time image data of the reaction is collected from multiple angles using a high-temperature resistant camera or industrial camera to capture the material's color, texture, transparency, bubbles, particle distribution, and wall adhesion. The collected data is cleaned and synchronized to ensure alignment of sensor data with image frame timestamps. Furthermore, to facilitate acquisition of real-time images and reaction data, the reactor or kettle used in this reaction is colorless or transparent.

[0070] S2. Extracting features from the real-time reaction data and the real-time image data based on a pre-trained deep learning model to obtain time series features and image features;

[0071] In one embodiment, step S2 includes:

[0072] extracting the temperature change rate and pressure fluctuation amplitude from the real-time reaction data using a pre-trained long short-term memory network to generate time series features;

[0073] The real-time image data is processed by a pre-trained convolutional neural network and image segmentation technology to obtain image features.

[0074] Specifically, the present invention calculates the temperature change rate and pressure fluctuation amplitude based on real-time reaction data collected in real time. This data is then normalized and fed into a pre-trained long-short-term memory network for processing to output time series features. The pre-trained convolutional neural network and image segmentation techniques are then used to segment the real-time image data to extract image features. This present invention processes the real-time reaction and image data using a deep learning model to capture the multi-dimensional dynamic characteristics of isoprene rubber polymerization reactions, reducing bias associated with a single data source and improving prediction accuracy.

[0075] In one embodiment, the real-time image data is processed using a pre-trained convolutional neural network and image segmentation technology to obtain image features, including:

[0076] Performing grayscale and binarization processing on the real-time image data to obtain a real-time processed image;

[0077] Using a shallow convolution kernel of the convolutional neural network to extract color distribution features of the real-time processed image to obtain a color histogram;

[0078] Extracting texture roughness of the real-time processed image through a middle convolutional layer of the convolutional neural network, and processing the texture roughness based on a deep convolutional layer of the convolutional neural network to obtain texture features;

[0079] Calculating the colloid displacement field between adjacent frames in the real-time processed image using an optical flow method to generate a dynamic viscosity change feature vector;

[0080] Detecting bubbles in the real-time processed image using a pre-trained segmentation model to obtain bubble features;

[0081] The color histogram, the texture feature, the dynamic viscosity change feature vector and the bubble feature are integrated to obtain image features.

[0082] Specifically, during the isoprene rubber polymerization reaction, characteristics of the reaction material, such as transparency, color change, presence of particles, and foam formation, are all related to the degree of reaction. For example, as the reaction progresses, viscosity increases, and different textures or flow characteristics appear on the surface of the material, which in turn changes the image captured by the camera. Based on this, the present invention uses machine learning and image segmentation techniques to extract image features from real-time image data, which includes the following steps:

[0083] The present invention first grayscales the real-time image data to reduce the amount of calculation and enhance contrast, thereby retaining key brightness information. It then uses a median filter to eliminate salt and pepper noise in the image to preserve the clarity of the bubble edges. Finally, a Gaussian filter is used to smooth the light scattering noise in the image to improve the accuracy of texture analysis. The denoised real-time image data is then binarized to separate the bubbles from the background and reduce interference from uneven illumination, resulting in a binary mask image, i.e., the real-time processed image.

[0084] Use the shallow convolution kernel of the pre-trained CNN (such as the first convolution layer of VGG) to extract the color distribution characteristics of the real-time processing image, and count the pixel value distribution of the RGB channels to generate a color histogram to reflect the overall hue change of the reaction liquid; and calculate the variance of each channel and average it through the middle convolution layer of the pre-trained CNN (such as Conv3 of VGG) to capture the scalar texture roughness value of the real-time processing image, and then use the deep convolution layer of the pre-trained CNN (such as res5c of VGG) to process the scalar texture roughness value of the image, and apply global average pooling to generate texture features to capture the microstructure of the formation or phase separation of intermediate products and precipitates.

[0085] Taking the real-time processed image as input, the Farneback algorithm is used to calculate the displacement field, then the displacement field amplitude is calculated, and the mean, variance and main direction histogram of the amplitude are statistically analyzed as the dynamic viscosity change feature vector, which can be used to quantify the macroscopic characteristics of fluid motion. In addition, viscosity sensors such as rotational viscometers can also be used to measure the material viscosity change data during the reaction process as the dynamic viscosity change feature vector.

[0086] Then, the real-time processing image is segmented using the pre-trained segmentation model, and the bubbles in each area are detected using the bubble kinematic model to obtain bubble features to indicate the intensity of the reaction; finally, the color histogram, texture features, dynamic viscosity change feature vector and bubble features are normalized and feature spliced ​​to obtain image features.

[0087] The present invention combines multi-dimensional features such as color, texture, dynamic viscosity, and bubbles to comprehensively describe the physical and chemical state of the reaction system, avoid the one-sidedness of a single feature, and improve the robustness of state judgment; uses grayscale and binarization processing to quickly simplify image complexity and reduce subsequent calculations; utilizes the hierarchical feature extraction capabilities of the pre-trained model to take into account both low-order color texture and high-order semantic information; and combines physical model-driven (optical flow method) and data-driven (image segmentation technology) methods to enhance the interpretability of dynamic features.

[0088] In one embodiment, the detecting bubbles in the real-time processed image using a pre-trained segmentation model to obtain bubble features includes:

[0089] Using a pre-trained MobileNetV3 model to process the real-time processing image to obtain a bubble segmentation mask;

[0090] Performing hole filling, noise filtering, and edge refinement processing on the bubble segmentation mask to obtain a bubble connected domain;

[0091] Calculating the area, circularity and centroid coordinates of each bubble connected domain to construct the static characteristics of the bubble;

[0092] Based on the centroid coordinates, a SORT algorithm is used to associate bubbles in adjacent frames to obtain continuous coordinates of each bubble to quantify the kinematic parameters of each bubble;

[0093] Stokes' law is used to verify each of the bubbles to mark abnormal bubbles;

[0094] The area, the circularity, the kinematic parameters and the abnormal bubbles are integrated to obtain bubble characteristics.

[0095] Specifically, the present invention uses MobileNetV3 as the encoder, which is pre-trained on a synthetic bubble dataset (Blender simulates different viscosities and lighting conditions), and uses it to process real-time processing images, outputting a pixel-by-pixel bubble mask (0 / 1 binary image); then the obtained bubble segmentation mask is subjected to bubble morphological post-processing, including: ‌hole filling‌: performing a morphological closing operation (5×5 circular kernel) on the segmentation mask to seal the internal holes of the bubbles; ‌noise filtering‌: removing tiny artifacts based on the connected domain area threshold (such as particles with a diameter of less than 3 pixels are judged as noise); ‌edge refinement‌: extracting sub-pixel bubble contours through the Sobel operator to improve positioning accuracy, thereby obtaining multiple bubble connected domains.

[0096] The area (pixel counting method), circularity (4π area / perimeter squared), and center of mass coordinates of each bubble's connected domain are calculated and used as static bubble features. Based on the center of mass coordinates of each bubble's connected domain, the SORT (Simple Online and Realtime Tracking) algorithm is used to associate bubbles in adjacent frames. The Kalman filter is used to predict the position of the next frame. The Hungarian algorithm is used to match the detection box with the predicted box (intersection over union threshold > 0.7), and the continuous trajectory coordinates (time series coordinates) of each bubble are recorded. Based on the calculated coordinates, kinematic parameters such as instantaneous velocity, acceleration, and trajectory curvature are calculated, and each bubble is verified using Stokes' law, with those that fail verification marked as abnormal bubbles. The bubble density (number of bubbles per unit area), average velocity variance (reflecting flow uniformity), and abnormal bubble proportion (indicating abnormal risk) in each frame are summarized and integrated with the area, circularity, kinematic parameters, and abnormal bubbles to obtain the bubble features.

[0097] The present invention quantifies the morphological characteristics of bubbles through static features (area, circularity) to distinguish bubble types (such as microbubbles and large bubbles); constructs dynamic features (movement speed, trajectory) and combines them with Stokes' law for verification to identify non-ideal moving bubbles caused by abnormal fluid viscosity or impurities; combines precise bubble characterization with dynamic tracking, and uses abnormal marking to improve monitoring reliability, avoid misjudgment (such as mistaking impurity particles for bubbles), improve the accuracy of subsequent predictions, and enhance the credibility of the physical model.

[0098] S3. Based on a cross-modal attention mechanism, the temporal features and the image features are fused to obtain a multimodal joint representation;

[0099] In one embodiment, step S3 includes:

[0100] The temporal self-attention layer of the pre-trained Transformer is used to encode the temporal features to obtain a temporal feature vector;

[0101] Processing the image features using a local block of a pre-trained Vision Transformer to generate an image feature vector;

[0102] Based on a cross-modal attention mechanism, a bidirectional interaction is performed between the time series feature vector and the image feature vector, and a learning parameter matrix is ​​introduced in the interaction process to adjust the interaction strength to generate a first enhanced representation and a second enhanced representation;

[0103] The first enhanced representation and the second enhanced representation are aggregated through a gating mechanism, and a conditional routing mechanism is introduced in the aggregation process to dynamically select a fusion path based on the modal correlation between the first enhanced representation and the second enhanced representation to obtain a multimodal joint representation.

[0104] Specifically, the present invention uses a variant of LSTM or the temporal self-attention layer of Transformer to encode temporal features to extract global temporal dependencies and generate a temporal feature vector; and uses the high-level features of the convolutional neural network or the local blocks of the Vision Transformer to extract the spatial structure information of the image features and generate an image feature vector.

[0105] A shared attention module is designed based on the cross-modal attention mechanism, allowing the time series feature vector and the image feature vector to serve as queries, keys, and values ​​for bidirectional interaction, namely, time series to image interaction: using the time series feature vector as the query and the image feature vector as the key and value to generate a time series enhanced representation of the image, and image to time series interaction: using the image feature vector as the query and the time series feature vector as the key and value to generate a time series enhanced image representation; then, a learning parameter matrix is ​​introduced in the interaction process to dynamically adjust the interaction strength between different modalities, thereby avoiding a single modality dominating the fusion process, and the time series enhanced representation of the image after the interaction is used as the first enhanced representation, and the image enhanced representation of the time series after the interaction is used as the second enhanced representation.

[0106] A gating mechanism (such as the gating unit or gated convolution of GRU) is used to calculate the modal correlation and its corresponding weights between the first enhanced representation and the second enhanced representation, and a conditional routing mechanism (such as soft routing based on feature similarity) is introduced to quantify the relationship between the difference in the weights of the two enhanced representations and the preset difference. When the calculated difference is less than the preset difference, the corresponding enhanced representations are weightedly fused using the calculated weights to obtain a multimodal joint representation. When the calculated difference is not less than the preset difference, the corresponding enhanced representations are gated and spliced ​​using the Sigmoid function and the calculated weights to obtain a multimodal joint representation.

[0107] The present invention uses a cross-modal attention mechanism to capture the potential correlation between temporal features and image features to perform deep fusion of complementary information; adopts a gating mechanism and conditional routing to dynamically adjust the fusion weights according to modal correlation, avoiding the modal dominance bias caused by fixed weights and improving the robustness of the model in complex scenarios; has high flexibility and can adapt to data distribution and task requirements of different modalities; has strong robustness and good fault tolerance for modal missing or noisy data.

[0108] S4. Processing the multimodal joint representation through a lightweight pre-trained model embedded with Arrhenius equation constraints to output a response prediction result;

[0109] In one embodiment, step S4 includes:

[0110] An Arrhenius equation is constructed based on the isoprene rubber polymerization reaction, and temperature data in the real-time reaction data is input as an explicit feature into the Arrhenius equation to obtain an Arrhenius auxiliary feature;

[0111] The multimodal joint representation, the temperature data, and the Arrhenius auxiliary features are spliced ​​into an input tensor, and the input tensor is input into the lightweight pre-training model for processing, and a reaction conversion rate prediction value and a viscosity change rate prediction value for a future preset time period are output as the reaction prediction result.

[0112] Specifically, the present invention constructs the Arrhenius equation based on the activation energy Ea and pre-exponential factor A of the isoprene rubber polymerization reaction: k(T)=A*e −Ea / (R*T) Where k(T) is the temperature-dependent reaction rate constant; T is the real-time temperature; and R is the gas constant. The activation energy can be determined by fitting the slope of the historical reaction data (temperature-conversion curve) to −Ea / R. The real-time temperature data from the real-time reaction data is then input as an explicit feature into the Arrhenius equation to calculate the linear combination of the logarithm of the rate constant and the inverse of the temperature. This yields the Arrhenius auxiliary feature FArr = [lnk(T), 1 / T], which is used to characterize the reaction kinetics.

[0113] The multimodal joint representation, real-time temperature data, and Arrhenius auxiliary features are concatenated into input tensors by time step or feature dimension for processing in a lightweight pre-trained model (TCN, GNN, etc.). The model outputs predicted reaction conversion rates and viscosity change rates for a preset time period (e.g., the next 5 minutes) as reaction prediction results. Alternatively, the lightweight pre-trained model can be a two-branch network; the physical constraint branch layer in this network is embedded in the Arrhenius equation computation layer to force the output to be strongly correlated with the reaction rate. The data-driven branch layer in this network uses 1D convolutional layers to extract temporal-spatial features, combined with a Transformer encoder to capture long-range dependencies. The resulting features are then mapped to conversion rates using fully connected layers. The Arrhenius equation is used to correct for biases. Dynamic viscosity is then output based on a temperature-viscosity-concentration multivariate model constructed using historical data. Residual connections are used to compensate for nonlinear effects, ultimately resulting in reaction prediction results.

[0114] The present invention explicitly embeds the Arrhenius equation to inject prior knowledge of reaction dynamics (the exponential effect of temperature on reaction rate) into the pre-trained model to avoid purely data-driven black-box predictions and improve the physical rationality and interpretability of the prediction results. The pre-trained model compresses the number of parameters through knowledge distillation while retaining the key information of multimodal features to meet the real-time reasoning needs of industrial scenarios. The integration of physical equation features (Arrhenius auxiliary features), multimodal joint representation (image + time series) and original temperature data can cover the coupling effect of macroscopic physical property changes and microscopic reaction dynamics, that is, temperature fluctuations quantify the reaction rate adjustment through the Arrhenius equation. The implicit impact of by-product accumulation is captured through multimodal features, thereby enhancing the prediction ability.

[0115] In another embodiment, the input layer of the lightweight pre-trained model is a high-dimensional input tensor of multimodal joint representation, temperature data, and Arrhenius auxiliary features. The encoder is constructed using depthwise separable convolution (MobileNet Block) and gated linear unit (GLU), and a dual-branch fully connected layer is used as the output layer to predict the conversion rate and viscosity change rate respectively. An explicit Arrhenius constraint layer (that is, a physically guided neural network layer) is added to the model for loss calculation.

[0116] The present invention optimizes the model parameters in stages to gradually enhance the influence of physical constraints, specifically:

[0117] In the pre-training stage (without physical constraints), historical data or ImageNet is used as the sample set to enable the model to initially learn the mapping relationship between multimodal features and response results. During the training process, only the task loss (MSE) function is used, and the Adam optimizer is used to optimize the model parameters to obtain the basic model parameters.

[0118] During the fine-tuning phase (introducing physical constraints), the MSE function is combined with a physical constraint loss function designed based on the Arrhenius equation and the differential equation of the conversion rate constructed to obtain a total loss function to optimize the basic model parameters during training. The physical constraint loss function is expressed as follows:

[0119]

[0120]

[0121] Where, L phy is the physical constraint loss function; N is the total number of iterations; i is the number of iterations; is the predicted value of reaction conversion rate; t For time; nis the kinetic order to which the isoprene rubber polymerization reaction belongs; C is the reaction conversion rate.

[0122] In the total loss function, the initial weight of the physical constraint loss function is set to 0.1, and then increases by 0.1 in each subsequent round. However, if the decrease rate of the physical constraint loss function is less than 5%, the weight corresponding to the physical constraint loss function is frozen until the next stage. If the MSE loss function increases, early stopping is triggered (rollback to the previous round parameters), and the RAdam optimizer is used to fine-tune the model parameters during training to obtain the adjusted parameters.

[0123] In the physical constraint strengthening stage (high weight constraint), the model is forced to strictly follow the Arrhenius equation, and the weight of the physical constraint loss function is fixed to 1.0, and high-order physical constraints are added to form a high-order loss function. L phy-high , which is expressed by the following formula:

[0124]

[0125] Where γ is a parameter with a value of 0.5, which is enabled only when the physical constraint loss function is less than 0.01.

[0126] During the training process, a second-order optimizer (L-BFGS) is used to improve convergence accuracy, and the input layer is frozen. Only the output head and Arrhenius constraint layer are fine-tuned to obtain updated parameters and verify the accuracy of the model after configuring the updated parameters. If its accuracy does not meet the requirements, the fine-tuning stage and the physical constraint strengthening stage are iteratively executed until the final model accuracy meets the preset requirements, and a trained lightweight pre-training model can be obtained.

[0127] The present invention adopts a multi-round training mechanism to train the lightweight pre-training model, wherein, in the pre-training stage, there are no constraints and multimodal feature learning is introduced; in the fine-tuning stage, the weight of the physical constraint loss function is gradually increased to balance physics and data; in the strengthening constraint stage, high-order physical terms are strictly dynamically consistent; by injecting physical constraints in stages, the lightweight pre-training model can learn complex patterns from the data during the training process and strictly follow the laws of reaction kinetics, which greatly improves the prediction accuracy and realizes high-precision and explainable reaction prediction in industrial scenarios.

[0128] S5. Based on the reaction prediction result, a reaction control strategy is generated and executed by a fuzzy control method and an anti-saturation integral PID control algorithm to achieve control of the isoprene rubber polymerization reaction;

[0129] In one embodiment, step S5 includes:

[0130] Obtaining a reaction error corresponding to the reaction prediction result according to a transfer learning method, and decomposing the reaction error into a short-term error and a long-term error using a sliding window averaging method;

[0131] Constructing a dynamic weight function through the long-term error and the short-term error, and using the dynamic weight function to quantify the first contribution ratio and the second contribution ratio corresponding to the fuzzy control method and the anti-saturation integral PID control algorithm respectively;

[0132] Based on the reaction error, generating a first control signal and a second control signal by a fuzzy control method and an anti-windup integral PID control algorithm, and weighting them by the first contribution ratio and the second contribution ratio to obtain a total control signal;

[0133] According to the total control signal and the reaction error, the pulse width, pulse interval and pulse amplitude are adjusted by a gradient optimization algorithm, and the reaction control strategy is generated and executed to achieve control of the isoprene rubber polymerization reaction.

[0134] Specifically, the present invention collects reaction conversion rates and viscosity change rates corresponding to different control inputs (such as temperature, catalyst flow rate, etc.) in historical reactions or historical simulations, and performs dimensionality reduction processing on these historical data to extract key features such as conversion rate peak and viscosity change slope, and then establishes a mapping relationship table between control input and response features based on the key features and the collected historical data to construct an input response database; then, Euclidean distance or cosine similarity is used to match the current reaction conversion rate prediction value and viscosity change rate prediction value with the historical response features, and the historical response with the highest similarity is selected, and its corresponding control input is extracted as the initial control strategy; then, according to the process requirements, the target conversion rate and target viscosity change rate are defined, and the error data of similar historical responses is used as the reaction error corresponding to the current reaction prediction result according to the transfer learning method, and the sliding window averaging method is used to decompose the reaction error into short-term error (reflecting the need for fast dynamic response) and long-term error (reflecting the need for steady-state accuracy).

[0135] The dynamic weight function is constructed by long-term error and short-term error, which is expressed as follows:

[0136]

[0137] Where, is the dynamic weight function; t For time; k 1. k 2 are adjustable parameters used to control the relative weights of short-term error and long-term error; e 1, e2 are short-term error and long-term error respectively.

[0138] Then, a fuzzy controller is designed, which takes the long-term error and the short-term error as input variables and outputs a control signal. The fuzzy rule base can be designed according to process experience, such as: if the short-term conversion rate error is large and the short-term viscosity error is small, then increase the reaction temperature; if the long-term viscosity error is large and the long-term conversion rate error is small, then adjust the catalyst concentration, etc.; and an anti-saturation integral PID controller is designed, which takes the reaction error as input and outputs a compensation signal. It adopts an anti-saturation integral strategy. When the output control signal approaches the actuator limit, it dynamically freezes the integral term accumulation to prevent integral saturation; then the control mode is dynamically selected according to the error scale characteristics: fuzzy dominant mode: when the short-term error is significant, the signal output by the fuzzy controller is dominant; PID dominant mode: when the long-term error is significant, the signal output by the anti-saturation integral PID controller is dominant; mixed mode: when both short-term and long-term errors cannot be ignored, the output signals are weighted summed according to their weights to obtain the first contribution ratio and second contribution ratio corresponding to the fuzzy control method and the anti-saturation integral PID control algorithm respectively.

[0139] The reaction error and its corresponding long-term and short-term errors are respectively input into the anti-saturation integral PID controller and the fuzzy controller for processing, and the second control signal corresponding to the anti-saturation integral PID controller and the first control signal corresponding to the fuzzy controller are generated respectively. The corresponding control signals are then weighted using the first contribution ratio and the second contribution ratio respectively, and the total control signal can be generated by combining the switching logic.

[0140] Finally, the total control signal is parameterized into three core variables: pulse width (PWM): controls the action time of the actuator; pulse interval (PI): controls the sleep time of the actuator; pulse amplitude (PA): controls the output intensity of the actuator; the objective function of the pulse control is defined by comprehensively considering the adjustment time (system response speed), overshoot (system stability) and energy consumption (actuator energy consumption), and the pulse width, pulse interval and pulse amplitude are optimized and adjusted through the gradient optimization algorithm. Then, combined with the state of the reaction error and the optimization results, the optimal pulse mode is dynamically selected as the reaction control strategy and sent to the actuator for execution to achieve control of the isoprene rubber polymerization reaction; among them, if the error of the optimal pulse mode is large and a fast response is required, the "high PWM+short PI" mode is selected; if the error is small and energy saving is required, the "low PWM+long PI" mode is selected, etc.

[0141] The present invention combines the nonlinear processing capability of fuzzy control with the steady-state accuracy of anti-saturation PID. Through a real-time dynamic weight adjustment strategy, it can not only respond quickly under severe disturbances, but also eliminate long-term steady-state errors and adapt to complex and changing reaction environments. It adopts a sliding window averaging method to separate short-term fluctuations from long-term trends, avoiding control oscillations or lags caused by single time scale errors. The gradient optimization algorithm dynamically adjusts pulse parameters to achieve a balance between energy consumption and response speed of actuators (such as heaters and stirring motors), reducing equipment losses while achieving effective control of isoprene rubber polymerization reactions.

[0142] In one embodiment, after step S4, the method further includes:

[0143] Based on the reaction prediction results, a pre-trained hierarchical architecture model is used to generate and execute a reaction control strategy to achieve control of the isoprene rubber polymerization reaction; wherein the hierarchical architecture model includes a bottom control layer, a middle control layer, and a high control layer;

[0144] In one embodiment, the reaction control strategy is generated and executed using a pre-trained hierarchical architecture model based on the reaction prediction result to achieve control of the isoprene rubber polymerization reaction, including:

[0145] A temperature controller is designed based on a PID algorithm to construct the bottom control layer, and a preset temperature upper limit is used as a first constraint. The deviation between the current temperature and the preset temperature is input into the bottom control layer for processing, and a control adjustment sequence is output;

[0146] The middle control layer is constructed based on a pre-trained LSTM neural network, the reaction prediction results are input into the middle control layer for processing, the predicted energy consumption is output, and a multi-objective optimization model is established with maximizing conversion rate, minimizing viscosity fluctuation, and minimizing energy consumption as the ultimate goals. The multi-objective optimization model is solved by a preset optimization algorithm with the dynamic balance of preset upper and lower limits of multiple reaction parameters and raw material feeding and discharging rates as the second constraint to obtain a prediction control sequence;

[0147] In the high-level control layer, a reward function is constructed based on the first constraint, the second constraint and the final goal, and a deep reinforcement learning algorithm is used to adjust the control adjustment sequence and the predictive control sequence, and a reaction control strategy is generated and executed to achieve control of the isoprene rubber polymerization reaction.

[0148] Specifically, the present invention can also use a pre-trained hierarchical architecture model to generate a reaction control strategy, including:

[0149] First, in the underlying control layer, the present invention designs a temperature controller based on the PID (proportional-integral-differential) algorithm. By adjusting the three parameters of proportion, integration and differentiation, it can achieve precise control of temperature, and set a preset temperature upper limit as the first constraint to ensure that the reaction temperature is not too high to avoid adverse effects on the reaction; then the deviation between the current temperature and the preset temperature in the real-time reaction data is input into the underlying control layer for processing through the PID algorithm, and then a control adjustment sequence is output. This sequence contains specific instructions for temperature adjustment, such as the intensity and time of heating or cooling, and the adjustment amount of cooling water flow.

[0150] Next, the present invention constructs a middle-level control layer based on the pre-trained LSTM neural network, which can process sequence data and is suitable for processing reaction prediction results. The reaction prediction results are input into the middle-level control layer for processing through the LSTM neural network, and the predicted energy consumption of the polymerization reaction in a preset future time period (such as the next 5 minutes) is output; then, based on the predicted energy consumption, a multi-objective optimization model is established with maximizing the conversion rate, minimizing the viscosity fluctuation, and minimizing the energy consumption as the ultimate goals, and preset upper and lower limits of temperature, preset upper and lower limits of pressure, and preset upper and lower limits of stirring speed, as well as the dynamic balance of raw material feeding and discharging rates are set as the second constraint to ensure the safety and stability of the reaction process. The multi-objective optimization model is solved using a preset optimization algorithm (such as a genetic algorithm, a particle swarm algorithm, etc.), and a predictive control sequence can be obtained, which includes specific control strategies for achieving multi-objective optimization under different constraints, such as the stirring speed adjustment amount and the raw material feed rate adjustment amount.

[0151] Finally, in the high-level control layer, the current conversion rate, viscosity, temperature, pressure, energy consumption, safety risk level, etc. are used as the state space, and the weight adjustment of the objective function (such as increasing the conversion rate priority or reducing the energy consumption priority) and fine-tuning of the PID parameters are used as the action space for environmental modeling. Hard constraints such as the first constraint and the second constraint are encoded as penalty terms of the reward function, and a reward function is constructed based on the conversion rate, energy consumption and safety risk in the final goal. Deep reinforcement learning algorithms (such as PPO and SAC) are combined with the safe reinforcement learning (Safe RL) mechanism to adjust the control adjustment sequence and predictive control sequence, generate a reaction control strategy and send it to the actuator for execution, so as to achieve effective control of the isoprene rubber polymerization reaction.

[0152] The present invention uses a layered architecture model in combination with a PID algorithm, an LSTM neural network, and a deep reinforcement learning algorithm to achieve precise control of the isoprene rubber polymerization reaction and improve product quality and production efficiency. When constructing the middle control layer, multiple objectives such as maximizing the conversion rate, minimizing viscosity fluctuations, and minimizing energy consumption are taken into consideration, and the results are solved through a multi-objective optimization model, making the control strategy more comprehensive and optimized. The application of the deep reinforcement learning algorithm in the high-level control layer enables the control strategy to be dynamically adjusted according to the real-time reaction situation, thereby improving the adaptability and robustness of the system. The design of the layered architecture model of the present invention makes each control layer relatively independent, facilitating maintenance and upgrading, and also improving the scalability of the system.

[0153] In addition, a hybrid algorithm of differential evolution and dynamic programming can be used to solve multi-objective optimization models. The differential evolution algorithm can generate new solutions through population difference vectors, quickly explore the global solution space, and avoid falling into local optimality. The dynamic programming algorithm can screen the optimal solution that meets the constraints through the state transition equation among the candidate solutions generated by the differential evolution algorithm, achieving local fine-tuning. Differential evolution provides diversity, while the dynamic programming algorithm improves the solution quality, thereby reducing invalid calculations. This hybrid algorithm includes:

[0154] Based on the predicted energy consumption, a multi-objective optimization model is constructed with the ultimate goals of maximizing conversion rate, minimizing viscosity fluctuation, and minimizing energy consumption. The preset upper and lower limits of temperature, pressure, and stirring speed are set, as well as the dynamic balance of raw material feeding and discharging rates as the second constraint.

[0155] In the differential optimization algorithm stage, an initial population is randomly generated, and each individual in the population satisfies the upper and lower limit constraints of the parameters; then, for each individual, three different individuals are randomly selected to generate a mutation vector; a crossover operation is performed on each individual and the mutation vector to generate a test vector; the objective function values ​​of the initial individual and the test vector are compared, and the better individual is selected to enter the next generation.

[0156] In the dynamic programming algorithm stage, the state is defined as the current reaction parameter combination, that is, the individual combination in the population, and the action is the parameter adjustment amount; based on the candidate solutions generated by the differential optimization algorithm, the objective function value of each state is calculated, and then the optimal transfer path is selected based on the objective function calculation result, and the solutions that do not meet the balance of input and output rates are eliminated through the pruning strategy.

[0157] Finally, from the Pareto frontier solution set generated by DE-DP, the optimal solution is selected according to the process requirements (such as prioritizing conversion rate while limiting viscosity fluctuation and energy consumption), and the optimal solution is mapped to a specific reaction parameter control sequence, such as temperature, pressure, catalyst concentration, etc., to obtain a predictive control sequence.

[0158] The DE-DP algorithm adopted in the present invention provides an efficient and reliable solution for the precise control of isoprene rubber polymerization reaction through the coordination of global and local optimization and the processing of dynamic constraints.

[0159] In addition, it should be noted that the various models used in the present invention are all trained models, which can obtain historical reaction data and historical image data in the isoprene rubber polymerization reaction process, as well as use online analyzers (such as near-infrared spectroscopy) to measure historical raw material conversion rates, and use laboratory sampling to mark the reaction degree of each key time point in the historical reaction, historical control strategies, etc. to form a variety of historical training sets after labeling these historical data. They can also be pre-trained in similar time series tasks (such as industrial sensor data), large data sets such as ImageNet, etc.

[0160] In the embodiment of the present application, based on the problem of how to effectively control the polymerization reaction of isoprene rubber and thus improve its reaction efficiency and safety, a method for controlling the polymerization reaction of isoprene rubber is designed. The method extracts the dynamic change law of key parameters such as temperature, pressure, conversion rate, etc. through a deep learning model, captures the reaction kinetics, uses a pre-trained model to analyze the spatial characteristics such as fluid morphology and bubble distribution in the reactor, reveals the local reaction heterogeneity, and dynamically associates the time series features with the image features based on the cross-modal attention mechanism to improve the comprehensiveness of the feature expression; the lightweight model embeds the Arrhenius equation as a priori knowledge to reduce the dependence on a large amount of data, improve the generalization ability in small sample scenarios, and Under abnormal working conditions such as agent deactivation, physical constraints can correct model prediction deviations and avoid the "black box" problem of pure data-driven methods; the fuzzy control method is used to dynamically adjust control parameters according to real-time input of multimodal joint representation, and the anti-saturation integral PID control method is used to supplement the performance degradation problem of traditional PID under integral saturation, ensuring smooth adjustment of the control quantity within a safe range; the present invention realizes the accurate perception-prediction-control closed loop of isoprene rubber polymerization reaction through the deep integration of data-driven (deep learning) and knowledge-driven (physical equations) combined with intelligent control (fuzzy + PID), which has significant advantages in improving product quality, process safety and economic benefits, and provides an innovative paradigm for the intelligent upgrading of complex chemical processes.

[0161] It should be noted that although the steps in the above flowchart are shown in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders.

[0162] In another embodiment, if Figure 2 As shown, the second aspect of the present invention provides an isoprene rubber polymerization reaction control system, comprising:

[0163] The data acquisition module 10 is used to collect real-time reaction data and real-time image data of the isoprene rubber polymerization reaction;

[0164] A feature extraction module 20 is configured to extract features from the real-time reaction data and the real-time image data based on a pre-trained deep learning model to obtain time series features and image features;

[0165] A feature fusion module 30 is configured to fuse the temporal features and the image features based on a cross-modal attention mechanism to obtain a multimodal joint representation;

[0166] a reaction prediction module 40, configured to process the multimodal joint representation using a lightweight pre-trained model embedded with Arrhenius equation constraints and output a reaction prediction result;

[0167] The strategy execution module 50 is used to generate and execute a reaction control strategy based on the reaction prediction result by using a fuzzy control method and an anti-saturation integral PID control algorithm to achieve control of the isoprene rubber polymerization reaction.

[0168] It should be noted that each module in the aforementioned isoprene rubber polymerization reaction control system can be implemented in whole or in part via software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor within a computer device in hardware form, or stored in a computer device memory in software form, allowing the processor to call and execute the corresponding operations of each module. For the specific definition of an isoprene rubber polymerization reaction control system, refer to the definition of an isoprene rubber polymerization reaction control method above. Both have the same functions and effects and will not be elaborated upon here.

[0169] In summary, the present invention relates to the field of data processing technology, and discloses a method and system for controlling an isoprene rubber polymerization reaction, which includes collecting real-time reaction data and real-time image data of an isoprene rubber polymerization reaction; performing feature extraction on the real-time reaction data and the real-time image data based on a pre-trained deep learning model to obtain time series features and image features; based on a cross-modal attention mechanism, fusing the time series features and the image features to obtain a multi-modal joint representation; processing the multi-modal joint representation through a lightweight pre-trained model embedded with Arrhenius equation constraints, and outputting a reaction prediction result; based on the reaction prediction result, generating and executing a reaction control strategy through a fuzzy control method and an anti-saturation integral PID control algorithm to achieve control of the isoprene rubber polymerization reaction; and achieving effective control of the isoprene rubber polymerization reaction through deep fusion of multi-source data and a variety of advanced algorithms.

[0170] Each embodiment in this specification is described in a progressive manner, and the same or similar parts of each embodiment can be directly referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. It should be noted that the various technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the various technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0171] The above-described embodiments merely represent several preferred implementations of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art could make several improvements and substitutions without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be based on the scope of protection of the claims.

Claims

1. A method for controlling an isoprene rubber polymerization reaction, characterized in that: include: Collecting real-time reaction data and real-time image data of isoprene rubber polymerization reaction; Based on the pre-trained deep learning model, feature extraction is performed on the real-time reaction data and the real-time image data to obtain time series features and image features; Based on the cross-modal attention mechanism, the temporal features and the image features are fused to obtain a multimodal joint representation; Processing the multimodal joint representation through a lightweight pre-trained model embedded with Arrhenius equation constraints to output a response prediction result; According to the reaction prediction result, a reaction control strategy is generated and executed through a fuzzy control method and an anti-saturation integral PID control algorithm to achieve control of the isoprene rubber polymerization reaction.

2. A method for controlling the polymerization reaction of isoprene rubber according to claim 1, characterized in that: The pre-trained deep learning model performs feature extraction on the real-time reaction data and the real-time image data to obtain time series features and image features, including: extracting the temperature change rate and pressure fluctuation amplitude from the real-time reaction data using a pre-trained long short-term memory network to generate time series features; The real-time image data is processed by a pre-trained convolutional neural network and image segmentation technology to obtain image features.

3. A method for controlling the polymerization reaction of isoprene rubber according to claim 2, characterized in that: The real-time image data is processed by the pre-trained convolutional neural network and image segmentation technology to obtain image features, including: Performing grayscale and binarization processing on the real-time image data to obtain a real-time processed image; Using a shallow convolution kernel of the convolutional neural network to extract color distribution features of the real-time processed image to obtain a color histogram; Extracting texture roughness of the real-time processed image through a middle convolutional layer of the convolutional neural network, and processing the texture roughness based on a deep convolutional layer of the convolutional neural network to obtain texture features; Calculating the colloid displacement field between adjacent frames in the real-time processed image using an optical flow method to generate a dynamic viscosity change feature vector; Detecting bubbles in the real-time processed image using a pre-trained segmentation model to obtain bubble features; The color histogram, the texture feature, the dynamic viscosity change feature vector and the bubble feature are integrated to obtain image features.

4. A method for controlling the polymerization reaction of isoprene rubber according to claim 3, characterized in that: The detecting of bubbles in the real-time processed image using a pre-trained segmentation model to obtain bubble features includes: Using a pre-trained MobileNetV3 model to process the real-time processing image to obtain a bubble segmentation mask; Performing hole filling, noise filtering, and edge refinement processing on the bubble segmentation mask to obtain a bubble connected domain; Calculating the area, circularity and centroid coordinates of each bubble connected domain to construct the static characteristics of the bubble; Based on the centroid coordinates, a SORT algorithm is used to associate bubbles in adjacent frames to obtain continuous coordinates of each bubble to quantify the kinematic parameters of each bubble; Stokes' law is used to verify each of the bubbles to mark abnormal bubbles; The area, the circularity, the kinematic parameters and the abnormal bubbles are integrated to obtain bubble characteristics.

5. A method for controlling the polymerization reaction of isoprene rubber according to claim 1, characterized in that: The cross-modal attention mechanism is based on fusing the temporal features and the image features to obtain a multimodal joint representation, including: The temporal self-attention layer of the pre-trained Transformer is used to encode the temporal features to obtain a temporal feature vector; Processing the image features using a local block of a pre-trained Vision Transformer to generate an image feature vector; Based on a cross-modal attention mechanism, a bidirectional interaction is performed between the time series feature vector and the image feature vector, and a learning parameter matrix is ​​introduced in the interaction process to adjust the interaction strength to generate a first enhanced representation and a second enhanced representation; The first enhanced representation and the second enhanced representation are aggregated through a gating mechanism, and a conditional routing mechanism is introduced in the aggregation process to dynamically select a fusion path based on the modal correlation between the first enhanced representation and the second enhanced representation to obtain a multimodal joint representation.

6. A method for controlling the polymerization reaction of isoprene rubber according to claim 1, characterized in that: The multimodal joint representation is processed by a lightweight pre-trained model embedded with Arrhenius equation constraints to output a reaction prediction result, including: An Arrhenius equation is constructed based on the isoprene rubber polymerization reaction, and temperature data in the real-time reaction data is input as an explicit feature into the Arrhenius equation to obtain an Arrhenius auxiliary feature; The multimodal joint representation, the temperature data, and the Arrhenius auxiliary features are spliced ​​into an input tensor, and the input tensor is input into the lightweight pre-training model for processing, and a reaction conversion rate prediction value and a viscosity change rate prediction value for a future preset time period are output as the reaction prediction result.

7. A method for controlling the polymerization reaction of isoprene rubber according to claim 1, characterized in that: The method generates and executes a reaction control strategy based on the reaction prediction result by using a fuzzy control method and an anti-saturation integral PID control algorithm to achieve control of the isoprene rubber polymerization reaction, including: Obtaining a reaction error corresponding to the reaction prediction result according to a transfer learning method, and decomposing the reaction error into a short-term error and a long-term error using a sliding window averaging method; Constructing a dynamic weight function through the long-term error and the short-term error, and using the dynamic weight function to quantify the first contribution ratio and the second contribution ratio corresponding to the fuzzy control method and the anti-saturation integral PID control algorithm respectively; Based on the reaction error, generating a first control signal and a second control signal by a fuzzy control method and an anti-windup integral PID control algorithm, and weighting them by the first contribution ratio and the second contribution ratio to obtain a total control signal; According to the total control signal and the reaction error, the pulse width, pulse interval and pulse amplitude are adjusted by a gradient optimization algorithm, and the reaction control strategy is generated and executed to achieve control of the isoprene rubber polymerization reaction.

8. A method for controlling the polymerization reaction of isoprene rubber according to claim 1, characterized in that: After processing the multimodal joint representation by a lightweight pre-trained model embedded with Arrhenius equation constraints and outputting a response prediction result, the method further includes: According to the reaction prediction results, a pre-trained hierarchical architecture model is used to generate and execute a reaction control strategy to achieve control of the isoprene rubber polymerization reaction; wherein the hierarchical architecture model includes a bottom control layer, a middle control layer and a high-level control layer.

9. A method for controlling the polymerization reaction of isoprene rubber according to claim 8, characterized in that: The method comprises generating and executing a reaction control strategy based on the reaction prediction result using a pre-trained hierarchical architecture model to control the isoprene rubber polymerization reaction, comprising: A temperature controller is designed based on a PID algorithm to construct the bottom control layer, and a preset temperature upper limit is used as a first constraint. The deviation between the current temperature and the preset temperature is input into the bottom control layer for processing, and a control adjustment sequence is output; The middle control layer is constructed based on a pre-trained LSTM neural network, the reaction prediction results are input into the middle control layer for processing, the predicted energy consumption is output, and a multi-objective optimization model is established with maximizing conversion rate, minimizing viscosity fluctuation, and minimizing energy consumption as the ultimate goals. The multi-objective optimization model is solved by a preset optimization algorithm with the dynamic balance of preset upper and lower limits of multiple reaction parameters and raw material feeding and discharging rates as the second constraint to obtain a prediction control sequence; In the high-level control layer, a reward function is constructed based on the first constraint, the second constraint and the final goal, and a deep reinforcement learning algorithm is used to adjust the control adjustment sequence and the predictive control sequence, and a reaction control strategy is generated and executed to achieve control of the isoprene rubber polymerization reaction.

10. An isoprene rubber polymerization reaction control system, characterized in that: include: A data acquisition module is used to collect real-time reaction data and real-time image data of the isoprene rubber polymerization reaction; A feature extraction module is used to extract features from the real-time reaction data and the real-time image data based on a pre-trained deep learning model to obtain time series features and image features; A feature fusion module, configured to fuse the temporal features and the image features based on a cross-modal attention mechanism to obtain a multimodal joint representation; a reaction prediction module, configured to process the multimodal joint representation using a lightweight pre-trained model embedded with Arrhenius equation constraints and output a reaction prediction result; The strategy execution module is used to generate and execute a reaction control strategy based on the reaction prediction result by using a fuzzy control method and an anti-saturation integral PID control algorithm to achieve control of the isoprene rubber polymerization reaction.

Images