Intelligent control method for imidazole crystallization reaction temperature based on operating condition correlation prediction

Abstract

This invention belongs to the field of industrial mechanism modeling technology, specifically involving an intelligent control method for imidazole crystallization reaction temperature based on operating condition correlation prediction. The method includes: collecting multidimensional time-series data within the batch cycle of imidazole crystallization and filtering for noise reduction; calculating the deviation between theoretical heat dissipation and the actual center temperature of the material based on thermodynamic equilibrium mechanism, obtaining the intrinsic latent heat release index of crystallization, and constructing a dynamic time-regularized distance matrix to achieve feature alignment optimization; using external electrical parameters to infer the material rheological state to obtain the apparent rheological resistance factor, which is then used as the covariance kernel function of a weighted adaptive reconstruction Gaussian process regression model; inputting the predicted trajectory into the model predictive controller to solve for the optimal command to drive valve action. This invention eliminates the temporal misalignment of operating condition variables, dynamically adapts to the nonlinear heat transfer resistance changes in the imidazole crystallization process, and achieves high-precision intelligent control of the reaction temperature.

Description

Technical Field

[0001] This invention relates to the field of industrial mechanism modeling technology. More specifically, this invention relates to a method for intelligent control of imidazole crystallization reaction temperature based on operating condition correlation prediction. Background Technology

[0002] Imidazole and its derivatives are key pharmaceutical intermediates. In actual industrial production, the crystallization and purification of imidazole is a typical complex thermodynamic and fluid dynamic coupling process. The core of the crystallization process lies in precisely controlling the cooling rate of the material inside the reactor by adjusting the opening of the cooling water regulating valve in the reactor jacket to form a specific temperature trajectory. The accuracy of controlling this material temperature trajectory directly determines the nucleation rate of imidazole crystals, the uniformity of particle size distribution, and the purity of the final product.

[0003] Currently, for material temperature control within crystallization reactors, the industry generally employs purely data-driven model predictive control (MMDC) technology. This technology predicts the material temperature change trend over subsequent control periods and outputs feedforward control commands. However, this control scheme faces significant technical bottlenecks under imidazole crystallization conditions. Specifically, the mapping of multidimensional operating variables during imidazole crystallization suffers from severe temporal asynchrony, and changes in external adjustment variables such as cooling water flow rate and stirring speed result in time delays in the response of the material's central temperature within the reactor. Existing dynamic time warping algorithms, when performing feature alignment processing on multidimensional historical time-series data, rely solely on numerical Euclidean distance to find the shortest matching path, severely deviating from the underlying thermodynamic physical constraint of latent heat release during crystallization. This causes the algorithm to forcibly align stable cooling nodes under non-exothermic conditions with abnormal cooling nodes during latent heat bursts when searching for the optimal data matching path, resulting in temporal misalignment that violates thermodynamic principles. Ultimately, this leads the predictive model to learn incorrect correlation logic.

[0004] However, the nonlinear abrupt changes in physical properties during imidazole crystallization are also a significant reason for the substantial decrease in the prediction accuracy of existing heat transfer prediction models. Specifically, when imidazole crystallization enters the secondary growth stage, the precipitation of a large number of crystals causes a sharp increase in the apparent viscosity of the solid-liquid mixture, leading to a nonlinear and sudden drop in the convective heat transfer coefficient of the boundary layer on the inner wall of the reactor. To cope with the deterioration of heat transfer, the control system must acquire the changes in the apparent viscosity of the material in the reactor in real time. The conventional method for acquiring apparent viscosity mainly relies on the direct installation of an immersion online viscometer. However, during the fluidization process, the crystallized slurry is prone to adhesion and scaling on the sensor probe, which not only leads to serious distortion of the directly measured viscosity data but also incurs high equipment maintenance costs. On the other hand, some technologies attempt to use pure mathematical state observers for soft measurement estimation of viscosity. However, due to the lack of in-depth analysis of the deep physical mapping relationship between motor load and rheological resistance, the estimation results often have a large delay and are difficult to effectively integrate with the underlying algorithm structure of subsequent temperature prediction models. Furthermore, existing implicit data prediction models typically use a globally constant covariance kernel function, which cannot dynamically sense and adapt to changes in heat transfer resistance caused by drastic changes in fluid viscosity. This situation inevitably leads to a significant decrease in the accuracy of temperature trajectory prediction in the later stages of crystallization, which in turn causes production anomalies such as severe lag in cooling commands from the control system, excessive overshooting of the reactor temperature, and localized overcooling and agglomeration of materials within the reactor. Summary of the Invention

[0005] To address the technical challenges of heat transfer prediction failure caused by asynchronous operating conditions and drastic viscosity changes, this invention proposes an intelligent control method for imidazole crystallization reaction temperature based on operating condition correlation prediction. This method can dynamically regulate time sequence differences and adaptively adjust the kernel function of the heat transfer model through theoretical deviation calibration and multi-physical quantity correlation fusion, thereby achieving high-precision intelligent closed-loop control of imidazole crystallization temperature.

[0006] This invention provides an intelligent control method for imidazole crystallization reaction temperature based on operating condition correlation prediction, comprising: collecting multidimensional time series data within a complete batch cycle of imidazole crystallization by a distributed control system; smoothing and denoising the multidimensional time series data using a filtering algorithm to obtain an initial feature sequence; calculating the deviation between the theoretical heat dissipation temperature in the reactor and the real-time center temperature change of the material in the reactor based on the thermodynamic equilibrium mechanism to obtain the intrinsic latent heat release index of crystallization; constructing a dynamic time-regularized distance matrix based on the intrinsic latent heat release index of crystallization; and performing feature alignment optimization on the initial feature sequence; and using external electrical parameters to back-calculate the reaction temperature within the reactor based on fluid dynamics principles. The rheological state of the material is used to obtain the apparent rheological resistance factor. This apparent rheological resistance factor is then used as a weighting adjustment variable to adaptively reconstruct the covariance kernel function of the Gaussian process regression model. The aligned and optimized initial feature sequence is then input into the reconstructed Gaussian process regression model to predict the real-time center temperature trajectory of the material in the reactor during the subsequent control period. This real-time center temperature trajectory of the material in the reactor during the subsequent control period is then used as a feedforward input variable and passed to the cost function of the model predictive controller. The optimal command is then obtained using a preset crystallization cooling curve as a reference to drive the external valves and achieve intelligent control of the imidazole crystallization reaction temperature.

[0007] This invention constructs an intrinsic latent heat release index of crystallization that includes the difference between the theoretical heat dissipation temperature drop and the actual change in the real-time center temperature of the material in the reactor. This index can accurately assess the actual thermodynamic state inside the crystallization system and strictly filter out the interference of conventional cold-state temperature fluctuations in the non-exothermic stage, providing reliable physical support for the accurate alignment of subsequent time-series characteristics.

[0008] Preferably, the intrinsic latent heat of crystallization release index is calculated as follows:

[0009]

[0010] in, The intrinsic latent heat of crystallization release index; This is the temperature normalization constant; This is a scalar value representing the inherent heat exchange capacity of the reactor; This represents the real-time flow rate of the cooling water. This represents the real-time center temperature of the material inside the reactor during the current sampling period. This represents the real-time center temperature of the material inside the reactor during the previous sampling period.

[0011] This invention introduces a physical state penalty coefficient for the difference in latent heat release index in the characteristic distance calculation of multidimensional operating condition variables. This can force the optimization algorithm to actively abandon erroneous matching paths with inconsistent thermodynamic states, effectively eliminate the pseudo-correlation phenomenon in the alignment process of multidimensional operating condition variables, and ensure the rigor of time axis alignment and the scientific nature of multidimensional data fusion.

[0012] Preferably, the method for calculating the dynamic time-warped distance matrix is ​​as follows:

[0013] in, This represents the state-aware distance between the current time node and historical time nodes in the dynamic time-warped distance matrix. and These are the multidimensional operating condition variables at the current time node and historical time nodes, normalized by standard deviation, at the 1st time node. Numerical values ​​in each feature dimension; The total number of feature dimensions for multidimensional operating condition variables; and These are the latent heat release indices corresponding to the two nodes; This is the physical state penalty coefficient.

[0014] This invention uses the product of the active power and rotation speed of the stirring motor to inversely deduce the rheological state of the material in the reactor. It can detect the decline in heat transfer efficiency caused by the increase in the viscosity of the solid-liquid mixture in real time and accurately. At the mathematical calculation level, it eliminates the risk of division-to-zero collapse and gives the prediction model a clear physical viscosity perception capability.

[0015] Preferably, the multidimensional time series data includes the real-time center temperature of the material inside the reactor, the real-time flow rate of cooling water on the jacket water inlet pipe, the real-time rotation speed of the agitator, and the real-time active power of the agitator motor in the power supply circuit.

[0016] Preferably, the apparent rheological resistance factor is calculated as follows:

[0017] in, It is the apparent rheological resistance factor; This refers to the real-time active power of the stirring motor; This refers to the real-time rotational speed of the agitator. This is the inherent basic mechanical friction loss constant of the equipment.

[0018] This invention measures the degree of latent heat release during crystallization by using theoretical deviations and filters out spurious feature alignment during multidimensional operating condition variable matching by combining thermodynamic constraints. This eliminates the problem of delayed cooling command output caused by errors in the correlation logic of the control model, and significantly improves the timing accuracy and physical logic rationality of the feedforward control.

[0019] Preferably, after obtaining the apparent rheological resistance factor, it is necessary to calculate the smoothing weighting coefficient, which is calculated as follows:

[0020] in, For smoothing weighting coefficients; This is the preset upper limit extreme value of the historical apparent rheological resistance factor; It is the apparent rheological resistance factor; This is the preset lower limit extreme value of the historical apparent rheological resistance factor; This is a preset protective constant.

[0021] Preferably, a dynamic combination kernel function is calculated based on the smoothing weight coefficients. The dynamic combination kernel function is equal to the product of the smoothing weight coefficients and the linear kernel function, plus the product of the difference between the constant 1 and the smoothing weight coefficients and the radial basis kernel function. The calculation method is as follows:

[0022] in, For dynamically combined kernel functions; and The input feature vector; For smoothing weighting coefficients; It is a linear kernel function, and its expression is: ,in For variance hyperparameters, This is the bias constant; Here is the radial basis function kernel, and its expression is: ;in For variance hyperparameters, This is a length scale hyperparameter.

[0023] This invention extracts smoothing weight coefficients and dynamically combines linear kernel functions and radial basis kernel functions, enabling the prediction model to adaptively switch network architectures according to the physical rheological state. Combined with log-marginal likelihood optimization based on historical data, the Gaussian process regression model can capture both the global linear cooling trend and accurately respond to nonlinear characteristics caused by dramatic local viscosity changes, effectively improving the model's generalization ability and prediction accuracy under complex crystallization conditions.

[0024] Preferably, the specific steps for smoothing and denoising the multidimensional time series data using a filtering algorithm are as follows: set the length of the sliding window to be multiple consecutive sampling periods, remove the maximum and minimum values ​​of the multidimensional time series data within the sliding window, calculate the average value of the remaining values, and output the smoothed engineering time series data block as the initial feature sequence.

[0025] Preferably, the real-time center temperature trajectory of the material in the reactor during the subsequent control period is transmitted to the cost function of the model predictive controller, including: constructing a target cost function that includes temperature tracking error and energy consumption of regulating valve control action, setting physical constraints, solving the optimal solution of the target cost function using a quadratic programming algorithm, and calculating the incremental sequence of cooling water pneumatic regulating valve opening during the subsequent control period. The target cost function is calculated as follows:

[0026] in, Let the objective cost function be... To predict the length of the time domain; To control the length of the time domain; The preset temperature error penalty weight; The preset control increment penalty weight; The subsequent real-time center temperature trajectory predicted by the Gaussian process regression model; The set temperature is used as a reference for the baseline trajectory; The increment of the opening of the pneumatic control valve for cooling water is the problem to be solved. The quadratic programming algorithm must satisfy the following physical constraints, including the absolute valve opening constraint: Valve opening change rate constraint: .

[0027] This invention introduces a multi-objective cost function that comprehensively evaluates temperature tracking error and control valve energy consumption into the model predictive controller, and innovatively adds physical limit constraints on the absolute valve opening and rate of change of action. This not only outputs more stable and energy-efficient feedforward control commands, but also ensures that the obtained optimal control strategy conforms to the actual mechanical execution capability of the pneumatic valve in the field, effectively preventing mechanical damage and command failure caused by excessive control actions.

[0028] Preferably, in the step of driving the external valve to achieve intelligent control of the imidazole crystallization reaction temperature, the first element of the opening increment sequence is extracted as the actual command issued in the current control cycle, and it is converted into a continuous electrical signal to drive the pneumatic diaphragm regulating valve on site to operate, and then cyclically executed in subsequent sampling cycles.

[0029] The beneficial effects of this invention are as follows: This invention deeply embeds thermodynamic equilibrium mechanisms and fluid dynamic physical properties into a data-driven algorithm. By eliminating the temporal misalignment problem caused by conventional distance algorithms during the exothermic crystallization period, it ensures that the multidimensional time series data samples obtained by the prediction model all possess true physical mapping logic, thereby improving the temporal accuracy of feedforward control.

[0030] Furthermore, this invention indirectly infers the viscosity variation of the material system within the reactor by utilizing the motor load characteristics, and adaptively reconstructs the covariance kernel function of the Gaussian process regression model accordingly, endowing the implicit data prediction model with the dynamic perception capability of deteriorating physical heat transfer efficiency. Even under conditions of drastic changes in solid-liquid state during the later stages of crystallization, it can still ensure the prediction accuracy of the real-time center temperature trajectory of the material within the reactor during subsequent control periods.

[0031] Meanwhile, this invention optimizes the control valve through closed-loop rolling and outputs precise feedforward control commands, thereby implementing thermodynamic intervention on the growth environment of imidazole crystals, avoiding temperature overshoot and local overcooling, and thus improving the batch quality of chemical products and the stability of automated production. Attached Figure Description

[0032] Figure 1 This is a flowchart of the intelligent control method for imidazole crystallization reaction temperature based on working condition correlation prediction in this invention; Figure 2 This is a scatter plot of the feature optimization path in the planar coordinate system of this invention. Detailed Implementation

[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0034] This invention discloses an intelligent control method for imidazole crystallization reaction temperature based on operating condition correlation prediction, referring to... Figure 1 This includes steps S1-S4: S1. Collect multidimensional time series data within the complete batch cycle of imidazole crystallization, and use a filtering algorithm to smooth and denoise the multidimensional time series data to obtain the initial feature sequence.

[0035] In an optional embodiment, throughout the complete batch production cycle of imidazole crystallization, a distributed control system deployed within the chemical plant is used to synchronously collect multidimensional time-series data from various hardware sensors at a fixed frequency. Specifically, firstly, the real-time center temperature of the material inside the reactor is obtained through a platinum resistance temperature transmitter inserted in the lower middle part of the reactor; this center temperature directly reflects the current thermodynamic macroscopic state of the material system. Secondly, the real-time flow rate of the cooling water is obtained through a mass flow meter installed on the jacket inlet water pipe; this real-time flow rate characterizes the heat dissipation intervention intensity of the external cooling water system. Thirdly, the real-time rotational speed of the agitator is obtained through the inverter communication interface connected to the agitator motor. Finally, the real-time active power of the agitator motor is obtained through a three-phase power transmitter in the power supply circuit. Thus, the raw signals obtained from these four dimensions together constitute the key physical dataset describing the imidazole crystallization process.

[0036] Furthermore, considering the high-frequency electromagnetic noise interference caused by the frequent start-up and shutdown of high-power equipment in chemical plants, the distributed control system automatically executes a sliding median filter algorithm to smooth and denoise the multidimensional time series data after receiving it. Specifically, a sliding window with a length of five consecutive sampling periods is set, and after removing the maximum and minimum values ​​of the multidimensional time series data within each sliding window, the average value of the remaining values ​​is calculated. Subsequently, the distributed control system outputs the smoothed engineering time series data block, which is used as the initial feature sequence for subsequent feature extraction and control algorithms.

[0037] Thus, by applying a sliding median filtering algorithm to the collected multidimensional time series data for smoothing and denoising, this denoising process can effectively eliminate abrupt false signals caused by high-frequency electromagnetic interference in chemical plants, thereby allowing the distributed control system to retain the initial characteristic sequence that can highly restore the real crystallization conditions, further ensuring the stability of subsequent data operations.

[0038] S2. Calculate the deviation between the theoretical heat dissipation temperature in the reactor and the real-time center temperature change of the material in the reactor based on the thermodynamic equilibrium mechanism to obtain the intrinsic latent heat release index of crystallization. Combine the intrinsic latent heat release index of crystallization to construct a dynamic time-regularized distance matrix and perform feature alignment optimization on the initial feature sequence.

[0039] In an optional embodiment, according to the law of conservation of energy in chemical thermodynamics, the actual rate of temperature change of the material inside the reactor is determined by the thermal equilibrium coupling effect between the physical heat dissipation of the cooling system and the latent heat of crystallization inside the material. Therefore, the intrinsic latent heat of crystallization release index is constructed, and its calculation method is as follows:

[0040] in, The intrinsic latent heat of crystallization release index; Let be the temperature normalization constant, and its dimensions are: ; This is a scalar value representing the inherent heat exchange capacity of the reactor, in units of... ; This is the real-time flow rate of cooling water, in units of... ; This is the real-time center temperature of the material inside the reactor during the current sampling period, in units of... ; This represents the real-time center temperature of the material inside the reactor during the previous sampling period, in units of... .

[0041] It should be noted that the latent heat release index formula is constructed by coupling the law of conservation of energy in chemical thermodynamics. The formula introduces... The non-negative cutoff function, adhering to the premise of only being effective during the exothermic crystallization period, effectively filters out the interference of cold-state temperature fluctuations during the normal cooling phase. The natural logarithmic function can smooth the transient abnormal peaks during the latent heat burst period. In terms of dimensions, the unit of the scalar product of cooling water flow rate and inherent heat transfer capacity is °C, and the unit of temperature difference is also °C; the sum of the above terms in °C is then compared with the unit in °C. - ¹Temperature normalization constant Multiplication achieves dimensional cancellation, thus making the calculated intrinsic latent heat release index of crystallization... As a dimensionless parameter, it ensures the dimensional rigor of the underlying thermodynamic mathematical calculations.

[0042] For example, at a certain moment, the inherent heat exchange capacity scalar of the reactor is set. 2 Real-time flow rate of cooling water 10 The theoretical temperature drop is 20°C. If the real-time center temperature of the material inside the reactor is... For 80 The real-time center temperature of the material in the reactor during the previous sampling period 85 Substitute it into the theoretical temperature drop deviation calculation formula: The calculated deviation value is 15. The calculation result, after being processed by a non-negative truncation function, outputs 15. Set the temperature normalization constant 1 Further substituting the intrinsic latent heat of crystallization release index .

[0043] Furthermore, a dynamic time-warped distance matrix incorporating physical constraints is constructed. The method for calculating the state-aware distance between the current node and historical nodes is as follows:

[0044] in, This represents the state-aware distance between the current time node and historical time nodes in the dynamic time-warped distance matrix. and These are the multidimensional operating condition variables at the current time node and historical time nodes, normalized by standard deviation, at the 1st time node. Numerical values ​​in each characteristic dimension, such as temperature, flow rate, rotational speed, and power; The total number of feature dimensions for multidimensional operating condition variables; and These are the latent heat release indices corresponding to the two nodes; This is the physical state penalty coefficient.

[0045] It should be noted that the state-aware distance formula is constructed by introducing a constraint term that includes the difference in latent heat release index in the calculation of the characteristic distance of multidimensional operating condition variables. The latter part of the formula utilizes a dimensionless state penalty coefficient. The absolute value of the difference between the latent heat release index and the latent heat release index actively amplifies the penalty cost under thermodynamic state mismatch conditions. At the dimensional level, the input multidimensional operating condition variables have all undergone standard deviation normalization beforehand, removing their original physical dimensions; simultaneously, the latent heat release index is also a dimensionless parameter. Therefore, the state-aware distance calculated using this formula... It is a dimensionless, purely numerical system that solves the problem of dimensional conflict when aligning and fusing physical quantities with different dimensions, such as temperature, flow rate, and rotation speed, in multidimensional feature integration.

[0046] For example, assuming the sum of squared differences in each dimension of the multidimensional operating condition variables after standard deviation normalization between the current time node and historical time nodes, i.e., the squared basic Euclidean distance, is 0.25; the latent heat release index corresponding to the current time node... The value is 2.77; the latent heat release index corresponding to the historical time node is 0; the physical state penalty coefficient is 2.77. If the value is 10, then the state-aware distance between the current time node and the historical time node in the dynamic time warping distance matrix is... Because this mechanism amplifies the penalty cost, the optimization algorithm can automatically abandon incorrectly aligned paths that do not match thermodynamic states.

[0047] Thus, by constructing a dynamic time-warped distance matrix that integrates physical constraints and optimizing the feature alignment of the initial feature sequence, the temporal misalignment problem caused by conventional distance algorithms during the exothermic crystallization period can be eliminated, ensuring that the multidimensional time series data samples obtained by the prediction model all have real physical mapping logic, thereby improving the temporal accuracy of feedforward control.

[0048] S3. Based on the principle of fluid dynamics, the rheological state of the material in the reactor is reverse-engineered using external electrical parameters to obtain the apparent rheological resistance factor. The apparent rheological resistance factor is used as a weight adjustment variable to adaptively reconstruct the covariance kernel function of the Gaussian process regression model. The aligned and optimized feature sequence is then input into the reconstructed Gaussian process regression model to predict the real-time center temperature trajectory of the material in the reactor during the subsequent control period.

[0049] In an optional embodiment, from a fluid dynamics perspective, the increased viscosity of the material inside the reactor causes a synchronous increase in the fluid shear stress resistance torque on the agitator, resulting in the agitator motor needing to output higher active power to maintain the rotational speed. Therefore, the apparent rheological drag factor is calculated as follows:

[0050] in, The apparent rheological drag factor is expressed in units of 1000 kJ / m². ; This represents the real-time active power of the mixing motor, in units of... ; This refers to the real-time rotational speed of the agitator, in units of... ; This is the inherent basic mechanical friction loss constant of the equipment, in units of... .

[0051] Furthermore, at a certain moment, the real-time active power of the stirring motor is set. 50 Real-time rotation speed of the agitator 100 Mechanical friction loss constant For 500 Therefore, the apparent rheological resistance factor This calculation process accurately assesses the viscosity of the material in the reactor as it transitions from a clear liquid to a thick slurry.

[0052] However, after obtaining the apparent rheological drag factor, it is necessary to calculate the smoothing weight coefficients and reconstruct the covariance kernel function. The method for calculating the smoothing weight coefficients is as follows:

[0053] in, For smoothing weighting coefficients; This is the preset upper limit extreme value of the historical apparent rheological resistance factor; This is the preset lower limit extreme value of the historical apparent rheological resistance factor; This is a preset protective constant.

[0054] It should be noted that the smoothing weighting coefficients are constructed by incorporating the objective laws governing changes in the physical state of the fluid within the reactor. As the material viscosity increases, the fluid shear stress resistance torque on the agitator rises, leading to the derivation of the apparent rheological resistance factor. And the smoothing weight coefficient Based on this, the mathematical logic of range normalization is adopted. The numerator of the formula is the difference between the upper limit extreme value and the current resistance factor, and the denominator is the historical range and the protective constant. The sum. In terms of dimensions, all parameters in the numerator and denominator inherit the same unit from the apparent rheological drag factor, i.e. After the quotient operation is performed between the two, the physical dimensions are eliminated, and a dimensionless weight value between 0 and 1 is stably output, providing a scientific numerical benchmark for the adaptive dynamic combination of linear and nonlinear kernel functions in the prediction model.

[0055] For example, setting an upper limit extreme value for the historical apparent rheological drag factor. The value is 0.01, representing the historical lower limit extreme value. The value is 0.001, a preset protective constant. The current apparent rheological drag factor is 0.0001. The value is 0.0047. Therefore, the smoothing weight coefficient... .

[0056] Based on this, the dynamic combination kernel function is equal to the product of the smoothing weight coefficient and the linear kernel function, plus the product of the difference between the constant 1 and the smoothing weight coefficient and the radial basis kernel function. Its calculation method is as follows:

[0057] in, For dynamically combined kernel functions; and The input feature vector; the linear kernel function expression is: It can capture global linear trends; the radial basis kernel function expression is: It can capture local highly nonlinear features, among which , For variance hyperparameters, The bias constant is This is a length scale hyperparameter.

[0058] Before performing trajectory prediction, a historical training dataset is first constructed. The system extracts multi-dimensional feature sequences from multiple batches of production records, aligned and optimized using a dynamic time warping algorithm, as input features for the model. The actual temperature trajectories at the corresponding sampling times were used as label sequences. Using this dataset, the variance parameter in the model was optimized by maximizing the log-marginal-likelihood function and employing the Adam optimization algorithm. , Bias constant and length scale parameters Automatic optimization is performed to complete the offline training of the prediction model.

[0059] As the fluid viscosity increases dramatically in the later stages of crystallization, the smoothing weight coefficient automatically approaches zero, allowing the nonlinear radial basis kernel function to take over. By inputting the optimized features of the current batch into the trained reconstruction model, a high-precision future temperature prediction sequence can be output.

[0060] Thus, by constructing an apparent rheological resistance factor and adaptively reconstructing the covariance kernel function accordingly, the implicit data prediction model can be given the ability to dynamically perceive the deterioration of physical heat transfer efficiency. Even under the condition of drastic changes in solid-liquid state in the later stage of crystallization, the prediction accuracy of the real-time center temperature trajectory of the material in the reactor during the subsequent control period can still be guaranteed.

[0061] S4. The real-time center temperature trajectory of the material in the reactor during the subsequent control period is passed as a feedforward input variable to the cost function of the model predictive controller. The optimal command is obtained by using the preset crystallization cooling curve as a reference, and the external valve is driven to achieve intelligent control of the imidazole crystallization reaction temperature.

[0062] In an optional embodiment, after acquiring the real-time center temperature trajectory of the material inside the reactor during the subsequent control period based on viscosity dynamic correction, the system integrates it into the model predictive control closed-loop system. The controller uses the optimal crystallization cooling curve predetermined in the enterprise process specifications as the core reference trajectory and employs a quadratic programming algorithm to mathematically search for the optimal solution to the objective cost function. This algorithm comprehensively considers the balance between the center temperature tracking error of the material inside the reactor and the energy consumption of the regulating valve control action; the formula for calculating the objective cost function is as follows:

[0063] in, Let the objective cost function be... To predict the length of the time domain; To control the length of the time domain; The preset temperature error penalty weight; The preset control increment penalty weight; The subsequent real-time center temperature trajectory predicted by the Gaussian process regression model; The set temperature is used as a reference for the baseline trajectory; Let be the opening increment of the cooling water pneumatic regulating valve to be solved.

[0064] It should be noted that the construction of the objective cost function incorporates the optimization theory of quadratic programming optimal control. This formula uses a preset temperature error penalty weight. Compared with the preset control increment penalty weight This approach transforms deviation terms with different physical meanings into a unified mathematical value, comprehensively considering the balance between center temperature tracking error and pneumatic control valve operating energy consumption. This construction method not only guides the algorithm to optimize and solve for feedforward control commands that closely track the reference trajectory, but also, through the adjustment of opening increments... The square term is used to apply a penalty, which effectively suppresses the violent oscillations and jumps in the command, ensuring that the final continuous electrical signal changes smoothly and conforms to the mechanical action law of the pneumatic diaphragm control valve on site.

[0065] To ensure that the solved control commands match the actual capabilities of the physical actuators on site, the quadratic programming algorithm must satisfy the following system physical constraints: Valve absolute opening constraint, for example, between 0% and 100%:

[0066] Valve opening change rate constraint, i.e., to prevent excessive valve actuation from damaging the mechanical structure:

[0067] Based on the aforementioned cost function and physical constraints, the incremental sequence of cooling water pneumatic regulating valve openings during subsequent control periods is calculated using quadratic programming solvers such as the effective set method or interior point method.

[0068] Furthermore, during the control execution process, the system extracts the first element of the opening increment sequence and directly uses it as the actual command issued in the current control cycle. The system's continuous electrical signal output module converts the digital command into a standardized current signal, directly driving the pneumatic diaphragm regulating valve on the field pipeline to precisely adjust the real-time flow of cooling water. In each subsequent sampling cycle, the control system strictly cyclically executes the complete process from data acquisition to command issuance, forming a rolling closed-loop control of the entire crystallization temperature trajectory.

[0069] In this way, by optimizing the closed-loop rolling mechanism and outputting precise feedforward control commands to drive the regulating valve, thermodynamic intervention on the growth environment of imidazole crystals can be implemented, avoiding temperature overshoot and local overcooling, thereby improving the batch quality of chemical products and the stability of automated production.

[0070] Reference Figure 2 The figure illustrates two optimization paths based on the dynamic time warping algorithm: one is an alignment path based on the fundamental Euclidean distance, and the other is an optimization path optimized by introducing an intrinsic latent heat release index. When the path enters the high latent heat release difference penalty region, the penalty is applied according to the physical state penalty coefficient. This amplifies the penalty for thermodynamic mismatch. Under this constraint, the optimization algorithm avoids the difference in latent heat release exponent. For excessively large regions, pseudo-correlation paths that violate thermodynamic principles are actively discarded, thus forming the optimal alignment path under physical constraints. This mechanism effectively eliminates the temporal misalignment caused by conventional distance algorithms during the latent heat of crystallization burst period, ensuring that the sample data on which the model prediction is based all have true physical mapping logic, and providing key physical support for high-precision prediction of temperature trajectories during subsequent control periods.

[0071] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for intelligent control of imidazole crystallization reaction temperature based on operating condition correlation prediction, characterized in that, include: The deviation between the theoretical heat dissipation temperature in the reactor and the real-time center temperature change of the material in the reactor is calculated based on the thermodynamic equilibrium mechanism to obtain the intrinsic latent heat release index of crystallization. A dynamic time-normalized distance matrix is ​​constructed in combination with the intrinsic latent heat release index of crystallization. Feature alignment optimization is performed on the initial feature sequence obtained after smoothing and denoising the collected multidimensional time series data. Based on the principle of fluid dynamics, the rheological state of the material in the reactor is reverse-engineered using external electrical parameters to obtain the apparent rheological resistance factor. The apparent rheological resistance factor is used as a weight adjustment variable to adaptively reconstruct the covariance kernel function of the Gaussian process regression model. The aligned and optimized initial feature sequence is input into the reconstructed Gaussian process regression model to predict the real-time center temperature trajectory of the material in the reactor during the subsequent control period. The real-time center temperature trajectory of the material in the reactor during the subsequent control period is used as a feedforward input variable and passed to the cost function of the model predictive controller. The optimal command is obtained by using the preset crystallization cooling curve as a reference and driving the external valve to achieve intelligent control of the imidazole crystallization reaction temperature.

2. The intelligent control method for imidazole crystallization reaction temperature based on operating condition correlation prediction according to claim 1, characterized in that, The intrinsic latent heat release index is calculated as follows: in, The intrinsic latent heat of crystallization release index; This is the temperature normalization constant; This is a scalar value representing the inherent heat exchange capacity of the reactor; This represents the real-time flow rate of the cooling water. This represents the real-time center temperature of the material inside the reactor during the current sampling period. This represents the real-time center temperature of the material inside the reactor during the previous sampling period.

3. The intelligent control method for imidazole crystallization reaction temperature based on operating condition correlation prediction according to claim 1, characterized in that, The method for constructing the dynamic time-warped distance matrix is ​​as follows: in, This represents the state-aware distance between the current time node and historical time nodes in the dynamic time-warped distance matrix. and These are the multidimensional operating condition variables at the current time node and historical time nodes, normalized by standard deviation, at the 1st time node. Numerical values ​​in each feature dimension; The total number of feature dimensions for multidimensional operating condition variables; and These are the latent heat release indices corresponding to the two nodes; This represents the physical state penalty coefficient.

4. The intelligent control method for imidazole crystallization reaction temperature based on operating condition correlation prediction according to claim 1, characterized in that, The multidimensional time series data includes the real-time center temperature of the material inside the reactor, the real-time flow rate of cooling water on the jacket inlet pipe, the real-time rotation speed of the agitator, and the real-time active power of the agitator motor in the power supply circuit.

5. The intelligent control method for imidazole crystallization reaction temperature based on operating condition correlation prediction according to claim 1, characterized in that, The method for calculating the apparent rheological resistance factor is as follows: in, It is the apparent rheological resistance factor; This refers to the real-time active power of the stirring motor; This refers to the real-time rotational speed of the agitator. This is the inherent basic mechanical friction loss constant of the equipment.

6. The intelligent control method for imidazole crystallization reaction temperature based on operating condition correlation prediction according to claim 1, characterized in that, After obtaining the apparent rheological resistance factor, it is necessary to calculate the smoothing weighting coefficient, which is calculated as follows: in, For smoothing weighting coefficients; This is the preset upper limit extreme value of the historical apparent rheological resistance factor; It is the apparent rheological resistance factor; This is the preset lower limit extreme value of the historical apparent rheological resistance factor; This is a preset protective constant.

7. The intelligent control method for imidazole crystallization reaction temperature based on operating condition correlation prediction according to claim 6, characterized in that, The dynamic combined kernel function is calculated based on the smoothing weight coefficients. The dynamic combined kernel function is equal to the product of the smoothing weight coefficients and the linear kernel function, plus the product of the difference between the constant 1 and the smoothing weight coefficients and the radial basis kernel function. The calculation method is as follows: in, For dynamically combined kernel functions; and The input feature vector; For smoothing weighting coefficients; It is a linear kernel function, and its expression is: ,in For variance hyperparameters, This is the bias constant; Here is the radial basis function kernel, and its expression is: ;in For variance hyperparameters, This is a length-scale hyperparameter.

8. The intelligent control method for imidazole crystallization reaction temperature based on operating condition correlation prediction according to claim 1, characterized in that, The specific steps for smoothing and denoising the multidimensional time series data using a filtering algorithm are as follows: set the length of the sliding window to be multiple consecutive sampling periods, remove the maximum and minimum values ​​of the multidimensional time series data within the sliding window, calculate the average value of the remaining values, and output the smoothed engineering time series data block as the initial feature sequence.

9. The intelligent control method for imidazole crystallization reaction temperature based on operating condition correlation prediction according to claim 1, characterized in that, The real-time center temperature trajectory of the material in the reactor during the subsequent control period is transmitted to the cost function of the model predictive controller. This includes: constructing a target cost function that includes temperature tracking error and energy consumption of the regulating valve control action, setting physical constraints, solving for the optimal solution of the target cost function using a quadratic programming algorithm, and calculating the incremental sequence of the cooling water pneumatic regulating valve opening during the subsequent control period. The target cost function is calculated as follows: in, Let the objective cost function be... To predict the length of the time domain; To control the length of the time domain; The preset temperature error penalty weight; The preset control increment penalty weight; The subsequent real-time center temperature trajectory predicted by the Gaussian process regression model; The set temperature is used as a reference for the baseline trajectory; The increment of the opening of the pneumatic control valve for cooling water is the problem to be solved. The quadratic programming algorithm must satisfy the following physical constraints, including the absolute valve opening constraint: Valve opening change rate constraint: .

10. The intelligent control method for imidazole crystallization reaction temperature based on operating condition correlation prediction according to claim 9, characterized in that, In the step of driving the external valve to achieve intelligent control of the imidazole crystallization reaction temperature, the first element of the opening increment sequence is extracted as the actual command issued in the current control cycle, which is converted into a continuous electrical signal to drive the pneumatic diaphragm regulating valve on site, and then executed cyclically in subsequent sampling cycles.

Images