An automatic pressure regulation system and method for an acetate reactor

Abstract

This application relates to the field of acetate reaction control technology, and proposes an automatic pressure regulation system and method for acetate reactors. The method includes: collecting and preprocessing multi-source data to construct a gas-liquid-solid three-phase rheological characteristic sequence; extracting fingerprints related to crystal growth and bubble rupture using a deep belief network, and obtaining the dynamic correlation index between viscosity change rate and mass transfer coefficient by combining manifold learning dimensionality reduction; establishing a phase change coupled pressure evolution model, estimating gas phase escape flux deviation using an extended Kalman filter, and using an LSTM model to predict pressure abrupt changes to generate a predicted trajectory containing phase change disturbances; differentiating the prediction from the target, using adaptive fuzzy sliding mode control to obtain the critical point compensation amount, dynamically adjusting the exhaust valve gain and generating decoupling feedforward commands; adjusting the condenser and valves according to the commands, and triggering pulse backflushing to clean the pressure tap after exceeding the threshold to execute correction control.

Description

Technical Field

[0001] This application relates to the field of acetate reaction control technology, and proposes an automatic pressure regulation system and method for acetate reactors. Background Technology

[0002] During the acetate reaction, the environment inside the reactor is a dynamically evolving three-phase system consisting of gas, liquid, and solid phases. During the reaction, especially in the crystallization stage, the physical state of the materials undergoes drastic and complex continuous changes. For example, the formation and growth of crystals cause nonlinear changes in the viscosity of the liquid phase, accompanied by the formation and escape of the gas phase. These dynamic evolutions of phase transformations and rheological properties within the reaction system directly translate into strong, nonlinear endogenous disturbances to the pressure inside the reactor.

[0003] Existing pressure control systems typically rely on direct feedback from pressure measurements for adjustment. These systems exhibit significant limitations when faced with pressure fluctuations caused by drastic changes in the internal state of the material. Because they cannot perceive and quantify the underlying rheological evolution processes driving pressure changes within the reactor in real time, existing control methods are essentially a delayed and passive response. Consequently, the control system struggles to anticipate and effectively suppress rapid pressure drifts or jumps caused by internal factors such as crystal growth or abrupt changes in bubble behavior, often resulting in significant overshoot, frequent oscillations, or sluggish response in pressure control, failing to meet the stringent pressure stability requirements of high-quality acetate production. Therefore, achieving proactive and highly precise stable control of reactor pressure under such endogenous strong disturbances is a pressing technical challenge that needs to be addressed.

[0004] In view of this, this application proposes an automatic pressure regulation system and method for an acetate reactor. Summary of the Invention

[0005] To achieve the above objectives, this application provides an automatic pressure regulation system and method for an acetate reactor, the specific technical solution of which is as follows:

[0006] An automatic pressure regulation method for an acetate reactor includes:

[0007] Step 1: Collect multi-source data from the acetate reactor, preprocess the collected multi-source data, and construct an original feature vector sequence describing the gas-liquid-solid three-phase rheological state;

[0008] Step 2: Input the original feature vector sequence into the deep belief network model to extract the rheological feature fingerprints that characterize the growth state of acetate crystals and the frequency of bubble rupture. Map the high-dimensional rheological features to the low-dimensional state space through the manifold learning algorithm to quantify the dynamic correlation index between the viscosity change rate of the reaction liquid and the mass transfer coefficient of the gas-liquid interface.

[0009] Step 3: Construct a phase transition coupling evolution model of reactor pressure based on dynamic correlation index, use extended Kalman filter algorithm to estimate gas phase escape flux deviation in real time during crystallization, combine long short-term memory network to predict pressure change trend in future time window, and generate pressure prediction trajectory containing phase transition interference factor.

[0010] Step 4: Compare the pressure prediction trajectory with the set target value using differential comparison, calculate the pressure compensation control quantity for the phase transition critical point using an adaptive fuzzy sliding mode controller, dynamically adjust the opening gain coefficient of the exhaust valve according to the gas phase escape flux deviation, and generate a feedforward control command sequence to decouple the influence of feed flow rate fluctuations.

[0011] Step 5: Based on the feedforward control command sequence, adjust the flow rate of the condenser cooling medium and the opening of the exhaust regulating valve. When the pressure prediction value deviates from the safety threshold, trigger the pulse backflushing mechanism to clean the pressure tap and perform pressure adjustment according to the corrected control parameters.

[0012] Preferably, multi-source data from the acetate reactor are collected, including acoustic emission data, stirring motor torque ripple data, and thermodynamic parameters; the collected multi-source data are preprocessed, including time synchronization and wavelet packet decomposition noise reduction.

[0013] Based on the preprocessed multi-source data, an original feature vector sequence describing the three-phase rheological state of gas, liquid and solid is constructed. At the aligned time point, the root mean square value, signal energy and peak amplitude of the acoustic emission signal are extracted from the noise-reduced signal. The statistical features of standard deviation, kurtosis and skewness are extracted from the noise-reduced torque ripple signal. At the same time, the instantaneous measurement values ​​of the basic thermodynamic parameters are used as feature quantities.

[0014] All features extracted at the aligned time points will be combined to form a high-dimensional original feature vector.

[0015] Preferably, after constructing the original feature vector sequence describing the three-phase rheological state of gas, liquid and solid, the original feature vector sequence is input into a pre-trained deep belief network model, which is composed of multiple stacked restricted Boltzmann mechanisms.

[0016] The deep belief network model utilizes a large-scale sequence of original feature vectors collected from acetate reactors under various historical operating conditions to pre-train the deep belief network. The historical operating conditions include the normal crystallization period, the rapid crystal growth period, the period of intense solvent vaporization, and the period of significant changes in material viscosity. The training adopts a greedy layer-by-layer unsupervised learning method, that is, using the contrastive divergence algorithm to independently train each layer of the restricted Boltzmann machine.

[0017] The original feature vector sequence is used as the input of the lowest layer restricted Boltzmann machine of the deep belief network. Through the forward propagation of the network, the feature vector is output in the top hidden layer. The feature vector is a rheological fingerprint that characterizes the growth state of acetate crystal and the frequency of bubble rupture at the current moment.

[0018] The isomap algorithm is used to reduce the dimensionality and visualize the rheological feature fingerprint sequence extracted by the deep belief network.

[0019] Preferably, based on the reduced low-dimensional state space, the dynamic correlation index between the viscosity change rate of the reaction liquid and the mass transfer coefficient at the gas-liquid interface is quantified.

[0020] Two proxy variables are defined for the dynamic correlation index: the viscosity change rate proxy index and the gas-liquid interface mass transfer coefficient proxy index.

[0021] The viscosity change rate proxy index is obtained by calculating the time derivative of the projection component of the low-dimensional state point in the pre-calibrated viscosity change direction; the gas-liquid interface mass transfer coefficient proxy index is estimated by the projection component of the point in the low-dimensional state space in the mass transfer change direction.

[0022] The interaction strength and relative relationship between the viscosity change rate proxy index and the gas-liquid interface mass transfer coefficient proxy index were constructed by constructing a dynamic correlation index.

[0023] Preferably, a phase change coupled evolution model is constructed to describe the dynamic changes in pressure inside the reactor, and the pressure change rate is expressed as a function of the net generation rate of gaseous substances.

[0024] The extended Kalman filter algorithm is used to estimate the gas phase escape flux deviation during the crystallization process. The gas phase escape flux deviation is defined as the difference between the actual exhaust flux and the theoretical exhaust flux calculated based on valve opening and pressure difference.

[0025] By combining the phase transition coupled evolution model, nonlinear discrete-time state transition equations and observation equations are constructed; the estimated value of gas phase escape flux deviation is obtained by using the extended Kalman filter algorithm.

[0026] Preferably, the pre-trained long short-term memory network is used to predict the nonlinear pressure change trend within the future time window. The multivariate information from the current and past multiple sampling times is constructed into a feature sequence and used as the input of the long short-term memory network. The feature sequence at a single time includes: measured pressure value, gas phase escape flux deviation, dynamic correlation index, opening degree of exhaust regulating valve and flow rate of condenser cooling medium.

[0027] The Long Short-Term Memory Network outputs a sequence of predicted stress values ​​for future moments.

[0028] Preferably, the pressure prediction trajectory is compared with the preset pressure target value to generate a prediction error sequence;

[0029] An adaptive fuzzy sliding mode controller is used to calculate the pressure compensation control quantity for strong nonlinear disturbances at the phase transition critical point; the opening gain coefficient of the exhaust valve is dynamically adjusted according to the gas phase escape flux deviation, and a reference gain coefficient is defined.

[0030] When the estimated gas phase escape flux deviation is positive, the opening gain coefficient is automatically reduced; when the deviation is negative, the opening gain coefficient is increased.

[0031] Preferably, a feedforward control command sequence is generated to decouple the effects of feed flow rate fluctuations. The required feedforward control quantity is calculated by monitoring the instantaneous flow rate of the feed pump online and based on the feedforward model describing the dynamic relationship between the feed flow rate and the reactor pressure response.

[0032] The pressure compensation control quantity output by the adaptive fuzzy sliding mode controller and the result after dynamic gain adjustment are superimposed with the feedforward control quantity to form a feedforward-feedback composite control command.

[0033] Preferably, the feedforward-feedback composite control command sequence is parsed and assigned to the actuators, including the condenser cooling medium flow regulator and the reactor top exhaust regulating valve;

[0034] A collaborative decoupling control strategy based on the current thermodynamic state of the reactor is adopted to analyze the rate of temperature change and the rate of pressure change within the reactor. By introducing a dynamic decoupling weighting factor, the feedforward-feedback composite control command sequence is decomposed into a control component for the cooling medium flow rate and a control component for the exhaust valve opening.

[0035] Safety monitoring is performed using the predicted pressure trajectory. Pressure safety thresholds are set, including high-pressure safety thresholds and low-pressure safety thresholds. In each control cycle, all predicted pressure values ​​within the entire prediction time window are checked. When the pressure at any predicted time point is found to be outside the range of the high-pressure safety threshold and the low-pressure safety threshold, the pulse backflush mechanism is triggered.

[0036] Based on the decoupled control components, the flow rate of the cooling medium supply and the instantaneous opening of the exhaust regulating valve are adjusted.

[0037] An automatic pressure regulation system for an acetate reactor, used in the aforementioned automatic pressure regulation method for an acetate reactor, includes: a multi-source information characterization module, a reaction state identification module, a pressure trend prediction module, a control strategy decision-making module, and a collaborative regulation execution module.

[0038] The multi-source information characterization module collects multi-source data from the acetate reactor, preprocesses the collected multi-source data, and constructs an original feature vector sequence describing the gas-liquid-solid three-phase rheological state.

[0039] The reaction state identification module inputs the original feature vector sequence into the deep belief network model, extracts the rheological feature fingerprints that characterize the growth state of acetate crystals and the frequency of bubble rupture, and maps the high-dimensional rheological features to the low-dimensional state space through the manifold learning algorithm, quantifying the dynamic correlation index between the viscosity change rate of the reaction liquid and the mass transfer coefficient of the gas-liquid interface.

[0040] The pressure trend prediction module constructs a phase transition coupling evolution model of reactor pressure based on dynamic correlation index, uses extended Kalman filter algorithm to estimate gas phase escape flux deviation in real time during crystallization, and combines long short-term memory network to predict pressure change trend within future time window, generating a pressure prediction trajectory containing phase transition interference factor.

[0041] The control strategy decision module compares the pressure prediction trajectory with the set target value, calculates the pressure compensation control quantity for the phase transition critical point through an adaptive fuzzy sliding mode controller, dynamically adjusts the opening gain coefficient of the exhaust valve according to the gas phase escape flux deviation, and generates a feedforward control command sequence to decouple the influence of feed flow rate fluctuations.

[0042] The coordinated adjustment execution module, based on the feedforward control command sequence, adjusts the flow rate of the condenser cooling medium and the opening of the exhaust regulating valve. When the predicted pressure value deviates from the safety threshold, it triggers a pulse backflushing mechanism to clean the pressure tap and performs pressure adjustment according to the corrected control parameters. The beneficial effects of this application are: Through multi-source data acquisition and preprocessing, this application forms a stable and complete three-phase rheological characteristic sequence, reducing the interference of noise and missing data on judgment, improving the observability and consistency of key state quantities, and providing a reliable data foundation for subsequent feature extraction and modeling.

[0043] This application utilizes deep belief networks to extract rheological fingerprints of crystal growth and bubble rupture, and then uses manifold learning to reduce the dimensionality, thereby obtaining a dynamic correlation index between viscosity change rate and mass transfer coefficient. This enables quantifiable characterization of complex mechanisms and improves the sensitivity and robustness of phase transition sign identification.

[0044] This application constructs a phase change coupled pressure evolution model, combines extended Kalman filtering to estimate gas phase escape flux deviation online, and uses LSTM to predict pressure change trends, outputting pressure trajectories containing perturbation factors in advance, thereby enhancing the predictability and safety margin of pressure fluctuations during the crystallization stage.

[0045] This application uses adaptive fuzzy sliding mode to achieve rapid compensation of critical points by differentiating the predicted trajectory from the target, and automatically adjusts the exhaust valve gain based on the escape flux deviation to generate decoupled feedforward commands, thereby reducing the impact of feed fluctuations and phase change disturbances on the accuracy and stability of pressure control.

[0046] This application coordinates the condenser medium flow rate and the exhaust valve opening according to feedforward commands to achieve smooth pressure tracking; when the predicted value approaches the safety threshold, it triggers pulse backflushing to clean the pressure tap, reducing the risk of pressure drift and blockage, and ensuring that the control parameters are effective and the execution is reliable.

[0047] This application realizes an integrated closed-loop control of pressure prediction, compensation and execution during the acetate crystallization reaction. It can intervene in advance before the phase transition critical point, significantly reducing the probability of pressure surge, venting over-limit and safety interlock triggering. It is beneficial to improve the uniformity of product crystal size and yield, reduce energy consumption and unplanned venting, and improve intrinsic safety and economy. Attached Figure Description

[0048] Figure 1 A flowchart of an automatic pressure regulation method for an acetate reactor provided in this application;

[0049] Figure 2 This is a schematic diagram of the sensor deployment in the acetate reactor provided in this application;

[0050] Figure 3 A schematic diagram of the collaborative decoupling execution and pulse backflush mechanism provided in this application;

[0051] Figure 4 This application provides a structural diagram of an automatic pressure regulation system for an acetate reactor. Detailed Implementation

[0052] To make the above-mentioned objectives, features and advantages of this application more readily understood, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0053] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0054] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of this application. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.

[0055] Example 1

[0056] Reference Figures 1 to 3 This is the first embodiment of the present application, such as Figure 1 As shown, an automatic pressure regulation method for an acetate reactor is provided.

[0057] Step 1: Collect multi-source data from the acetate reactor, preprocess the collected data, and construct a sequence of original feature vectors describing the gas-liquid-solid three-phase rheological state. (See also...) Figure 2 This is a schematic diagram of the sensor deployment in the acetate reactor for this step.

[0058] To comprehensively capture the complex dynamic information of the gas-liquid-solid three-phase flow field during acetate crystallization, multi-source heterogeneous data from the reaction process are collected. At the data acquisition level, at least two broadband piezoelectric acoustic emission sensors are installed on the outer wall of the acetate reactor, away from structural stress concentration areas such as the inlet, outlet, and welds, using magnetic chucks or welded bases. An acoustic coupling agent is coated between the sensors and the reactor wall to ensure signal transmission fidelity. The frequency response range of the acoustic emission sensors is set to 100kHz to 1MHz, effectively covering the microscopic elastic energy release events generated during the growth, aggregation, and breakup of acetate crystals, as well as the acoustic events caused by the generation, merging, and collapse of bubbles within the reactor. The data acquisition card rapidly acquires the output voltage signals of the acoustic emission sensors, forming a high-frequency acoustic emission waveform raw data stream. Simultaneously, the instantaneous torque signal of the stirring motor is acquired at a sampling frequency of 1kHz by monitoring the DC bus current of the inverter driving the agitator or by directly installing a high-precision dynamic torque sensor on the stirring shaft. The focus is on the high-frequency fluctuations superimposed on the torque signal, i.e., the instantaneous torque ripple data, because the amplitude and frequency distribution of the torque ripple directly reflect the viscosity change caused by the increase in solid content in the reaction liquid and the interaction strength between the particulate phase and the agitator. In addition, thermal sensors are deployed at pre-set measurement ports inside the reactor to obtain basic thermodynamic parameters, specifically including: measuring the pressure of the gas phase space at the top of the reactor using a diaphragm pressure transmitter; measuring the temperature of the liquid phase using a resistance temperature sensor; monitoring the pH of the reaction liquid in real time using an in-situ insertion pH meter; and reading the actual rotational speed of the agitator from the frequency converter feedback signal.

[0059] Data preprocessing is performed on the acquired multi-source data. Since the acoustic emission, torque ripple, and thermodynamic parameters originate from different sensors, their sampling frequencies differ from the data recording clocks, necessitating timestamp alignment. A data acquisition server based on Network Time Protocol (NTP) synchronization can be used to provide a unified, high-precision time reference for all acquisition channels. For each acquired data stream, using the time axis with the highest sampling rate (i.e., the acoustic emission data sampling rate) as the reference, linear interpolation or forward padding methods are used to resample the torque ripple data and basic thermodynamic parameter data with lower sampling rates, generating a time-strictly aligned multi-channel data matrix. Timestamp alignment ensures that at any given time point, all data accurately reflect the physicochemical state of the reactor at the same moment.

[0060] Wavelet packet decomposition was performed to denoise the time-aligned high-frequency acoustic emission waveform data and instantaneous torque ripple data. Considering that mechanical vibration and electromagnetic interference in the reactor operating environment can contaminate the effective signal, wavelet packet decomposition technology allows for more refined time-frequency localization analysis of the signal. The original signal... conduct Layer wavelet packet decomposition can yield... The decomposition process of a sub-band signal covering the entire frequency band can be expressed as: ;in, This represents the raw signal to be processed (acoustic emission or torque ripple signal). The number of wavelet packet decomposition levels; for example, The value can be set to 4 or 5 depending on the signal characteristics and computing resources. For the decomposition after the first Layer frequency band index; This represents the original signal at the th... Components in each frequency band. For each frequency band component... The wavelet packet coefficients in the signal are subjected to soft-threshold quantization using an adaptive thresholding rule based on SURE (Stan Unbiased Risk Estimation), with coefficients below the threshold set to zero to filter out noise interference. The thresholded wavelet packet coefficients are then used to reconstruct the signal, yielding the denoised acoustic emission signal and torque ripple signal. This denoising step significantly improves the signal-to-noise ratio, allowing the weak characteristic signals generated by crystal growth and bubble activity to be highlighted.

[0061] Furthermore, based on the preprocessed multi-source data, a sequence of original feature vectors describing the three-phase rheological state of gas, liquid, and solid is constructed. At each aligned time point... The root mean square (RMS), signal energy, and peak amplitude time-domain features were extracted from the denoised acoustic emission signal. Statistical features such as standard deviation, kurtosis, and skewness were extracted from the denoised torque ripple signal to quantify the severity and distribution of torque fluctuations. Simultaneously, instantaneous measurements of fundamental thermodynamic parameters (pressure, temperature, pH, and stirring speed) were directly used as feature quantities. These parameters at specific time points were then analyzed. All extracted features are combined to form a high-dimensional original feature vector. :

[0062] ;

[0063] in, and These represent the root mean square value and energy of the acoustic emission signal, respectively. and These represent the standard deviation and kurtosis of the torque ripple signal, respectively. The pressure in the reactor; Temperature of the reaction solution; pH value of the reaction solution; The stirring speed. The continuously generated feature vectors over time. This constitutes the original sequence of eigenvectors describing the evolution of the three-phase rheological state within the acetate reactor over time.

[0064] This step collects and integrates data reflecting the core dynamics of the acetate crystallization process from multiple physical dimensions, and generates a high-fidelity, high-dimensional, and time-synchronized original feature vector sequence through refined preprocessing techniques. This original feature vector sequence not only comprehensively quantifies the complex changes inside the reactor, from microscopic crystals and bubble events to macroscopic fluid properties, but also provides a high-quality and information-rich data foundation for subsequent steps to accurately extract rheological fingerprints and gain insights into the reaction state using deep learning models.

[0065] Step 2: Input the original feature vector sequence into the deep belief network model to extract the rheological feature fingerprints that characterize the growth state of acetate crystals and the frequency of bubble rupture. Map the high-dimensional rheological features to the low-dimensional state space through the manifold learning algorithm to quantify the dynamic correlation index between the viscosity change rate of the reaction liquid and the mass transfer coefficient of the gas-liquid interface.

[0066] After constructing the original feature vector sequence describing the three-phase rheological state of gas, liquid, and solid, the original feature vector sequence is input into a pre-trained deep belief network (DBN) model to achieve high-level abstract feature extraction of complex process data.

[0067] The Deep Belief Network (DBN) consists of multiple stacked Restricted Boltzmann Machines (RBMs). The RBM structure excels at unsupervised layer-by-layer feature learning. During offline training, the DBN is pre-trained using a large-scale sequence of raw feature vectors collected from an acetate reactor under various historical operating conditions (including normal crystallization, rapid crystal growth, intense solvent vaporization, and significant changes in material viscosity). Training employs a greedy layer-by-layer unsupervised learning method, using the contrastive divergence algorithm (CD-k) to independently train each layer of the RBM, ensuring that each hidden layer learns effective feature representations of its input layer data. The energy function of a RBM is described. Defined as: ;in, It is the state vector of the visible layer unit, corresponding to the original feature vector of the input. ; It is the state vector of the hidden layer unit; and These are the bias vectors for the visible layer and the hidden layer, respectively; It is the weight matrix connecting the visible layer and the hidden layer.

[0068] Deep belief networks can automatically learn a more compact and information-dense set of deep features from high-dimensional original features. In application, the sequence of original feature vectors generated in step one is used... As the input to the lowest layer RBM of the deep belief network, through the forward propagation of the network, it finally outputs a low-dimensional but highly abstract feature vector at the top hidden layer. This vector is the rheological fingerprint that characterizes the current growth state of acetate crystals (such as nucleation rate and changes in crystal size distribution) and the frequency of bubble bursting (reflecting the intensity of gas-liquid mass transfer and phase transition). This transformation from raw data to abstract fingerprints can effectively filter out redundant information and reveal the intrinsic connections between the core physical processes hidden behind multi-source data.

[0069] To further understand and quantify the dynamic evolution trajectory of the reaction process, the Isomap algorithm is used to reduce the dimensionality and visualize the rheological feature fingerprint sequence extracted by the deep belief network. Specifically, the core idea of ​​the Isomap algorithm is to assume that high-dimensional data points are distributed on a low-dimensional nonlinear manifold and to reduce dimensionality by preserving the geodesic distance between data points. The Isomap algorithm first maps each rheological feature fingerprint in the time series... Treat each point as a point in a high-dimensional space, and construct a neighborhood graph for each point, that is, connect each point to its nearest neighbor. The algorithm identifies several neighboring points. Then, it uses Dijkstra's or Floyd-Warshall's algorithms to calculate the shortest path length between any two points in the neighborhood graph, which serves as an approximation of the geodesic distance between them. Finally, it performs classical multidimensional scaling (MDS) analysis on the geodesic distance matrix to find a low-dimensional embedding space (exemplarily a two-dimensional or three-dimensional space) that best preserves these distances, thereby embedding the high-dimensional rheological fingerprint. Mapped to a point in a low-dimensional state space As time goes by, the point The trajectories formed in low-dimensional space intuitively demonstrate the evolution path of the overall state of the acetate reactor. In this way, the complex reaction process is transformed into a low-dimensional dynamic system that is easy to analyze, providing a basis for quantifying the correlation of key parameters.

[0070] Based on the reduced-dimensional state space, a dynamic correlation index between the viscosity change rate of the reaction liquid and the gas-liquid interface mass transfer coefficient is quantified. In the low-dimensional state space, different regions and directions of motion correspond to different dominant physicochemical processes. The factors most relevant to viscosity change and gas-liquid mass transfer in the low-dimensional space are identified, and two proxy variables are defined accordingly: the viscosity change rate proxy index. Gas-liquid interface mass transfer coefficient proxy index The viscosity change rate proxy index is calculated by using low-dimensional state points. The time derivative of the projected component in the pre-calibrated direction of viscosity change is obtained; the gas-liquid interface mass transfer coefficient proxy index is obtained through the state point. The projection component along the mass transfer change direction is used for estimation. Subsequently, a dynamic correlation index is constructed. The formula for quantifying the strength and relative relationship between these two interactions is as follows:

[0071] ;

[0072] in, It is a dynamic correlation index; It is a proxy index for the rate of change in the viscosity of the reaction liquid, reflecting the rate at which solid-phase crystallization precipitation affects the fluid properties; It is a proxy index for the gas-liquid interface mass transfer coefficient, reflecting the intensity of gas phase generation and escape. It is a very small positive number set to prevent the denominator from being zero.

[0073] For example, suppose that, through calibration, the viscosity change rate proxy index is correlated with the low-dimensional state vector. The relationship is , and These are two components of a low-dimensional state vector, and the surrogate index for the gas-liquid interface mass transfer coefficient is... If in time, At the previous moment (time interval) ), ,and ,but ,and Therefore, the dynamic correlation index .

[0074] This step transforms multi-source, heterogeneous sensor data into a single, highly condensed index—the Dynamic Correlation Index—through progressive abstraction and refinement using deep learning and manifold learning. This index precisely and intuitively reflects the dynamic coupling relationship between the two core physical phenomena of solid-phase precipitation and gas-phase formation during acetate crystallization. The Dynamic Correlation Index greatly simplifies the understanding and judgment of complex reaction process states, providing crucial, highly condensed decision-making support for subsequently constructing accurate pressure evolution models and achieving precise control.

[0075] Step 3: Construct a phase transition coupling evolution model of reactor pressure based on the dynamic correlation index, use the extended Kalman filter algorithm to estimate the gas phase escape flux deviation during the crystallization process in real time, and combine it with the long short-term memory network to predict the pressure change trend within the future time window, generating a pressure prediction trajectory containing phase transition interference factors.

[0076] The dynamic correlation index between the rate of change of the reaction liquid viscosity and the mass transfer coefficient at the gas-liquid interface was obtained. Subsequently, a phase change coupled evolution model describing the dynamic changes in pressure inside the reactor was constructed. This model is a nonlinear state-space model, its core being the combination of macroscopic thermodynamic equilibrium relationships with microscopic phase change kinetic disturbances characterized by dynamic correlation indices. The establishment of the phase change coupled evolution model is based on the conservation of mass and energy in the gas phase space at the top of the reactor, and incorporates the pressure change rate... It is expressed as a function of the net formation rate of gaseous substances. The specific equation of state is constructed as follows:

[0077] ;

[0078] in, Is the reactor at a certain time? Internal pressure; It is the instantaneous rate of change of pressure; It is the ideal gas constant; It is the temperature of the gas phase space in the reactor, which is obtained from basic thermodynamic parameters. It is the instantaneous rate of change of temperature in the gas phase space; It is the effective volume of the gas phase space of the reactor, which is considered as an approximately constant parameter; It is the gas molar flux removed through the exhaust valve and condenser, and is the direct manipulated object of the control process; It is the total molar flux of gas generated within the reactor, and is the primary endogenous driving force for pressure. Crucially, It is modeled as the superposition of the basic reaction kinetics term and the phase transition coupling perturbation term, i.e. ,in It depends on the reactant concentration. and temperature The basic gas generation rate, and The term represents the nonlinear gas generation fluctuation caused by the coupling between crystal growth (affecting viscosity and interface renewal) and bubble behavior (affecting mass transfer). The coupling coefficient is identified through historical data. A dynamic correlation index is introduced. This evolutionary model can more realistically reflect the abrupt changes in the mass transfer characteristics of the phase interface caused by the intensification or deceleration of solid phase precipitation, and thus the impact on pressure.

[0079] Furthermore, to accurately grasp the unknown disturbances in the actual exhaust process, the Extended Kalman Filter (EKF) algorithm is used to estimate the gas phase escape flux deviation during crystallization in real time. The gas phase escape flux deviation is defined as the difference between the actual exhaust flux and the theoretical exhaust flux calculated based on valve opening and pressure difference. It is mainly caused by factors that are difficult to measure directly, such as pressure tap blockage, valve sticking, or fluctuations in condensation efficiency. Therefore, the gas phase escape flux deviation to be estimated is... The vector is augmented to the system state vector to form a new state vector. Assuming flux deviation is a slowly varying stochastic process over time, its dynamics can be modeled as a random walk. Combining this with the aforementioned pressure evolution model, the following nonlinear discrete-time state transition equations and observation equations are constructed:

[0080] ;

[0081] in, yes The state vector at any given time; It is a control input vector that includes the theoretical exhaust flux; It is a nonlinear state transition function discretized from the pressure evolution model; It is process noise, and its covariance matrix is... It characterizes the uncertainty of the model; yes The observed value at a given time, i.e., the actual measurement value of the pressure sensor. ; It is the observation function, in this case, ; It is observation noise, and its covariance matrix is... This characterizes the sensor's measurement error. The extended Kalman filter algorithm iterates through prediction steps (predicting the current state using the state transition equation) and update steps (correcting the predicted state using actual observations and Kalman gain), outputting a real-time state vector. Optimal estimate In this way, a smoothed estimate of the actual pressure is obtained. More importantly, it yielded a key perturbation that cannot be directly measured, namely the gas phase escape flux deviation. The real-time estimated value.

[0082] This study combines a pre-trained Long Short-Term Memory (LSTM) network to predict nonlinear stress shifts within a specific future time window (e.g., the next 30 seconds). LSTM networks, due to their unique gating mechanisms (input gate, forget gate, output gate), are particularly well-suited for learning and predicting time-series data with long-term dependencies and complex nonlinear dynamics. A feature sequence is constructed from multivariate information from the current and multiple past sampling times and used as input to the LSTM network. The input feature vector sequence at time... The composition is ,in This is the actual measured pressure value. It is the gas-phase escape flux deviation estimated by the extended Kalman filter. It is a dynamic correlation index, and and These are the two main control variables: the opening degree of the exhaust regulating valve and the flow rate of the condenser cooling medium. This is achieved through a dynamic correlation index that incorporates core information about phase change disturbances. and the deviation in gas phase escape flux reflecting actuator abnormalities Simultaneously, as input, the Long Short-Term Memory (LSTM) network can learn how these disturbance factors trigger future stress fluctuations. For example, if the current time is... The network will input from arrive of The feature vector sequence at each time step, output from arrive of A sequence of predicted pressure values ​​for future times. , This represents the predicted stress value at the first future moment. Indicates the first The output sequence represents the predicted pressure values ​​at future moments. This sequence constitutes the predicted pressure trajectory, which includes the influence of the phase transition disturbance factor.

[0083] This step deeply integrates the mechanistic model with the data-driven model, constructing a phase transition coupled evolution model that provides a framework based on physicochemical principles. Extended Kalman filtering then enables precise visualization of key unknown perturbations. Finally, using this real-time information through a long short-term memory network, it makes forward-looking predictions about future nonlinear behavior. This allows the control mechanism of the acetate reactor pressure regulation process to move beyond a merely lagging response based on the current state, enabling it to anticipate pressure surges caused by complex factors such as crystallization phase transitions.

[0084] Step 4: Compare the pressure prediction trajectory with the set target value using differential comparison, calculate the pressure compensation control quantity for the phase transition critical point using an adaptive fuzzy sliding mode controller, dynamically adjust the opening gain coefficient of the exhaust valve according to the gas phase escape flux deviation, and generate a feedforward control command sequence to decouple the influence of feed flow rate fluctuations.

[0085] Obtaining the stress prediction trajectory within future time windows predicted by the Long Short-Term Memory Network. Then, the pressure prediction trajectory is compared with the preset pressure target value. Differential comparisons are performed to generate a prediction error sequence. Specifically, in each control cycle... Calculate the error in the prediction time domain ,in From 1 to The prediction step size. This error sequence not only reflects the degree of deviation from the current state, but more importantly, it predicts the trend and magnitude of future pressure changes, providing crucial predictive information for the controller's decision-making.

[0086] Next, an adaptive fuzzy sliding mode controller (AFSMC) is used to calculate the pressure compensation control quantity for strongly nonlinear disturbances at the phase transition critical point. A sliding mode surface function is defined. Sliding surface function Incorporating both the pressure prediction error and its rate of change, its expression is as follows: ;in, It is the pressure prediction error at the current moment, i.e. ; It is the rate of change of the prediction error, which is obtained by difference calculation of the prediction error sequence; As a positive constant, its value determines the rate at which the system state of the adaptive fuzzy sliding mode controller converges to the sliding surface. Pressure compensation control variable. From the equivalent control part and switching control section Composition. Equivalent control Used to maintain the system state on the sliding surface, while switching control This is used to quickly drive the system state from any position to the sliding surface, and its basic form is: ,in To switch the gain, This is the sign function. To suppress the chattering problem inherent in traditional sliding mode control and adaptively address the time-varying characteristics of parameters during the phase transition process, the gain is switched. Instead of a fixed value, it is adjusted online via a fuzzy sliding mode controller. This fuzzy inference uses a sliding surface function. and its derivative As two inputs, the data is processed through a pre-defined fuzzy rule base (e.g., if...). large and Large, then Increase; if Small, then (Reduce), and output a dynamically switching gain in real time. This design enables the controller to exhibit strong robustness and rapid response when far from the equilibrium point, while becoming smoother as it approaches the equilibrium point. This allows for the calculation of pressure compensation control quantities that can quickly compensate for pressure fluctuations while avoiding excessive wear on the actuator. .

[0087] Based on the calculated pressure compensation control quantity, the gas phase escape flux deviation estimated by the extended Kalman filter algorithm in step three is further used. The opening gain coefficient of the exhaust valve is dynamically adjusted. A reference gain coefficient is defined. The dynamic gain coefficient that actually acts on the control command : ;in, It is the sensitivity coefficient that is adjusted. When the estimated gas-phase escape flux deviation... When the value is positive (indicating that the actual exhaust volume is greater than the theoretical calculation value, which may indicate internal leakage or poor sealing of the valve), the opening gain coefficient is automatically reduced. This makes the control action more conservative, preventing pressure over-adjustment; conversely, when the deviation is negative (indicating that the actual exhaust volume is less than the theoretical value, possibly indicating valve sticking or a precursor to pressure tap blockage), the opening gain coefficient is increased to ensure the exhaust command is effectively executed with stronger control. This mechanism effectively compensates for changes in the physical characteristics of the actuator and unmodeled disturbances, significantly improving the adaptability and robustness of the control process.

[0088] To eliminate periodic or random disturbances introduced by the feed operation during acetate synthesis, a feedforward control command sequence decoupled from the effects of feed flow rate fluctuations is generated. This is achieved by online monitoring of the instantaneous flow rate of the feed pump. It is based on a pre-identified feedforward model that describes the dynamic relationship between the feed rate and the reactor pressure response. Calculate the required feedforward control quantity For example, if the feedforward model is a simple first-order element with a delay, then the feedforward control quantity can be obtained from... Calculations show that the pressure changes caused by temperature variations and reaction rate fluctuations due to the addition of new materials are used to preemptively offset these changes. Finally, the pressure compensation control quantity output by the adaptive fuzzy sliding mode controller is... The result after dynamic gain adjustment, and the feedforward control quantity The commands are superimposed to form the final feedforward-feedback composite control command. This sequence of instructions is sent to the downstream actuators, enabling high-performance control of the pressure.

[0089] This step constructs an advanced control strategy combining predictive, adaptive, and feedforward compensation. Pressure prediction trajectories are used in controller design, enabling proactive control actions that effectively address the lag in complex processes such as phase transitions. The application of an adaptive fuzzy sliding mode controller ensures robust control and rapid response while also resolving control chattering issues. Furthermore, online compensation for gas phase escape flux deviation and feedforward decoupling of feed disturbances further enhance adaptability to real-world operating conditions. This allows pressure control to proactively predict and offset disturbances, rather than passively responding to errors, thus providing strong technical support for achieving high-precision and high-stability automatic pressure regulation in acetate reactors.

[0090] Step 5: Based on the feedforward control command sequence, adjust the flow rate of the condenser cooling medium and the opening of the exhaust regulating valve. When the predicted pressure value deviates from the safety threshold, trigger the pulse backflushing mechanism to clean the pressure tap, and perform pressure adjustment according to the corrected control parameters. See also... Figure 3 This is a schematic diagram of the collaborative decoupling execution and pulse backflush mechanism for this step.

[0091] Upon receiving the feedforward-feedback composite control command sequence generated in step four Next, this step involves processing the feedforward-feedback composite control command sequence. The signal is analyzed and distributed to the actuators: the condenser cooling medium flow regulator and the reactor top exhaust regulating valve. This process is not a simple signal equalization, but rather employs a collaborative decoupling control strategy based on the current thermodynamic state of the reactor. The rate of temperature change within the reactor is analyzed. With pressure change rate By introducing a dynamic decoupling weighting factor , will send master control command Decomposed into control components for cooling medium flow rate and control components for exhaust valve opening The weighting factor It dynamically changes within the range [0,1], and its value is correlated with the evaporation load inside the vessel. When a pressure increase is detected, it is mainly accompanied by a sharp temperature rise (i.e., A significantly positive value indicates that the pressure fluctuation is mainly caused by the large-scale evaporation of the solvent due to the exothermic reaction or crystallization. In this case, increasing the value... The value of [value] allocates greater control weight to the condenser. Conversely, if the pressure rises while the temperature remains relatively stable, it is judged to be due to the accumulation of non-condensable gases or the escape of product gaseous components, and the [value] is reduced. This allows the exhaust regulating valve to undertake the main regulating task. The specific control components are calculated as follows:

[0092] ;

[0093] in, It is the control component allocated to the flow rate of the cooling medium in the cooler. It is the control component assigned to the exhaust regulating valve. These are dynamic decoupling weighting factors. These two control components are then converted into specific setpoints for the corresponding actuators, for example, by... Convert the inverter frequency setting value of the cooling water circulating pump or the opening setting value of the cooling water regulating valve to the set value. This is converted to the valve position percentage setpoint of the exhaust control valve. The cooperative decoupling method can precisely select the most effective control method based on the root cause of the disturbance, avoiding the saturation of the single actuator's regulating capacity and optimizing energy utilization efficiency.

[0094] At the same time, the pressure predicted in step three is used to predict the trajectory. Implement proactive safety monitoring. Set pressure safety thresholds, including high-pressure safety thresholds. and low-voltage safety threshold In each control cycle, examine all pressure forecasts within the entire forecast time window. When any forecast point is found... (in If any safety threshold is reached or exceeded, a preset pulse backflush mechanism is triggered. This mechanism is designed to address potential blockages in the pressure measurement taps caused by acetate slurry or high-viscosity materials. Such blockages can distort pressure sensor readings, leading to control failure. The pulse backflush mechanism injects a burst of high-pressure inert gas (such as nitrogen) instantaneously (e.g., for 100-200 milliseconds) into the tap by controlling a high-speed solenoid valve connected to the pressure tap line. This instantaneous high-pressure gas flow flushes and cleans any potential deposits from the tap's inner wall. This predictive action intervenes before the actual pressure reaches a dangerous level, ensuring the long-term reliability and accuracy of the pressure measurement signal.

[0095] Based on the decoupled control components and The control platform (such as a distributed control system (DCS) or a programmable logic controller (PLC) issues high-time-resolution adjustment commands to the field actuators according to the corrected control parameters. The corrected control parameters refer to the final actuator setpoints generated after feedforward-feedback composite control, dynamic gain adjustment, and collaborative decoupling strategies. The control platform typically adjusts the flow rate of the cooling medium and the instantaneous opening of the exhaust regulating valve with a millisecond-level response speed. For example, it sends a 4-20mA signal or fieldbus command to the valve intelligent positioner, enabling it to accurately reach the target opening degree in a very short time. Thus, from multi-source data acquisition, status identification, trend prediction, intelligent decision-making to final collaborative execution, a complete, fast, and intelligent closed-loop automatic adjustment loop is formed, achieving precise control of the acetate reactor pressure.

[0096] This step efficiently and reliably translates the decision-making results of advanced upper-level control algorithms into precise control of physical equipment. The collaborative decoupling control strategy enhances the targeting and resource utilization of control; the prediction-based pulse backflush mechanism fundamentally improves the robustness and safety of the reaction process, preventing safety accidents caused by sensor malfunctions. Rapid execution ensures that the entire closed-loop control process can effectively cope with complex phase transition disturbances during acetate crystallization, maintaining the reactor pressure stably within the optimal process window. This has significant practical implications for ensuring product quality uniformity and improving production safety.

[0097] Example 2

[0098] Reference Figure 4 This is the second embodiment of the present application, which provides an automatic pressure regulation system for an acetate reactor.

[0099] The system includes: a multi-source information characterization module, a reaction state identification module, a pressure trend prediction module, a control strategy decision-making module, and a collaborative regulation execution module.

[0100] The multi-source information characterization module collects multi-source data from the acetate reactor, preprocesses the collected multi-source data, and constructs an original feature vector sequence describing the gas-liquid-solid three-phase rheological state.

[0101] The reaction state identification module inputs the original feature vector sequence into a deep belief network model, extracts rheological feature fingerprints that characterize the growth state of acetate crystals and the frequency of bubble rupture, and maps the high-dimensional rheological features to a low-dimensional state space through a manifold learning algorithm, quantifying the dynamic correlation index between the viscosity change rate of the reaction liquid and the mass transfer coefficient at the gas-liquid interface.

[0102] The pressure trend prediction module constructs a phase transition coupling evolution model of reactor pressure based on a dynamic correlation index, uses an extended Kalman filter algorithm to estimate the gas phase escape flux deviation during the crystallization process in real time, and combines a long short-term memory network to predict the pressure mutation trend within the future time window, generating a pressure prediction trajectory that includes phase transition interference factors.

[0103] The control strategy decision module performs differential comparison between the pressure prediction trajectory and the set target value, calculates the pressure compensation control quantity for the phase transition critical point through an adaptive fuzzy sliding mode controller, dynamically adjusts the opening gain coefficient of the exhaust valve according to the gas phase escape flux deviation, and generates a feedforward control command sequence to decouple the influence of feed flow rate fluctuations.

[0104] The coordinated adjustment execution module adjusts the flow rate of the condenser cooling medium and the opening of the exhaust regulating valve according to the feedforward control command sequence. When the pressure prediction value deviates from the safety threshold, it triggers the pulse backflushing mechanism to clean the pressure tap and performs pressure adjustment according to the corrected control parameters.

[0105] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings or direct couplings or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0106] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of this application without departing from the spirit and scope of protection of the claims. All of these variations are within the protection scope of this application.

Claims

1. A method for automatic adjustment of pressure in an acetic acid reactor characterized by, include: Step 1: Collect multi-source data from the acetate reactor, preprocess the collected multi-source data, and construct an original feature vector sequence describing the gas-liquid-solid three-phase rheological state; Step 2: Input the original feature vector sequence into the deep belief network model to extract the rheological feature fingerprints that characterize the growth state of acetate crystals and the frequency of bubble rupture. Map the high-dimensional rheological features to the low-dimensional state space through the manifold learning algorithm to quantify the dynamic correlation index between the viscosity change rate of the reaction liquid and the mass transfer coefficient of the gas-liquid interface. Step 3: Construct a phase transition coupling evolution model of reactor pressure based on dynamic correlation index, use extended Kalman filter algorithm to estimate gas phase escape flux deviation in real time during crystallization, combine long short-term memory network to predict pressure change trend in future time window, and generate pressure prediction trajectory containing phase transition interference factor. Step 4: Compare the pressure prediction trajectory with the set target value using differential comparison, calculate the pressure compensation control quantity for the phase transition critical point using an adaptive fuzzy sliding mode controller, dynamically adjust the opening gain coefficient of the exhaust valve according to the gas phase escape flux deviation, and generate a feedforward control command sequence to decouple the influence of feed flow rate fluctuations. Step 5: Based on the feedforward control command sequence, adjust the flow rate of the condenser cooling medium and the opening of the exhaust regulating valve. When the pressure prediction value deviates from the safety threshold, trigger the pulse backflushing mechanism to clean the pressure tap and perform pressure adjustment according to the corrected control parameters.

2. The method for automatic pressure adjustment of an acetate reactor according to claim 1, characterized in that, Multi-source data were collected from the acetate reactor, including acoustic emission data, stirring motor torque ripple data, and thermodynamic parameters; the collected multi-source data were preprocessed, including time synchronization and wavelet packet decomposition noise reduction. Based on the preprocessed multi-source data, an original feature vector sequence describing the three-phase rheological state of gas, liquid and solid is constructed. At the aligned time point, the root mean square value, signal energy and peak amplitude of the acoustic emission signal are extracted from the noise-reduced signal. The statistical features of standard deviation, kurtosis and skewness are extracted from the noise-reduced torque ripple signal. At the same time, the instantaneous measurement values ​​of the basic thermodynamic parameters are used as feature quantities. All features extracted at the aligned time points will be combined to form a high-dimensional original feature vector.

3. The method for automatic pressure adjustment of an acetate reactor according to claim 2, characterized in that, After constructing the original feature vector sequence describing the three-phase rheological state of gas, liquid and solid, the original feature vector sequence is input into the pre-trained deep belief network model, which is composed of multiple stacked restricted Boltzmann mechanisms. The deep belief network model utilizes a large-scale sequence of original feature vectors collected from acetate reactors under various historical operating conditions to pre-train the deep belief network. The historical operating conditions include the normal crystallization period, the rapid crystal growth period, the period of intense solvent vaporization, and the period of significant changes in material viscosity. The training adopts a greedy layer-by-layer unsupervised learning method, that is, using the contrastive divergence algorithm to independently train each layer of the restricted Boltzmann machine. The original feature vector sequence is used as the input of the lowest layer restricted Boltzmann machine of the deep belief network. Through the forward propagation of the network, the feature vector is output in the top hidden layer. The feature vector is a rheological fingerprint that characterizes the growth state of acetate crystal and the frequency of bubble rupture at the current moment. The isomap algorithm is used to reduce the dimensionality and visualize the rheological feature fingerprint sequence extracted by the deep belief network.

4. The method for automatic pressure adjustment of an acetate reactor according to claim 3, characterized in that, Based on the reduced-dimensional state space, the dynamic correlation index between the viscosity change rate of the reaction liquid and the mass transfer coefficient at the gas-liquid interface is quantified. Two proxy variables are defined for the dynamic correlation index: the viscosity change rate proxy index and the gas-liquid interface mass transfer coefficient proxy index. The viscosity change rate proxy index is obtained by calculating the time derivative of the projection component of the low-dimensional state point in the pre-calibrated viscosity change direction; the gas-liquid interface mass transfer coefficient proxy index is estimated by the projection component of the point in the low-dimensional state space in the mass transfer change direction. The interaction strength and relative relationship between the viscosity change rate proxy index and the gas-liquid interface mass transfer coefficient proxy index were constructed by constructing a dynamic correlation index.

5. The method for automatic pressure adjustment of an acetate reactor according to claim 4, characterized in that, A phase transition coupled evolution model is constructed to describe the dynamic changes in pressure inside the reactor, and the pressure change rate is expressed as a function of the net generation rate of gaseous substances. The extended Kalman filter algorithm is used to estimate the gas phase escape flux deviation during the crystallization process. The gas phase escape flux deviation is defined as the difference between the actual exhaust flux and the theoretical exhaust flux calculated based on valve opening and pressure difference. By combining the phase transition coupled evolution model, nonlinear discrete-time state transition equations and observation equations are constructed; the estimated value of gas phase escape flux deviation is obtained by using the extended Kalman filter algorithm.

6. The method for automatic pressure adjustment of an acetate reactor according to claim 5, characterized in that, By combining a pre-trained long short-term memory network to predict the nonlinear stress change trend within future time windows, multivariate information from multiple sampling moments in the present and past is constructed into a feature sequence, which is then used as the input to the long short-term memory network. The characteristic sequence at a single moment includes: measured pressure value, gas phase escape flux deviation, dynamic correlation index, opening degree of exhaust regulating valve and flow rate of condenser cooling medium; The Long Short-Term Memory Network outputs a sequence of predicted stress values ​​for future moments.

7. The method for automatic pressure adjustment of an acetate reactor according to claim 6, characterized in that, The predicted pressure trajectory is compared with the preset pressure target value to generate a prediction error sequence. An adaptive fuzzy sliding mode controller is used to calculate the pressure compensation control quantity for strong nonlinear disturbances at the phase transition critical point; the opening gain coefficient of the exhaust valve is dynamically adjusted according to the gas phase escape flux deviation, and a reference gain coefficient is defined. When the estimated gas phase escape flux deviation is positive, the opening gain coefficient is automatically reduced; when the deviation is negative, the opening gain coefficient is increased.

8. The method for automatic pressure adjustment of an acetate reactor according to claim 7, characterized in that, Generate a feedforward control command sequence to decouple the effects of feed flow rate fluctuations. Calculate the required feedforward control quantity by monitoring the instantaneous flow rate of the feed pump online and based on a feedforward model that describes the dynamic relationship between the feed flow rate and the reactor pressure response. The pressure compensation control quantity output by the adaptive fuzzy sliding mode controller and the result after dynamic gain adjustment are superimposed with the feedforward control quantity to form a feedforward-feedback composite control command.

9. The method for automatic pressure adjustment of an acetate reactor according to claim 8, characterized in that, The feedforward-feedback composite control command sequence is parsed and assigned to the actuators, including the condenser cooling medium flow regulator and the reactor top exhaust regulating valve; A collaborative decoupling control strategy based on the current thermodynamic state of the reactor was adopted to analyze the rate of temperature change and the rate of pressure change within the reactor. By introducing a dynamic decoupling weighting factor, the feedforward-feedback composite control command sequence is decomposed into a control component for the cooling medium flow rate and a control component for the exhaust valve opening. Safety monitoring is performed using the predicted pressure trajectory. Pressure safety thresholds are set, including high-pressure safety thresholds and low-pressure safety thresholds. In each control cycle, all predicted pressure values ​​within the entire prediction time window are checked. When the pressure at any predicted time point is found to be outside the range of the high-pressure safety threshold and the low-pressure safety threshold, the pulse backflush mechanism is triggered. Based on the decoupled control components, the flow rate of the cooling medium supply and the instantaneous opening of the exhaust regulating valve are adjusted.

10. An automatic pressure regulation system for an acetate reactor, used to implement the automatic pressure regulation method for an acetate reactor as described in any one of claims 1 to 9, characterized in that, include: The system includes a multi-source information characterization module, a reaction state identification module, a pressure trend prediction module, a control strategy decision-making module, and a coordinated regulation and execution module. The multi-source information characterization module collects multi-source data from the acetate reactor, preprocesses the collected multi-source data, and constructs an original feature vector sequence describing the gas-liquid-solid three-phase rheological state. The reaction state identification module inputs the original feature vector sequence into the deep belief network model, extracts the rheological feature fingerprints that characterize the growth state of acetate crystals and the frequency of bubble rupture, and maps the high-dimensional rheological features to the low-dimensional state space through the manifold learning algorithm, quantifying the dynamic correlation index between the viscosity change rate of the reaction liquid and the mass transfer coefficient of the gas-liquid interface. The pressure trend prediction module constructs a phase transition coupling evolution model of reactor pressure based on dynamic correlation index, uses extended Kalman filter algorithm to estimate gas phase escape flux deviation in real time during crystallization, and combines long short-term memory network to predict pressure change trend within future time window, generating a pressure prediction trajectory containing phase transition interference factor. The control strategy decision module compares the pressure prediction trajectory with the set target value, calculates the pressure compensation control quantity for the phase transition critical point through an adaptive fuzzy sliding mode controller, dynamically adjusts the opening gain coefficient of the exhaust valve according to the gas phase escape flux deviation, and generates a feedforward control command sequence to decouple the influence of feed flow rate fluctuations. The coordinated adjustment execution module adjusts the flow rate of the condenser cooling medium and the opening of the exhaust regulating valve according to the feedforward control command sequence. When the pressure prediction value deviates from the safety threshold, it triggers the pulse backflushing mechanism to clean the pressure tap and performs pressure adjustment according to the corrected control parameters.

Images