PTA mixed feeding control method and system

Abstract

The invention relates to the technical field of PTA production control, and discloses a PTA mixed feeding control method and system. The system comprises a mechanism modeling module, a data driving prediction module, a dynamic fusion module, a rolling optimization module and an online learning module. The mechanism modeling module constructs a sub-model based on a PTA mixed reaction kinetic equation, and outputs a theoretical feeding constraint value; the data driving prediction module receives temperature, pressure and raw material flow data collected by the real-time sensor and outputs a real-time feeding prediction value. The dynamic fusion module performs weighted fusion on the two values to generate a fused feeding prediction value; the rolling optimization module generates a feeding parameter adjustment instruction through a rolling optimization algorithm. And when the deviation between the adjusted sensor data and the fused predicted value exceeds a preset threshold value, the online learning module performs online fine adjustment on the parameters of the data driving model. The system combines a reaction mechanism and real-time working condition optimization feeding control, reduces manual intervention, and adapts to a complex production environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of PTA production control, in particular to a PTA mixed feeding control method and system. BACKGROUND

[0002] In the production process of purified terephthalic acid (PTA), the stability of the mixed feeding link directly affects the subsequent reaction efficiency and product quality. Traditional PTA mixed feeding control relies mainly on a single control method, which is difficult to adapt to complex and variable production conditions. Among them, the control method based on mechanism modeling is more common. This kind of method determines the feeding parameter range by constructing the PTA mixed reaction kinetics equation. Although it can provide basic constraints according to the theoretical reaction law, there are many variables in actual production that cannot be completely covered by the mechanism model. For example, fluctuations in raw material purity, local temperature differences caused by equipment aging, and sensor transient interference factors will cause deviations between the theoretical feeding constraint value output by the mechanism model and the actual production requirements. If only the constraint value is used for feeding control, the reaction may not be complete or raw materials may be wasted.

[0003] With the development of industrial automation technology, some PTA production enterprises have begun to use data-driven models for feeding prediction. This kind of model can capture some implicit correlations in the production process by collecting historical production data for training, but its limitations are also obvious. The prediction effect of the data-driven model is highly dependent on the integrity and representativeness of the training data. When the production conditions change suddenly, such as raw material type replacement and production load adjustment, the relevance of historical data and real-time data decreases, and the accuracy of the real-time feeding prediction value output by the model will decrease significantly. In addition, the data-driven model lacks consideration of the internal mechanism of the PTA mixed reaction, and may produce prediction values that conform to the data law but violate the reaction kinetics principle, resulting in incorrect adjustment of the feeding parameters and affecting the production continuity.

[0004] In actual production, the traditional control method also has the problem of insufficient dynamic adjustment capability. Whether it is mechanism modeling or data-driven modeling, when used alone, it is difficult to realize real-time optimization of the feeding parameters. When deviations occur in the production process, manual intervention is required to adjust the parameters, which not only slows down the response speed, but also may result in inconsistent adjustment effects due to differences in the experience of the operators. At the same time, the existing control scheme lacks an effective online learning mechanism. Once the model parameters are determined, they are difficult to adaptively update according to the changes in real-time production data. The model accuracy gradually decreases over time, and cannot meet the requirements of PTA production for stability and refinement. The existence of these problems makes it difficult for the traditional PTA mixed feeding control method to meet the needs of modern production, and a new type of control system that can integrate the advantages of multiple control methods and has dynamic adjustment and online optimization capabilities is needed. SUMMARY

[0005] The present application aims to provide a PTA mixed feed control method and system to solve the problems in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides a PTA mixed feed control system, which comprises:

[0007] a mechanism modeling module for constructing a mechanism sub-model based on a PTA mixed reaction kinetics equation and outputting a theoretical feed constraint value;

[0008] a data-driven prediction module for receiving temperature, pressure and raw material flow data collected by real-time sensors through a data-driven model and outputting a real-time feed prediction value;

[0009] a dynamic fusion module for weighting and fusing the theoretical feed constraint value output by the mechanism sub-model and the real-time feed prediction value output by the data-driven model to generate a fused feed prediction value;

[0010] a rolling optimization module for generating a feed parameter adjustment instruction through a rolling optimization algorithm according to the fused feed prediction value;

[0011] an online learning module for online fine-tuning parameters of the data-driven model when the deviation between the sensor data collected after executing the feed parameter adjustment instruction and the fused feed prediction value exceeds a preset threshold.

[0012] Preferably, the mechanism modeling module constructs a mechanism sub-model based on a PTA mixed reaction kinetics equation and outputs a theoretical feed constraint value, which comprises:

[0013] setting temperature, pressure and raw material concentration process boundary conditions according to a reaction rate equation of a PTA slurry preparation process;

[0014] solving a theoretical feed constraint value through a material balance equation based on the process boundary conditions.

[0015] Preferably, the data-driven prediction module receives temperature, pressure and raw material flow data collected by real-time sensors through a data-driven model and outputs a real-time feed prediction value, which comprises:

[0016] inputting the temperature, pressure and raw material flow data into a multivariate time series neural network to extract dynamic coupling features;

[0017] outputting a real-time feed prediction value with the same dimension as the theoretical feed constraint value through a regression layer based on the dynamic coupling features.

[0018] Preferably, the dynamic fusion module weights and fuses the theoretical feed constraint value and the real-time feed prediction value to generate a fused feed prediction value, which comprises:

[0019] According to the real-time working condition error, the weight coefficients of the mechanism sub-model and the weight coefficients of the data-driven model are calculated through an adaptive weight distribution algorithm;

[0020] Based on the weight coefficients, the theoretical feed constraint value and the real-time feed prediction value are weighted and summed to generate the fused feed prediction value.

[0021] Preferably, the rolling optimization module generates a feed parameter adjustment instruction based on the fused feed prediction value through a rolling optimization algorithm, including:

[0022] Based on the fused feed prediction value within the rolling time window, a multi-objective optimization function is updated;

[0023] By solving the multi-objective optimization function, the current optimal temperature control value and the raw material flow set value are obtained to generate a feed parameter adjustment instruction.

[0024] Preferably, when the deviation between the adjusted sensor data collected after executing the feed parameter adjustment instruction and the fused feed prediction value exceeds a preset threshold, the online learning module includes:

[0025] According to the adjusted temperature, pressure and raw material flow data, an actual feed value is calculated through a material metering model;

[0026] The relative deviation absolute value of the actual feed value and the fused feed prediction value is calculated, and when the absolute value exceeds a set threshold, online fine tuning is triggered.

[0027] Preferably, the online learning module performs online fine tuning on the parameters of the data-driven model, including:

[0028] The parameters of the mechanism sub-model and the hidden layer parameters of the data-driven model are frozen;

[0029] Based on the adjusted temperature, pressure and raw material flow data, the gradient of the output layer parameters of the data-driven model is calculated;

[0030] According to the gradient, the output layer parameters are updated until the relative deviation absolute value of the actual feed value and the fused feed prediction value is lower than a set threshold.

[0031] Preferably, the system further includes:

[0032] An intelligent detector networking module for multi-agent collaborative detection and analysis of the PTA mixed feed process;

[0033] The intelligent detector networking module performs work rationality analysis based on the decision divergence point sequence of multiple detectors to generate a process state evaluation result.

[0034] Preferably, the intelligent detector networking module works on the basis of decision divergence point sequence of multiple detectors, including:

[0035] Selecting a reference divergence point from the decision divergence point sequence;

[0036] Tracing the operation trajectory of the detector corresponding to the reference divergence point in the historical period;

[0037] Traversing the trajectory features in the operation trajectory, and when the traversed trajectory feature matches the preset trigger feature, starting the replay verification mechanism;

[0038] Generating the process state evaluation result based on the replay verification result.

[0039] Preferably, the present application also includes a PTA mixed feed control method, which includes all modules and method processes of the above-mentioned PTA mixed feed control system.

[0040] Compared with the prior art, the present application has the following beneficial effects:

[0041] Through the synergistic effect of the mechanism modeling module and the data-driven prediction module, the limitations of traditional single control mode are effectively overcome. The mechanism modeling module constructs a mechanism sub-model based on the PTA mixed reaction kinetics equation, which can provide theoretical feed constraint values from the nature of the reaction, ensure that the feed parameter adjustment conforms to the inherent law of the reaction, avoid control instructions that violate the principles of kinetics, and provide a basic guarantee for the stability of the production process; and the data-driven prediction module can capture various dynamic variables and hidden correlations in the production process by receiving temperature, pressure and raw material flow data collected by real-time sensors, and the output real-time feed prediction value can reflect the actual working condition changes in time, and the combination of the two makes the determination of the feed parameters have both theoretical basis and actual production needs.

[0042] The dynamic fusion module generates a fused feed prediction value by weighting the theoretical feed constraint value and the real-time feed prediction value. This process is not a simple numerical superposition, but a dynamic weight distribution according to the reliability of the two prediction results under different working conditions. When the working condition is stable and the raw materials and equipment state meet the assumption conditions of the mechanism model, the weight of the theoretical feed constraint value can be increased to ensure the theoretical rigor of the control decision; when the working condition fluctuates, such as changes in raw material purity or abnormal temperature fluctuations detected by the sensor, the weight of the real-time feed prediction value can be increased to quickly respond to changes in working conditions, avoid control lag caused by single dependence on theoretical values, and realize flexible adaptation to complex and variable production environments.

[0043] The rolling optimization module generates the feedstock parameter adjustment instruction through the rolling optimization algorithm according to the fused feedstock prediction value, and can decompose the long-term control target into a series of short-term optimization problems. In each control cycle, the algorithm will combine the latest fused prediction value with the current production state to recalculate and adjust the feedstock parameters, rather than using a fixed control strategy. This method can effectively reduce the impact of cumulative deviation on the production process, ensure that the feedstock parameters are always in the optimal interval, improve the reaction efficiency, reduce the waste of raw materials, and reduce the risk of production failure caused by parameter deviation.

[0044] The setting of the online learning module further improves the adaptive ability of the system. When the deviation between the adjusted sensor data collected after the execution of the feedstock parameter adjustment instruction and the fused feedstock prediction value exceeds the preset threshold, the system can automatically fine-tune the parameters of the data-driven model online. This process does not require human intervention and can timely correct the deviation of the model caused by changes in working conditions or data drift, ensuring that the data-driven model maintains high prediction accuracy for a long time. Through continuous online learning, the model can gradually adapt to the change law under different production conditions, further enhancing the reliability of the real-time feedstock prediction value and providing higher-quality input data for the dynamic fusion module, forming a closed-loop mechanism of "prediction-control-feedback-optimization".

[0045] The entire control system realizes the automation and intelligentization of PTA mixed feedstock control, reduces the dependence on manual operation, and avoids the problem of inconsistent control effect caused by the experience difference of operators. The system can process multi-dimensional sensor data in real time, quickly generate and execute adjustment instructions, and has a much faster response speed than traditional manual adjustment methods. It can intervene in time at the initial stage of working condition fluctuation to prevent deviation from expanding and ensure the continuity and stability of the production process. At the same time, the modular design of the system allows the functional modules to be flexibly adjusted and expanded according to actual production needs, which can adapt to PTA production scenes of different scales and different process conditions, and has strong universality and practicality. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 The timing diagram of the PTA mixed feedstock control system described in the present application;

[0047] Figure 2 The flowchart of the trigger condition of the online learning module;

[0048] Figure 3 The flowchart of the fine-tuning process of the online learning module. DETAILED DESCRIPTION

[0049] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts are within the scope of the present application.

[0050] Please refer to Figure 1 The present application provides a PTA mixed feed control method and system, the system includes mechanism modeling module, data-driven prediction module, dynamic fusion module, rolling optimization module and online learning module.

[0051] The mechanism modeling module constructs a mechanism sub-model based on the PTA mixed reaction kinetics equation. The sub-model calculates the theoretical feed constraint value under the set process boundary conditions according to the reaction rate equation and the material balance relationship. The data-driven prediction module processes the temperature, pressure and raw material flow data collected by the real-time sensor through a multivariate time series neural network, extracts dynamic coupling features and outputs real-time feed prediction values. The dynamic fusion module uses an adaptive weight distribution algorithm to calculate the weight coefficients of the mechanism sub-model and the data-driven model according to the real-time working condition error, and performs weighted fusion on the theoretical feed constraint value and the real-time feed prediction value to generate a fused feed prediction value. The rolling optimization module updates the multi-objective optimization function based on the fused feed prediction value within the rolling time window, obtains the current optimal temperature control value and raw material flow set value by solving the function, and generates a feed parameter adjustment instruction. The online learning module collects the adjusted sensor data after executing the adjustment instruction, calculates the actual feed value through the material metering model, and if the absolute value of the relative deviation between the actual feed value and the fused feed prediction value exceeds the preset threshold, triggers the online fine-tuning mechanism, freezes the mechanism sub-model parameters and the data-driven model hidden layer parameters, updates the data-driven model output layer parameters based on the gradient, and continues until the deviation is below the threshold. The system can also optionally include an intelligent detection body networking module, which performs work rationality analysis based on the decision divergence point sequence of multiple detection bodies, selects a reference divergence point, traces the operation trajectory, matches the trajectory features and trigger features, starts the replay verification mechanism and generates a process state evaluation result.

[0052] Embodiment 1: The mechanism modeling module constructs a mechanism submodel based on PTA mixed reaction kinetics equations. The process first defines the reaction rate equation of the PTA slurry preparation process. The reaction rate equation describes the relationship between reaction rate and temperature in the Arrhenius form, and introduces a pressure correction factor to represent the influence of high pressure environment on reaction equilibrium. The setting of process boundary conditions covers three dimensions of temperature, pressure and raw material concentration. The temperature boundary is set to a closed interval of 180-220 degrees Celsius according to the PTA decomposition characteristics, the pressure boundary is limited to the range of 0.8-1.2 megapascals according to the design pressure capacity of the reaction kettle, and the raw material concentration boundary is determined to be 75-85 weight percent through solubility experiment. The solution process of material balance equation establishes double constraints of mass conservation and molar balance. The mass conservation equation tracks the total mass flow rate of PTA solid particles and solvent, and the molar balance equation quantifies the esterification reaction process of terephthalic acid and ethylene glycol, while integrating the generation kinetics model of byproduct water. The solution of differential equations uses the fourth-order Runge-Kutta method to iteratively calculate the theoretical feed constraint value with a fixed step size of 1 minute. The output results include the theoretical optimal feed flow rate value and the PTA slurry target concentration value, with a numerical accuracy controlled within ±0.5% error band.

[0053] The real-time prediction process of the data-driven prediction module starts from the sensor data acquisition layer. The temperature data is obtained by distributed thermocouples, the pressure data is measured by piezoresistive sensors, and the raw material flow data is collected by Coriolis mass flow meters. Three-level filtering operations are implemented in the raw data preprocessing stage: first, sliding mean filtering is performed to eliminate high-frequency noise, then median filtering is performed to eliminate pulse interference, and finally Z-score standardization processing is performed to realize dimension unification. The multivariate time series neural network adopts a bidirectional long short-term memory (BiLSTM) architecture. The network input layer receives 10 groups of historical data slices organized in time series, each slice containing data points of temperature, pressure, and three raw material flow dimensions. The hidden layer is designed as a 64-unit bidirectional LSTM structure. The forward layer and the backward layer capture historical and future time series features, respectively. Dynamic coupling feature extraction is achieved through a gating mechanism. The input gate controls the feature update strength, the forgetting gate adjusts the historical memory retention ratio, and the output gate selects effective features to pass to the next layer. The regression layer adopts a fully connected neural network structure, and the activation function is selected as the identity function to ensure the continuity of the output. The final generated real-time feed prediction value contains two output nodes: the first node corresponds to the PTA solid flow prediction value, and the second node corresponds to the solvent flow prediction value. The numerical dimension is completely consistent with the theoretical constraint value output by the mechanism sub-model. The hardware interface of the data acquisition system uses the industrial Ethernet protocol, and the sampling frequency is fixed at 10 Hz. The data buffer area implements ring queue management to avoid data overflow. The historical working condition data set is used in the neural network training stage, covering normal, transition, and abnormal working conditions. The Adam optimizer is used in the training process, with a batch size of 128 and 200 training rounds. In the feature extraction process, dynamic coupling features enhance the weight allocation of key parameters through attention mechanisms. The attention score is calculated based on the product of parameter fluctuation amplitude and process influence coefficient. After the prediction value is output, post-processing operations are performed, including numerical range limiting and gradient smoothing. The limiting range refers to the equipment safety operation threshold, and the gradient smoothing uses a first-order low-pass filter to suppress the setting value mutation.

[0054] The synchronization of theoretical feed constraints and real-time feed predictions is ensured by a timestamp alignment mechanism. When the sensor data transmission delay exceeds 200 milliseconds, a linear interpolation compensation is activated to compensate for the missing data points. The prediction module operates at a cycle that is synchronized with the scanning period of the control system, both set at 500 milliseconds, ensuring that each control decision is based on the latest prediction results. The effectiveness of real-time feed predictions is verified using residual analysis. If the prediction residuals exceed the 2σ range for three consecutive cycles, the model self-checking procedure is triggered. The self-checking procedure diagnoses the source of data anomalies by comparing the output values of the mechanism model at the same period. The byproduct processing module in the material balance equation is modeled separately. A proportional relationship model is established between the byproduct water generation rate and the main reaction rate, with the proportionality coefficient calibrated through laboratory kinetic experiments to be in the interval of 0.15-0.25. A variable step size adaptive algorithm is used in the numerical solution process. When the concentration gradient change rate exceeds 5% / min, the step size is automatically shortened to 0.2 minutes to ensure the stability of the calculation during the intense reaction stage. The theoretical constraint value output interface uses the OPCUA communication protocol, and the data is packaged in a structured array format and transmitted to the fusion module. The array elements include timestamp, constraint value type identifier, and floating point value. The gradient disappearance problem of the neural network hidden layer is alleviated by layer normalization technology. The normalization layer is inserted after the LSTM unit to process the hidden state with zero mean and unit variance. The regression layer weight initialization uses the Xavier method, and the bias term initial value is set to the mean of historical data. The output layer neuron activation range is limited within the feasible region of the process, and the values outside the range are automatically truncated to the boundary value. The online diagnosis function of the prediction module monitors the feature extraction effect in real time. When the L2 norm of the feature vector continuously decreases to 70% of the threshold, the feature importance reevaluation process is automatically triggered. The reevaluation is achieved by observing the output sensitivity through input parameter perturbation. The dynamic threshold method is used for outlier removal in the preprocessing stage. The threshold is dynamically updated based on the standard deviation of the last 100 sampling points, and the data points exceeding the 3σ range are marked as invalid. The data filling mechanism uses the Kalman filter algorithm based on the state space to predict the reasonable value of the missing data points. The time sequence alignment of input data is achieved through cross-correlation analysis, ensuring that multivariate data enters the neural network processing process under the same time reference. The reaction progress parameter in the theoretical constraint value calculation process is implemented online correction. The correction signal comes from the online component analyzer at the outlet of the reaction kettle. The component analysis data is updated every 5 minutes to correct the coefficient. The boundary condition updating mechanism of the material balance equation is linked with the DCS system. When the DCS switches the production formula, the corresponding boundary parameter set is automatically loaded. The convergence of the numerical solver is monitored by setting double criteria: the upper limit of the iteration number is 500 and the residual norm threshold is 1e-6. Either condition satisfies the calculation is terminated and the output result is obtained.

[0055] The confidence evaluation of real-time prediction value is realized by an ensemble learning method, and three heterogeneous neural network models (LSTM, GRU, and TCN) are run in parallel. The dispersion degree of the prediction result is output as a confidence index. When the confidence is less than 0.8, the weight distribution of the prediction value in the fusion module is automatically reduced, and a prediction reliability alarm is sent to the operator console. The model updating mechanism sets a version management function. When the online learning module triggers parameter updating, the prediction model automatically saves the historical version and records the change log, supporting rollback to the previous stable version if necessary.

[0056] In embodiment 2, the dynamic fusion module receives the theoretical feed constraint value output by the mechanism sub-model and the real-time feed prediction value output by the data-driven prediction module. The adaptive weight allocation algorithm takes the real-time working condition error as an input variable, which is calculated by comparing the mean square error of the current temperature sensor reading and the prediction values of the two models. The weight coefficient update period is set to 500 milliseconds synchronously with the control scanning period. The weight distribution ratio is adjusted using an exponential decay rule. The initial weight coefficient of the mechanism sub-model is set to 0.7, and the initial weight coefficient of the data-driven model is set to 0.3. The decay factor is dynamically adjusted according to the error change rate. When the working condition fluctuation amplitude exceeds 5%, the fast response mode is started, and the decay factor is adjusted from 0.9 to 0.6 to accelerate the weight transfer. The weighted summation operation is executed element by element at the data dimension level, and the weighted value is calculated independently for each data point. The numerical range verification is performed during the generation of the fused feed prediction value, and the boundary truncation protection is automatically triggered when the weighted result exceeds the process feasible region. The truncation threshold is set according to the equipment safety operation manual.

[0057] The rolling optimization module establishes a rolling time window mechanism with a time window length of 5 minutes, which contains 60 groups of continuous sampling points of the fused feed prediction value. A multi-objective optimization function is constructed, which includes three core target items: the first target item quantifies the steam consumption, which is calculated by the product of the heat exchanger power and the flow rate; the second target item evaluates the reaction conversion rate based on the deviation of the real-time concentration prediction value from the theoretical equilibrium concentration; and the third target item monitors the product quality fluctuation by measuring the variance of the outlet material characteristic parameters. The constraint condition system integrates the device physical limits and process safety rules, with the temperature control range limited to 185-215 degrees Celsius, the raw material flow rate change rate not exceeding 3% per minute, and the reaction kettle pressure fluctuation amplitude controlled within ±0.05 MPa. The optimization solving process uses a constrained sequential quadratic programming algorithm with adaptive iteration step adjustment. Convergence is determined when the target function improvement rate is less than 0.1%. Sensitivity analysis is performed during the optimal solution output stage, with the temperature control value precision retained to one decimal place and the raw material flow rate set value rounded to 0.01 tons / hour units.

[0058] The generation of the feedstock parameter adjustment instruction adopts a structured data packaging format, and the instruction package contains four parts: time stamp, target device address code, parameter type identifier and set value. The instruction transmission is implemented through industrial Ethernet, the communication protocol adopts Modbus TCP standard, and the transmission period is aligned with the control system scanning period. The actuator interface layer is configured with an instruction parser to convert digital instructions into 4-20 mA analog signal output. The temperature control signal drives the heater thyristor power regulation module, and the flow setting signal adjusts the pneumatic control valve positioner. The instruction execution feedback mechanism monitors the actuator response state in real time. When the actual change value of the valve opening degree deviates from the instruction by more than 2%, the adjustment instruction is automatically reissued and the execution abnormal event is marked. The error monitoring module of the adaptive weight distribution algorithm sets a double-check mechanism. While comparing the original sensor data with the model prediction value, the moving average of the adjacent three sampling points is additionally introduced as a reference benchmark. The weight coefficient smoothing process adopts first-order inertia filtering, and the time constant is set to 30 seconds, effectively suppressing the weight jump caused by sensor transient disturbance. The weight distribution history record function saves the change curve of the last 24 hours, supporting the operator to trace back and analyze the correlation between weight change and working condition fluctuation. The data management of the rolling time window adopts a ring buffer structure, and the new data covers the oldest data to realize dynamic update. The weight coefficient of the multi-objective optimization function can be configured online, and the default setting is energy consumption weight 0.4, conversion rate weight 0.4, and mass stability weight 0.2, supporting dynamic adjustment of priority according to production needs. The initial point selection strategy of the optimization solver adopts the historical optimal solution extrapolation method, taking the last period optimization result as the starting point to accelerate the convergence process. The solving process is implemented in real time. When the iteration number exceeds 50 times and still does not converge, it is automatically switched to the simplified model for recalculation.

[0059] The security protection mechanism after instruction generation includes an out-of-limit truncation function. When the temperature control value of the optimization output exceeds the safety boundary, the boundary value is automatically replaced by the optimization result and an out-of-limit warning is issued. Gradient limitation is implemented for flow set value changes. When the change amplitude of adjacent periods exceeds 2%, slope adjustment is started to gradually approach the target value at a rate of 0.5% per minute. The integrity check of instruction transmission uses CRC-16 cyclic redundancy check code. When the receiving end fails to pass the check, a data retransmission request is initiated. The maximum number of retransmissions is set to 3. The data synchronization technology in the dynamic fusion process uses hardware timestamp marking. The data streams of each module are kept time-synchronized through IEEE1588 precision clock protocol. When the data arrival time deviation exceeds 200 milliseconds, data reconstruction based on Lagrange interpolation method is started. The quality evaluation module of the fusion result calculates the correlation coefficient of the theoretical constraint value and the real-time predicted value. When the coefficient is lower than 0.6, the weight of the data-driven model is automatically increased by 10%. The constraint processing of the rolling optimization module uses the penalty function method to convert the constraint conditions into penalty items of the objective function. The penalty coefficient is dynamically adjusted according to the constraint violation degree. The optimization algorithm has a multi-thread parallel computing architecture. The target function calculation and constraint verification are distributed to different processor cores for simultaneous execution. Feasibility verification is performed before the output of the optimization result. A simplified mechanism model is used to simulate the execution of the optimization instruction to predict whether the key parameters will exceed the safety threshold. The lag compensation model in the instruction execution stage predicts the response delay of the actuator. A lead correction link is added to the temperature control instruction. The correction amount is set to 15% of the current change rate based on the thermal inertia test data. The flow regulation instruction implements feedforward compensation based on pressure sensor data to predict the change of pipe network resistance and adjust the valve opening instruction value in advance. The execution effect evaluation period is set to 90 seconds after the instruction is issued. The actual working condition data and the expected effect are compared. When the difference exceeds 3%, the optimization parameter re-adjustment mechanism is triggered. The fault protection mode of the adaptive weight distribution algorithm is activated when the sensor fails. It automatically switches to a fixed weight ratio (mechanism model 0.8, data model 0.2) and issues a warning. The rolling optimization module has an offline simulation interface that supports importing historical data sets for optimization strategy pre-validation. All control instruction generation logs contain complete decision chain information, recording time window data, optimization target weight, constraint condition state, and final instruction value. Log files are stored by batch and retained for three months.

[0060] Example 3: see Figure 2The online learning module starts the monitoring process after the execution of the feedstock parameter adjustment instruction. The collection of sensor data after adjustment is achieved through the distributed I / O module. The temperature data is obtained from three groups of thermocouples at different heights of the reactor, the pressure data is obtained from the pressure transmitter at the top of the reactor, and the raw material flow data is obtained through the Coriolis mass flowmeter. The data collection frequency is set to 2 times per second, the collection window duration is 180 seconds after the execution of the adjustment instruction, and the collected data is stored in the temporary buffer area after being filtered by the sliding average filter. The material metering model adopts a calculation method based on the principle of mass conservation. The actual feedstock value is calculated by the following formula:

[0061]

[0062] wherein: represents the actual feedstock value (kg / min), F i represents the volume flow rate of the i-th raw material (m 3 / min), represents the density of the i-th raw material (kg / m 3 ), C i represents the concentration of the i-th raw material (%) and n represents the number of raw material types, represents the metering time interval (min). The density parameter is measured in real time by an online density meter, the concentration parameter is obtained from an online spectral analyzer at the feed inlet of the reactor, and the metering time interval is set to 1 minute to ensure data stability.

[0063] The comparison between the actual feedstock value and the fused feedstock prediction value is calculated by the absolute value of the relative deviation, and the calculation formula is wherein represents the fused feedstock prediction value. The deviation calculation is performed every 60 seconds, and the calculation window contains the moving average of the last 10 sampling points. The preset threshold is set according to the process requirements: the first-level threshold is 5%, the second-level threshold is 8%, and the third-level threshold is 12%. When the absolute value of the relative deviation exceeds the first-level threshold for three consecutive times, the system triggers a yellow warning and starts the deviation reason analysis program; when the deviation value exceeds the second-level threshold, the system triggers an orange warning and reduces the weight of the data-driven model by 10%; when the deviation value reaches the third-level threshold, the system immediately triggers the online fine-tuning mechanism. The deviation reason analysis program first checks the reliability of the sensor data, and judges the data validity by comparing the consistency of adjacent sensor readings. If the maximum difference of the temperature sensor exceeds 2°C or the fluctuation amplitude of the pressure sensor reading exceeds 0.03 MPa, it is determined that the sensor is abnormal and the learning process is suspended. Secondly, the change of the raw material physical property parameters is analyzed, and when the density measurement value deviates from the standard value by more than 5% or the concentration measurement value fluctuates by more than 3%, the material property re-calibration program is started. Finally, the working condition stability is evaluated, and when the temperature variance within the last 30 minutes exceeds 4°C 2or pressure variance exceeds 0.01 MPa 2 When the above conditions are met, the system determines that the operating condition is unstable and delays the learning process.

[0064] The trigger condition of the online fine-tuning mechanism sets multiple verifications, in addition to the relative deviation absolute value exceeding the limit, it also needs to meet the following conditions: no parameter adjustment has been made in the last 15 minutes, the data-driven model running time exceeds 24 hours, and the current operating condition is in the steady-state operating interval (temperature fluctuation less than ±1℃, pressure fluctuation less than ±0.02MPa). Before fine-tuning is started, data quality verification is performed to check the integrity and consistency of sensor data, and if missing data exceeds 5%, the fine-tuning operation is automatically postponed. After verification, the system saves the current data-driven model parameter snapshot and creates a backup file with version number "Model_Vx.y", where x is the major version number and y is the sub-version number. The calculation of the actual feed value implements a real-time verification mechanism, which verifies the data reliability by comparing the results of three independent calculation methods: the first is the direct measurement value based on the mass flow meter, the second is the calculated value based on the pump speed and displacement, and the third is the estimated value based on the valve opening and pressure drop. When the difference between the results of the three methods exceeds 2%, a request for calibration of the metering system is triggered, and the weighted average value is used as the final actual feed value, with the weight distribution being 0.6 for direct measurement, 0.3 for calculation, and 0.1 for estimation. The dynamic adjustment of the metering time interval is automatically optimized according to the flow stability, when the flow fluctuation coefficient exceeds 0.05, the interval is automatically shortened to 30 seconds, and when the fluctuation coefficient is less than 0.01, the interval is lengthened to 2 minutes.

[0065] The calculation of the absolute value of the relative deviation introduces temperature and pressure compensation correction, taking into account the change characteristics of material density with temperature and pressure. The density parameter in the calculation of the actual feed value is real-time corrected by the temperature and pressure compensation formula. The compensation formula coefficients are determined by laboratory calibration experiments, and the compensation parameter table is updated every half year. The deviation evaluation implements multi-time scale analysis. In addition to the immediate deviation calculation, the 60-minute moving average deviation and the 24-hour cumulative deviation are also calculated to evaluate the long-term performance trend of the system. The threshold management adopts an adaptive adjustment strategy. Different threshold levels are set according to the importance of the production process: for key process parameters such as PTA main raw material flow, the threshold is set to 4%; for auxiliary materials such as catalyst flow, the threshold is set to 8%. The threshold value is reviewed and revised every quarter, and the revision is based on historical deviation data analysis, equipment performance change evaluation, and process optimization requirements. The record of deviation overrun events uses a classified storage method, which is classified and marked in three dimensions according to the deviation degree, occurrence frequency, and impact level, supporting subsequent statistical analysis. The system state check before the online fine-tuning trigger includes a hardware self-checking link, which checks indicators such as processor load rate (which needs to be less than 70%), memory available space (which needs to be greater than 1GB), and storage write speed (which needs to be higher than 100MB / s). If the self-checking fails, it automatically enters standby mode and waits for the system resources to recover before restarting the checking process. The data protection during the fine-tuning process uses a double buffer technology. Real-time data is written to buffer A, while buffer B saves the last period data as a recovery benchmark, ensuring that the fine-tuning can quickly return to a stable state in case of interruption.

[0066] The sensor verification in the deviation analysis process implements a cross-validation mechanism. The temperature sensor readings are correlated with the heater power output, the pressure sensor readings are consistent with the inlet valve opening, and the flow sensor readings are matched with the pump speed. If the verification fails, it automatically switches to redundant sensor data, or uses predicted values based on historical data if there is no redundant data. When the material property re-calibration program is started, the latest laboratory analysis results are automatically used as the benchmark, and a new sampling analysis request is triggered to update the calibration parameters. The time alignment of the actual feed value calculation uses high-precision timestamp technology. All sensor data are synchronized through the IEEE1588 precision clock protocol, with a time deviation of within milliseconds. The units used in the calculation process are unified according to the International System of Units. The density unit is kg / m 3 , the flow unit is m 3 / min, and the concentration unit is mass percentage. The data validity check implements range verification. Density values outside the range of 800-1200 kg / m 3 , flow values outside the range of ±20% of the rated flow, and concentration values outside the range of 70-90% are automatically marked as invalid and excluded from the calculation.

[0067] Example 4: See Figure 3 The online fine-tuning process of the data-driven model parameters starts from the parameter freezing operation, the parameter freezing of the mechanistic sub-model is achieved by setting the parameter update flag to false, and the gradient freezing technique is used to force the gradient calculation function return value of the hidden layer weight matrix to zero. The gradient calculation of the output layer parameters is based on the adjusted temperature, pressure and raw material flow data set, which contains 360 groups of samples collected within 180 seconds after the execution of the instruction, each group containing temperature measurement, pressure measurement, three raw material flow values and corresponding actual feed values. The gradient calculation uses the back propagation algorithm, the loss function is defined as the mean square error of the actual feed value and the model prediction value, the initial value of the learning rate is set to 0.001, and the momentum factor is set to 0.9. The parameter update process uses the mini-batch gradient descent method, the batch size is set to 32 samples, and the output layer weight matrix is updated immediately after each batch training is completed. The updated parameters need to be checked for numerical validity, the weight values outside the range of [-5, 5] are automatically truncated to the boundary values, and when the gradient value appears NaN abnormal, it is rolled back to the previous step parameter and the learning rate is reduced by 50%. The iteration stop condition is set with three judgment criteria: the relative deviation absolute value is less than the set threshold (5%), the loss function decrease rate of the last three batches is less than 0.1%, or the maximum iteration number is reached 100 times. Any one of the conditions will terminate the fine-tuning process, and the parameter freezing state will be released and the model will resume normal operation.

[0068] The data management in the fine-tuning process uses a double verification mechanism, the training data set is randomly divided into 80% training set and 20% validation set, and the loss function value on the validation set is calculated after each training period. When the validation set loss increases for three consecutive periods, the early stopping mechanism is triggered to prevent overfitting, and the learning rate is automatically reduced to 70% of the original value. The training progress real-time monitoring interface displays key indicators such as current iteration number, loss function value, relative deviation absolute value, etc., and the operator can pause or adjust the fine-tuning parameters at any time. See Table 1 for training data of a batch in the fine-tuning process.

[0069] Table 1: Output layer parameter fine-tuning training data record

[0070] Sample No. Temperature (°C) Pressure (MPa) PTA flow rate (t / h) Solvent flow rate (t / h) Catalyst flow rate (L / h) Actual feed value (kg / min) Predicted feed value (kg / min) Loss value B20240918-001 195.2 1.05 12.8 8.2 45.3 352.6 368.9 265.7 B20240918-002 195.3 1.06 12.7 8.3 45.1 351.8 367.2 238.2 B20240918-003 195.1 1.04 12.9 8.1 45.5 353.2 370.5 298.4 B20240918-004 195.4 1.05 12.6 8.4 44.9 350.3 365.8 240.3 B20240918-005 195.0 1.05 13.0 8.0 45.6 354.1 371.2 292.4

[0071] The abnormal handling mechanism in the fine-tuning process contains multiple protection layers, when hardware failure causes training interruption, the current training state is automatically saved to the persistent storage, including the current iteration number, parameter snapshot, loss function history and other key information. After training recovery, it can continue from the breakpoint without the need to start over. The software level exception capture includes memory overflow check, zero division error prevention, array out-of-bound protection, etc., any exception will trigger training pause and notify the technical personnel.

[0072] The numerical stability in the gradient calculation process is guaranteed by gradient clipping technology. When the L2 norm of the gradient vector exceeds the set threshold 1.0, the gradient value is automatically scaled to a reasonable range in proportion. The learning rate adjustment adopts the cosine annealing strategy, and the initial learning rate 0.001 gradually decays with the increase of the iteration number, and the minimum learning rate is set to 0.0001. The dynamic adjustment of the momentum factor is based on the consistency of the gradient direction. When the gradient direction is consistent for three consecutive batches, the momentum factor is increased to 0.95, and when the gradient direction oscillates, it is reduced to 0.85. The preprocessing of the training data set includes the outlier filtering link, and the isolated forest algorithm is used to detect abnormal samples. Samples with an abnormal score exceeding 0.6 are automatically excluded from the training set. The data standardization process uses dynamically adjusted mean and variance, which updates the standardization parameters once every batch to ensure the consistency of the training data distribution. The feature engineering section adds interaction features, such as the product of temperature and pressure, and the ratio between flow and other derived features, to enhance the model's expression ability. The verification test after parameter update includes three dimensions: first, test the model performance on the validation set, requiring the absolute value of the relative deviation to be reduced to below 4%; second, perform regression testing on the historical data set to ensure that the prediction accuracy of the model on known data does not decrease; and finally, test the real-time inference speed, requiring the time consumed by a single prediction to be no more than 50 milliseconds. Any test that fails will trigger the parameter rollback mechanism, automatically restoring to the model version before fine-tuning. Resource monitoring during the fine-tuning process tracks CPU utilization, memory occupancy, and disk I speed in real time, and automatically reduces the training batch size or extends the training interval when system resources are tight. The training log records detailed parameter changes for each batch, including weight matrix norm changes, gradient distribution statistics, learning rate adjustment records, and other information. The log file is stored in binary format to reduce storage space occupancy. The model version management adopts an incremental storage strategy, generating a new model version after each successful fine-tuning, and retaining the last 10 versions for quick rollback. The version metadata records training time, data batch, final accuracy indicators, and other key information, supporting retrieval of historical versions by time range or accuracy requirements. The model deployment adopts the blue-green deployment method, and the new version model runs in shadow mode for a period of time before switching to the official model after confirmation of stability. The fine-tuning effect evaluation not only focuses on the absolute value of the relative deviation, but also monitors the stability indicators of the model prediction results, including the variance, range, and coefficient of variation of the predicted values. Long-term performance tracking tracks the average and maximum deviation values every week, and draws trend charts to analyze the performance change law of the model. When the model performance is found to be continuously declining, a comprehensive retraining process is automatically triggered, and all parameters are retrained using a larger-scale data set. The visualization monitoring interface during the training process displays the loss function curve, accuracy change curve, and parameter distribution histogram in real time, and operators can adjust the training parameters or manually terminate the training through interface interaction.The alarm system sets multiple thresholds. When the training time exceeds the expected value, the loss function decreases slowly, or numerical anomalies occur, the relevant technical personnel are notified through sound and light alarms and short messages.

[0073] In embodiment 5, the intelligent detection body networking module constructs a multi-agent collaborative detection architecture. The system deploys five detection bodies with different functions: a process state detection body responsible for monitoring the temperature and pressure parameters of the reaction kettle, a quality index detection body tracking product concentration and viscosity data, an equipment health detection body analyzing the working state of the pump and valve actuators, an energy efficiency detection body statistics steam and power consumption mode, and a safety warning detection body evaluating the risk level of abnormal working conditions. Each detection body independently runs a decision algorithm and outputs an evaluation conclusion. The decision divergence point sequence is generated by comparing the differences in the output results of each detection body. When the evaluation conclusions of two or more detection bodies on the same working condition differ by more than the allowed range, it is recorded as a divergence point. The reference divergence point is selected using a weighted voting mechanism, considering three factors: divergence intensity, duration, and impact range. The weight distribution ratio is 4:3:3. The divergence point with the highest score is selected as the reference point and enters the deep analysis process.

[0074] After the working rationality analysis is started, the historical operation trajectory of the detection body corresponding to the reference divergence point is first traced back. The operation trajectory of the process state detection body includes the temperature control instruction sequence, pressure regulation record, and stirring speed adjustment log, covering the time range from 24 hours before the divergence point to 1 hour after the divergence point. The trajectory data of the quality index detection body includes online analyzer readings, laboratory sampling results, and product quality rating records. The equipment health detection body provides pump and valve maintenance records, vibration spectrum data, and actuator response time statistics. The trajectory feature extraction uses a multi-scale analysis method to mine operation rules from three time dimensions: minute, hour, and day. The minute-level feature focuses on the fluctuation frequency of control instructions, the hour-level feature analyzes the trend slope of parameter changes, and the day-level feature statistics the periodic mode of equipment operation. The preset trigger feature library contains twenty typical abnormal patterns, such as continuous positive and negative alternation of temperature control instructions, sudden acceleration of pressure regulation frequency, and continuous extension of pump and valve response time.

[0075] The replay verification mechanism is implemented on the digital twin platform, building a virtual model that is identical to the physical reaction kettle, including thermodynamic transfer processes, fluid dynamics characteristics, and chemical reaction kinetics equations. The replay input data uses actual operation records two hours before and after the divergence point, with a simulation step size of 0.1 seconds. The output parameters include temperature distribution field, concentration gradient field, and flow velocity field. The verification result is generated by comparing the matching degree of the virtual model output and the actual sensor data. The matching degree calculation uses a multi-index weighted scoring method, with temperature matching weight 0.4, pressure matching weight 0.3, and concentration matching weight 0.3. The process state evaluation result uses a five-level classification system: level one represents the optimal system operation, level two represents slight deviation that needs attention, level three suggests that the operation parameters need to be adjusted, level four warns of potential fault risk, and level five alarms for immediate shutdown inspection. The decision divergence point sequence management uses a time window sliding mechanism, detecting a new divergence point every five minutes and updating the sequence queue, with the sequence length maintaining the last 50 divergence point records. The confirmation of the benchmark divergence point requires independent verification by at least three detection bodies, including the timeliness (generation time not more than 10 minutes), severity (impact score more than 70 points), and reproducibility (reproducible in the test environment) of the divergence point. The traceability depth of the operation trajectory is dynamically adjusted according to the divergence level, with the last 4 hours of data for level one divergence, 8 hours for level two divergence, and 24 hours of complete data for level three and above divergence. The trajectory feature analysis uses a machine learning classification algorithm, with a training data set containing 1000 groups of historical normal operation trajectories and abnormal operation trajectories, and feature dimension expanded to 50 including statistical features, frequency domain features, and time series features. The preset trigger feature implements an online update mechanism, extracting at least three new features from newly occurring abnormal cases each month to add to the feature library, while eliminating old features with a usage frequency of less than 10 times. The feature matching process uses a similarity calculation algorithm, and when the cosine similarity of the real-time trajectory features and the preset features exceeds 0.85, it is determined as a successful match, and the replay verification process is immediately started. The model accuracy of the replay verification platform is calibrated regularly, with the latest laboratory data used to update the reaction kinetics parameters every week, and the mechanical performance parameters adjusted according to the equipment maintenance records every month. The simulation result verification sets double standards: the primary standard is that the average absolute error of key parameters does not exceed 2%, and the secondary standard is that the correlation coefficient of parameter change trend reaches more than 0.9. When model deviation is found during the verification process, the parameter correction program is automatically started, and the model parameters are optimized by least squares method until the accuracy requirements are met.

[0076] The output of the process state evaluation result adopts a structured report format, including four parts of evaluation level, main basis, influencing factors and improvement suggestions. The evaluation result is pushed to the operator console in real time, and stored in the historical database for trend analysis. Different evaluation levels are set with corresponding treatment processes: first and second evaluation results only need to be recorded and filed, third evaluation triggers operation parameter optimization suggestions, fourth evaluation starts preventive maintenance inspection, and fifth evaluation immediately executes emergency shutdown program. Multi-agent collaborative detection adopts a distributed computing architecture, with each detection body deployed on an independent computing node and data exchange through high-speed industrial Ethernet. The communication protocol between detection bodies adopts a custom binary format, with a fixed data transmission period of 100 milliseconds and a communication delay controlled within 5 milliseconds. System reliability is guaranteed by a heartbeat detection mechanism, which detects the running state of each detection body once a second and automatically switches to a backup node when an anomaly is found. The working reason analysis process implements continuous optimization, with an effect evaluation of the analysis algorithm performed once a month to test the analysis accuracy using historical divergence point data. The evaluation indicators include divergence point recognition accuracy, benchmark point selection rationality and evaluation result effectiveness, and any indicator below 90% will trigger the algorithm optimization process. Optimization methods include adjusting feature weights, updating classification model parameters and expanding training data sets, etc., to ensure continuous improvement of analysis capabilities. The decision logic of the intelligent detection body implements version control, with the old version decision logic retained as a comparison benchmark after each algorithm update. The new version logic needs to run in the test environment for more than 72 hours to confirm stability before being put into formal use. The key parameter settings in the decision-making process are managed by hierarchical management, with ordinary operators only able to view parameter settings, modification authority requiring engineer-level account authorization, and important parameter changes requiring approval by the technical director.

[0077] While embodiments of the present application have been shown and described, it is to be understood that the embodiments described are merely divergences of the principles and spirit of the present application, and that many changes, modifications, substitutions and variations can be made thereto without departing from the spirit and scope of the present application, which are defined by the following claims and their equivalents.

Claims

1. A control system for a PTA blend feed, characterized by, The method comprises the following steps: a mechanism modeling module is used to construct a mechanism sub-model based on a PTA mixed reaction kinetics equation, and output a theoretical feed constraint value; a data-driven prediction module is used to receive temperature, pressure and raw material flow data collected by a real-time sensor through a data-driven model, and output a real-time feed prediction value; a dynamic fusion module is used to weight and fuse the theoretical feed constraint value output by the mechanism sub-model and the real-time feed prediction value output by the data-driven model, and generate a fused feed prediction value; a rolling optimization module is used to generate a feed parameter adjustment instruction through a rolling optimization algorithm according to the fused feed prediction value; an online learning module is used to perform online fine-tuning on parameters of the data-driven model when a deviation between adjusted sensor data collected after the feed parameter adjustment instruction is executed and the fused feed prediction value exceeds a preset threshold.

2. The control system for a PTA hybrid feedstock of claim 1, wherein, The mechanism modeling module constructs a mechanism sub-model based on a PTA mixed reaction kinetics equation, and outputs a theoretical feed constraint value, which comprises the following steps: According to the reaction rate equation of the PTA slurry preparation process, the process boundary conditions of temperature, pressure and raw material concentration are set; Based on the process boundary conditions, the theoretical feed constraint value is calculated through the material balance equation.

3. The control system for a PTA hybrid feedstock of claim 1, wherein, The data-driven prediction module receives temperature, pressure and raw material flow data collected by a real-time sensor through a data-driven model, and outputs a real-time feed prediction value, which comprises the following steps: The temperature, pressure and raw material flow data are input into a multivariate time series neural network to extract dynamic coupling features; Based on the dynamic coupling features, a real-time feed prediction value with the same dimension as the theoretical feed constraint value is output through a regression layer.

4. The control system for a PTA hybrid feedstock of claim 1, wherein, The dynamic fusion module weights and fuses the theoretical feed constraint value and the real-time feed prediction value to generate a fused feed prediction value, which comprises the following steps: According to the real-time working condition error, the weight coefficient of the mechanism sub-model and the weight coefficient of the data-driven model are calculated through an adaptive weight distribution algorithm; Based on the weight coefficients, the theoretical feed constraint value and the real-time feed prediction value are weighted and summed to generate the fused feed prediction value.

5. The control system for a PTA hybrid feedstock of claim 1, wherein, The rolling optimization module generates a feed parameter adjustment instruction through a rolling optimization algorithm according to the fused feed prediction value, which comprises the following steps: The multi-objective optimization function is updated based on the fused feed prediction value within the rolling time window; The current optimal temperature control value and the raw material flow set value are obtained by solving the multi-objective optimization function to generate the feed parameter adjustment instruction.

6. The control system for a PTA hybrid feedstock of claim 1, wherein, When the deviation between the adjusted sensor data collected after the feed parameter adjustment instruction is executed and the fused feed prediction value exceeds a preset threshold, the online learning module comprises the following steps: According to the adjusted temperature, pressure and raw material flow data, the actual feed value is calculated through a material metering model; The relative deviation absolute value of the actual feed value and the fused feed prediction value is calculated, and online fine-tuning is triggered when the absolute value exceeds the set threshold.

7. A control system for a PTA hybrid feedstock according to claim 6, wherein, The online learning module performs online fine-tuning on the parameters of the data-driven model, which comprises the following steps: The parameters of the mechanism sub-model and the hidden layer parameters of the data-driven model are frozen; Based on the adjusted temperature, pressure and raw material flow data, the gradient of the data-driven model output layer parameters is calculated; According to the gradient, the output layer parameters are updated until the absolute value of the relative deviation between the actual feed value and the fused feed prediction value is lower than the set threshold.

8. The control system for a PTA hybrid feedstock of claim 1, wherein, Further comprising: An intelligent detector networking module for multi-agent collaborative detection and analysis of the PTA mixed feed process; The intelligent detector networking module performs working rationality analysis based on the decision divergence point sequence of multiple detectors to generate a process state evaluation result.

9. A control system for a PTA hybrid feedstock according to claim 8, wherein, The working rationality analysis based on the decision divergence point sequence of multiple detectors includes: Selecting a reference divergence point from the decision divergence point sequence; Tracing the operation trajectory of the detector corresponding to the reference divergence point in the historical period; Traversing the trajectory features in the operation trajectory, and when the traversed trajectory feature matches the preset trigger feature, starting the replay verification mechanism; Generating the process state evaluation result based on the replay verification result.

10. A method of controlling a PTA blend feed, characterized by, All modules and method processes of the PTA mixed feed control system according to any one of claims 1 to 9.

Images