Pyrolysis reactor carbon deposition multi-source data fusion early warning and decoking control method and system

Abstract

The application provides a pyrolysis reactor carbon deposition multi-source data fusion early warning and decoking control method and system, belonging to the technical field of mixed plastic pyrolysis, and used for solving the problems that in the related technology, the raw material component fluctuation leads to difficult prediction of carbon deposition and difficult adaptation of decoking strategy. Through the closed loop architecture of "working condition prediction-weight adaptation-data purification-model evolution-decision feedback", combined with LSTM, dynamic weight adjustment, multi-physical field coupling and other technologies, the method accurately outputs the carbon deposition state prediction result and the stepwise control strategy, realizes accurate early warning and efficient control of carbon deposition, improves the accuracy of carbon deposition judgment and the adaptability of the system working condition, and balances the economy and reliability of carbon deposition control.

Description

Technical Field

[0001] This application relates to the field of pyrolysis reactor operation control, and in particular to a method and system for early warning and decoking control of carbon buildup in pyrolysis reactors using multi-source data fusion. Background Technology

[0002] Pyrolysis of mixed plastics is an important technological pathway for resource recycling. The pyrolysis reactor, as the core equipment, is particularly vulnerable to carbon buildup, which directly impacts operational efficiency and product quality. With increasing environmental demands, the proportion of difficult-to-process components such as PVC in mixed plastics is rising, further exacerbating the difficulty of controlling carbon buildup.

[0003] Existing carbon buildup control technologies mostly employ single-parameter monitoring and fixed-strategy decoking, such as triggering decoking actions based on temperature and pressure thresholds, with some technologies incorporating basic data acquisition and simple algorithm prediction. These technologies can be initially applied under stable raw material composition conditions, but lack a deep integration with the pyrolysis reaction mechanism.

[0004] Existing technologies have significant drawbacks: they cannot adapt to fluctuations in mixed plastic components, and predictions are prone to failure when the PVC ratio changes; they do not consider the dynamic evolution of the physical properties of carbon deposits, leading to long-term degradation of model accuracy; and the decoking strategy neglects multi-objective balance, which can easily result in energy waste or product loss. Summary of the Invention

[0005] This application provides a method and system for early warning and decoking control of carbon buildup in pyrolysis reactors using multi-source data fusion, which can adapt to fluctuating conditions in mixed plastic pyrolysis, improve the accuracy of carbon buildup control, and achieve efficient early warning and economical decoking.

[0006] Firstly, this application provides a method for early warning and decoking control of multi-source data fusion for carbon deposition in a pyrolysis reactor. The method involves real-time acquisition of data on temperature, HCl concentration, carbon deposition thickness, carbon deposition density, and raw material composition within the pyrolysis reactor; based on the temperature, HCl concentration, and raw material composition data, identifying the current pyrolysis reaction stage, which includes at least a dechlorination stage and a main pyrolysis stage; according to the identified pyrolysis reaction stage, invoking a pre-set feature weight allocation strategy corresponding to that stage to perform weighted fusion of the acquired multi-source data to obtain weighted feature data; inputting the weighted feature data into a digital twin model to obtain a predicted carbon deposition state; the digital twin model integrates the carbon deposition growth kinetic equation and the reactor heat transfer equation; generating an early warning signal or a decoking control command based on the predicted carbon deposition state; and feeding back and correcting the parameters of the feature weight allocation strategy or the digital twin model based on the execution result of the decoking control command.

[0007] By adopting the above technical solutions, the data processing strategy is dynamically matched with the pyrolysis reaction process through stage identification and stage-based weight allocation; the physical reliability of the model is improved by using a digital twin model that integrates the mechanism equations for state prediction; and the control results are fed back to the front-end model to form an adaptive optimization closed loop, thereby effectively solving the problems of difficult prediction of carbon deposition state and difficulty in adapting control strategies under the fluctuation of mixed plastic raw material composition, and realizing accurate early warning and adaptive control of carbon deposition.

[0008] Further, the identification of the current pyrolysis reaction stage includes: using a Long Short-Term Memory (LSTM) network as a time-series processing unit, inputting a time-series feature vector containing temperature, HCl concentration, and raw material components; embedding a pyrolysis reaction stage identifier into the input layer of the LSTM in a unique thermal encoding form, the stage identifier being determined based on real-time temperature and HCl concentration thresholds; calculating the PVC dechlorination mechanism activating factor and the hydrogen free radical mechanism activating factor, the PVC dechlorination mechanism activating factor being calculated based on the dechlorination reaction kinetic equation and real-time HCl concentration; introducing a data quality score, the data quality score being calculated based on data integrity, noise level, and consistency with dechlorination reaction kinetics; dynamically adjusting the gating attention weight matrix according to the stage identifier to strengthen the importance of the raw material component variation coefficient and the features corresponding to high activation factors, the raw material component variation coefficient being pre-calculated based on the raw material component data; and outputting the extreme operating condition type, the predicted occurrence time, and the predicted confidence level that integrates the operating condition probability and the data quality score.

[0009] By adopting the above technical solution and integrating time series models, mechanistic activation factors and staged attention mechanisms, it is possible to accurately distinguish different pyrolysis stages and quantify the risks and confidence levels of extreme working conditions, providing accurate scenario judgment basis for subsequent processing.

[0010] Furthermore, the determination of the feature weight allocation strategy includes: constructing a four-dimensional weight evaluation matrix, wherein the four dimensions include: pyrolysis mechanism correlation, fingerprint spectrum similarity, sensor bidirectional calibration deviation, and carbon deposition physical property adaptability; calculating a dynamic correction factor based on real-time carbon deposition density and thermal conductivity to correct the basic weights; setting constraints to ensure that the total weight of features associated with the core pyrolysis mechanism is not lower than a preset threshold; and selecting different nonlinear correction functions to adjust the corrected weights in stages according to the current pyrolysis reaction stage, wherein the correction function used in the dechlorination stage strengthens features with high mechanism correlation, and the correction function used in the main pyrolysis stage strengthens features with high correlation to carbon deposition characteristics.

[0011] By adopting the above technical solutions, through multi-dimensional evaluation and dynamic correction, and by forcibly ensuring the dominant position of core mechanism data, the weight allocation can be intelligently adapted to the carbon deposition state and reaction process, thereby improving the correlation between input data and current control objectives.

[0012] Furthermore, before inputting the weighted feature data into the digital twin model, a data purification step is included: calculating the sensitivity derivatives of the carbon deposition growth rate to temperature, HCl concentration, and hydrogen radical concentration; incorporating the sensitivity derivatives into the calculation of feature information gain to construct a pyrolysis kinetics weighted model and enhance the input features; establishing a multiphysics residual formula coupling the heat transfer coefficient, carbon deposition thickness, and dechlorination reaction rate to calculate the data residuals; selecting the corresponding residual threshold according to the current pyrolysis reaction stage; if the data residuals exceed the threshold or simultaneously meet the judgment conditions for incomplete dechlorination, the data is marked as abnormal and removed and supplemented.

[0013] By adopting the above technical solutions, through mechanism-driven sensitivity analysis and multi-physics field coupling verification, abnormal data can be accurately identified and eliminated from the level of physical laws, especially pyrolysis-specific abnormal types (such as incomplete dechlorination), providing highly reliable input for the model.

[0014] Furthermore, after data acquisition and before weight allocation, a two-way sensor calibration step is included: for the HCl concentration sensor, a corrosion drift failure factor driven by cumulative HCl exposure and temperature is calculated; for the carbon deposit thickness sensor, a thickness measurement error factor driven by carbon deposit density, adhesion coefficient, and time is calculated; based on the failure factors, the current pyrolysis reaction stage, and the physical characteristics of the carbon deposit, adaptive sensor data correction coefficients are calculated; using the calibrated data, combined with the pyrolysis mechanism correlation weights, the heat transfer, reaction, and mass transfer coupling coefficients in the digital twin model are back-optimized using the gradient descent method.

[0015] By adopting the above technical solution, sensor data is calibrated using a failure model designed for harsh pyrolysis environments, and the calibration data is used to optimize the parameters of the mechanism model in reverse, thereby achieving a two-way accuracy improvement between the sensor and the digital twin model.

[0016] Furthermore, the method includes a mechanism consistency verification step: constructing a pyrolysis fingerprint library containing different mixing ratios of PVC, PE, and PP, wherein the spectral features in the pyrolysis fingerprint library include standardized temperature, gas concentration, carbon deposition characteristics, and encoded reaction stages; assigning different attention weights to temperature, gas concentration, carbon deposition characteristics, and reaction stage characteristics according to the current pyrolysis reaction stage; using a dynamic time warping algorithm that integrates the attention weights and mechanism deviation constraints to calculate the similarity between real-time data and the benchmark fingerprint library; and updating the fingerprint library under mechanism constraints when the incremental data reaches a scale threshold and its distribution differs significantly from the benchmark fingerprint.

[0017] By adopting the above technical solution, and through multi-ratio fingerprint spectrum library and staged weighted similarity matching, a mechanism consistency criterion is provided for real-time data. In addition, incremental data is combined to realize spectrum self-evolution, enabling the system to adapt to changes in raw material ratio.

[0018] Furthermore, the parameter self-evolution of the digital twin model includes: embedding the Arrhenius equation of carbon deposition growth as a mechanistic constraint term into the reinforcement learning reward function; embedding the carbon deposition density not exceeding the physical upper limit as a penalty term into the reward function; triggering the model evolution process when the coefficient of variation of the raw material components exceeds a first threshold, or the coefficient of change of the carbon deposition growth rate exceeds a second threshold; updating the carbon deposition growth kinetic parameters and multiphysics coupling coefficients in the digital twin model with the goal of maximizing the cumulative reward through a gradient descent method driven by the prediction error and multiphysics residuals; and dynamically adjusting the allowable boundary range of the kinetic parameters according to the degree of fluctuation of the raw material components.

[0019] By adopting the above technical solutions, using physical laws as the evolutionary guide, and triggering adaptive updates when operating conditions fluctuate, the digital twin model can continuously fit the actual pyrolysis process, maintaining long-term prediction accuracy and rapid response capability to operating condition fluctuations.

[0020] Furthermore, obtaining the predicted value of carbon deposit state includes: dynamically selecting convolutional kernels of different sizes for multi-scale feature extraction based on the current pyrolysis reaction stage, wherein small-sized convolutional kernels are used in the dechlorination stage, medium-sized convolutional kernels are used in the main pyrolysis stage, and large-sized convolutional kernels are used in the high-temperature pyrolysis stage; performing stage-adaptive weighted fusion on the extracted multi-scale features; for short-term prediction, inputting the fused features into a recurrent neural network for time-series prediction; for long-term prediction, correcting the prediction by combining the baseline prediction value output by the digital twin model and the operating condition prediction result; and assigning different weights to the prediction loss functions of carbon deposit thickness, density, and growth rate based on the current pyrolysis reaction stage.

[0021] By adopting the above technical solution and through a staged multi-scale feature extraction and prediction architecture, both short-term fluctuations and long-term trends of data are taken into account, and high-precision prediction of multiple targets such as carbon deposition thickness, density, and growth rate is achieved, meeting the control requirements of different time scales.

[0022] Furthermore, the generation of early warning signals or decoking control commands includes: constructing a coupled early warning function, whose inputs include six standardized indicators: carbon deposit thickness, growth rate, fingerprint spectrum similarity, calibration deviation, carbon deposit characteristics, and data quality, and introducing constraint penalty terms for reactor coating tolerance temperature and product yield loss; classifying five early warning levels according to the value of the coupled early warning function; generating a stepped decoking control strategy for different early warning levels, ranging from enhanced monitoring, operating condition adjustment, light decoking to deep decoking, with the strategy generation based on multi-objective optimization of decoking effect, energy consumption, and product yield; comparing the measured value of carbon deposit state after strategy execution with the predicted value, and feeding the obtained correction coefficient back into the model of the pyrolysis reaction stage identification step.

[0023] By adopting the above technical solutions, and through multi-dimensional comprehensive early warning and tiered strategy generation, the system balances equipment safety, desiccation effect and economic cost, and continuously optimizes the accuracy of the system's front-end prediction model through closed-loop feedback.

[0024] Secondly, this application provides a multi-source data fusion early warning and decoking control system for carbon buildup in pyrolysis reactors. It includes: an HCl-resistant data acquisition sensor group for acquiring data on temperature, HCl concentration, carbon buildup thickness, and density within the reactor; an edge computing device communicatively connected to the data acquisition sensor group, the edge computing device being equipped with an FPGA acceleration chip and a high-temperature heat dissipation unit, and configured to execute the method described in any of the first aspects above; a cloud server communicatively connected to the edge computing device for storing and updating a pyrolysis fingerprint library and digital twin model parameters; and a decoking actuator communicatively connected to the edge computing device for receiving and executing decoking control commands.

[0025] By adopting the above technical solutions, through corrosion-resistant sensors, edge computing devices with dedicated acceleration and heat dissipation capabilities, cloud collaboration and execution mechanisms, reliable hardware support and execution paths are provided for the core algorithm, ensuring the efficient and stable operation and implementation of the method in harsh industrial environments.

[0026] In summary, this application has at least the following beneficial effects: A carbon buildup control solution adapted to the fluctuating conditions of mixed plastic pyrolysis is provided, which realizes accurate early warning and adaptive decoking, and improves the stability of reactor operation; By using mechanistic data-driven weight allocation and a model self-evolution mechanism, the long-term accuracy and adaptability of the system under various operating conditions are ensured. By employing a tiered multi-objective coking control strategy and closed-loop feedback, coking energy consumption and product loss were significantly reduced, achieving economical operation.

[0027] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0028] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein: Figure 1 The diagram illustrates the principle of a multi-source data fusion early warning and decoking control system for carbon buildup in a pyrolysis reactor, as described in an embodiment of this application.

[0029] Figure 2 A flowchart of a multi-source data fusion early warning and decoking control method for carbon buildup in a pyrolysis reactor, as described in an embodiment of this application, is shown. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0031] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0032] In a first aspect, embodiments of this application disclose a multi-source data fusion early warning and coke removal control system for carbon buildup in pyrolysis reactors.

[0033] Figure 1 The diagram illustrates the principle of a multi-source data fusion early warning and decoking control system for carbon buildup in a pyrolysis reactor, as described in an embodiment of this application.

[0034] Reference Figure 1 The system includes an HCl corrosion-resistant data acquisition module, an edge computing module, a cloud collaboration module, and an execution module, with each module communicating with each other in sequence.

[0035] Among them, the HCl corrosion resistant data acquisition module is a module used to collect multi-dimensional pyrolysis-specific data. It is equipped with a corrosion resistant carbon deposition physical property monitoring sensor and a high-temperature reaction stage identification sensor to collect multi-dimensional pyrolysis-specific data such as carbon deposition physical properties and reaction stage identification, and transmits the collected data to the edge computing module. The edge computing module is connected to the HCl-resistant data acquisition module for acquiring data output from the HCl-resistant data acquisition module. It incorporates the pyrolysis reactor carbon deposition multi-source data fusion early warning and coking control method disclosed in the second aspect of this application, and is equipped with an FPGA acceleration array and a high-temperature heat dissipation unit. It analyzes and processes the received data and generates a step-by-step coking control strategy, which is then transmitted to the execution module. At the same time, it interacts with the cloud collaboration module. The cloud collaboration module communicates with the edge computing module to enable the cloud collaboration module to obtain relevant data output by the edge computing module for model training and fingerprint map updates, and transmits the updated model and map to the edge computing module. The execution module communicates with the edge computing module to enable the execution module to obtain the stepped focus control strategy output by the edge computing module and to execute the stepped focus control strategy.

[0036] Secondly, embodiments of this application disclose a method for early warning and decoking control based on multi-source data fusion of carbon deposits in a pyrolysis reactor.

[0037] Figure 2 A flowchart of a multi-source data fusion early warning and decoking control method for carbon buildup in a pyrolysis reactor, as described in an embodiment of this application, is shown.

[0038] Reference Figure 2 The method specifically includes the following steps: S1: To address the fluctuating composition characteristics of pyrolysis raw materials for mixed plastics, a five-stage closed-loop algorithm architecture is constructed, consisting of "operating condition prediction, weight adaptation, data purification, model evolution, and decision feedback".

[0039] In this five-stage closed-loop algorithm architecture, 'condition prediction' corresponds to 'identifying the current pyrolysis reaction stage'; 'weight adaptation' corresponds to 'calling the feature weight allocation strategy to perform weighted fusion of the collected multi-data'; 'data purification' is a key processing step before inputting the weighted feature data into the digital twin model; the core of 'model evolution' is reflected in 'feedback correction of the parameters of the digital twin model'; and 'decision feedback' covers the complete process of 'generating early warning signals or decoking control commands' and providing feedback based on the execution results. Subsequent steps S2 to S9 will elaborate on each stage of this architecture.

[0040] The core logic of this architecture is a dynamic loop of "data input - mechanism processing - status output - control feedback." Each module interacts in real time through data interfaces. The output of the operating condition prediction module serves as a pre-trigger signal for the weight adaptation module, and the execution result of the decision feedback module corrects the parameters of the operating condition prediction module, forming a closed-loop iteration. Fluctuations in the composition of mixed plastic raw materials are controlled through the coefficient of variation of the raw material composition. Quantification, its calculation formula is as follows In the formula The standard deviation of the component content of PVC, PE, and PP over a certain period of time. For the average content of the corresponding components, when A fast response mechanism for the time-triggered architecture.

[0041] The pyrolysis-specific mechanisms, such as the temperature-HCl concentration correlation mechanism in the PVC dechlorination stage, are deeply coupled to each module through the embedding of mechanism constraint terms into the algorithm formula and the allocation of staged feature weights. Driven by multi-dimensional pyrolysis-specific data such as carbon deposition physical characteristics and reaction stage identifiers, and combined with digital twin virtual mapping, the entire carbon deposition state self-evolution is realized.

[0042] The temperature-HCl concentration correlation mechanism in the PVC dechlorination stage is quantified using the dechlorination reaction kinetic equation, with the core formula being: In the formula HCl formation rate (unit: mol / (L·s)) The dechlorination reaction rate constant (value range 1.2 × 10⁻⁶) -3 ~3.5×10 -3 (calibrated by experiment) This represents the molar concentration of PVC in the mixed raw materials (unit: mol / L). The activation energy for the dechlorination reaction is denoted as R (valued at 240–260 kJ / mol, based on PVC pyrolysis experimental data fitting), R is the gas constant (8.314 J / (mol·K)), and T is the reactor temperature (unit: K). This equation is embedded in various modules as a mechanistic constraint term. For example, in the data purification module, it is used to determine the matching between HCl concentration and temperature. When the relative error between the measured HCl concentration and the calculated value exceeds 15%, it is marked as abnormal data.

[0043] Stage-based feature weight allocation is achieved through stage identifier factors. In practice, S(t) is dynamically determined based on the pyrolysis process: during the dechlorination stage. Main pyrolysis stage High-temperature pyrolysis stage Different stages correspond to different feature weight vectors. ,in Let be the weight of the i-th feature, and satisfy . Multi-dimensional pyrolysis-specific data includes carbon deposition physical properties data (density). Thermal conductivity ), reaction stage identification data HCl concentration data and hydrogen radical concentration data These data are collected in real time by sensors, and the sampling frequency is adjusted according to the stage. The sampling frequency is 10Hz in the dechlorination stage and increased to 20Hz in the main pyrolysis stage to ensure the timeliness of the data.

[0044] Digital twin virtual mapping, based on a three-dimensional model of a pyrolysis reactor, constructs a virtual model of the evolution of coke deposition states. The core of this model is the fusion of a coke growth rate model and a multiphysics coupling model. The coke growth rate model is... In the formula The carbon deposition growth rate is expressed in mm / h. The carbon deposition growth rate constant (value 5.2 × 10⁻⁶) -4 ~8.7×10 -4 (Determined from carbon deposition experimental data) The activation energy for carbon deposition growth (values ​​range from 180 to 200 kJ / mol). The maximum density of carbon deposits is 1.7 g / cm³. 3 (Based on experiments on the characteristics of carbon deposition from pyrolysis). The multiphysics coupling model connects the heat transfer field, flow field, and reaction field, and solves for the temperature field distribution T(x,y,z,t) and flow velocity distribution using the finite element method. This provides a virtual environment to support the evolution of carbon deposit states, enabling real-time synchronization between physical entities and virtual models.

[0045] The specific methods in this step include: the five-stage closed-loop algorithm architecture is constructed with each module working in sequence; the pyrolysis-specific mechanism is integrated into each module through two methods: embedding mechanism constraint terms and staged weight allocation; multi-dimensional pyrolysis-specific data provides input for each module; digital twin virtual mapping provides a virtual carrier for the evolution of carbon deposition state; and finally, the entire chain of self-evolution from data input to control output is realized.

[0046] The collaborative relationships between the modules are defined through a data flow matrix. Let the set of modules be... (Working condition prediction) (Weight adaptation) (Data purification) (Model evolution) (Decision Feedback) Data flow matrix middle Representation module To module Output data, This indicates no data interaction, specifically in the following form: Clearly define the input-output relationships of each module. The core of end-to-end self-evolution is the dynamic updating of model parameters, achieved through iterative optimization of the objective function. To achieve, in the formula To account for the error in predicting carbon deposit thickness, For multi-physics residuals, and The weighting coefficients (0.6 and 0.4 respectively, determined through grid search) are used to update the model parameters after each closed-loop iteration, making... Gradually reduce the size to ensure continuous optimization of algorithm performance.

[0047] S2: The working condition prediction adopts a stage perception-gated attention fusion algorithm, which is based on a time-series prediction model. It introduces pyrolysis reaction stage identifiers such as dechlorination and main pyrolysis, as well as mechanism activation factors such as PVC dechlorination and hydrogen free radicals and data quality scores. It dynamically adjusts the gated attention weight according to the pyrolysis stage, strengthens key features such as the coefficient of variation of raw material components, and outputs the extreme working condition type, occurrence time and prediction confidence.

[0048] The time-series prediction model uses a Long Short-Term Memory (LSTM) network to construct the basic time-series processing unit. Its core is to control the forgetting and updating of information through a gating mechanism. The cell state update formula for LSTM is: , In the formula, and These represent the cell states at the current and previous moments, respectively. Output for the current moment. Represents element-wise multiplication; forget gate This is used to control the proportion of cells retained from the previous moment. It is the sigmoid activation function. The forget gate weight matrix (dimensions [128, 256], obtained through training optimization after random initialization) is shown. This is the concatenated vector of the previous output and the current input. Bias for the forget gate (initial value set to 0.1); Input gate It is used to control the proportion of the current input information received. (Dimensions [128, 256]) and (Initial value 0.1) represent the weight matrix and bias of the input gate, respectively; candidate cell state. , (Dimensions [128, 256]) and (Initial value 0) is its weight matrix and bias; output gate , (Dimensions [128, 256]) and (Initial value 0.1) represents the weight matrix and bias of the output gate. The input feature vector at time t contains 16-dimensional features such as raw material composition, temperature, and HCl concentration.

[0049] The pyrolysis reaction stage is identified through stage characteristic functions. Quantification is performed using a function derived from the reactor temperature. and HCl concentration Joint decision: When At ppm, it is determined to be in the dechlorination stage. ;when At ppm, it is determined to be in the main pyrolysis stage. ;when At that time, it was determined to be in the high-temperature pyrolysis stage. . As a unique thermal encoding vector ([1,0,0] for the dechlorination stage, [0,1,0] for the main pyrolysis stage, and [0,0,1] for the high-temperature pyrolysis stage), it is embedded in the LSTM input layer, enabling the model to distinguish the reaction characteristics of different stages.

[0050] Mechanistic activators include PVC dechlorination activators. and hydrogen radical activators These are used to quantify the activity levels of the PVC dechlorination reaction and the hydrogen free radical carbon deposition inhibition reaction, respectively. The calculation is based on the dechlorination reaction kinetics, and the formula is: In the formula The maximum observed HCl concentration during the dechlorination stage is 500 ppm, obtained from historical experimental data. (250 kJ / mol) is the activation energy of the dechlorination reaction, R is the gas constant (8.314 J / (mol·K)), and T(t) is the current temperature; when At that time, it was determined that the PVC dechlorination reaction was violent, triggering feature weight enhancement. In the formula This represents the current concentration of hydrogen free radicals. The reference value for hydrogen radical concentration is 10 ppm, determined based on experiments on the critical concentration for inhibiting carbon deposition. When the hydrogen radical carbon deposition suppression ability is insufficient, the weight of related features is increased.

[0051] The data quality score Q(t) is used to evaluate the reliability of input data, comprehensively considering three dimensions: data integrity, noise level, and mechanistic consistency. The calculation formula is as follows: . For completeness scoring, when all 16-dimensional features are complete at time t. For each missing 1-dimensional feature, 0.2 is deducted, with a minimum of 0; , Let be the standard deviation of the eigenvectors at time t. The mean of the feature vector is used to quantify the noise level of the data; the closer the value is to 1, the lower the noise. To score the consistency of the mechanism, by judging and Does it conform to the kinetic relationship of dechlorination reaction (i.e.) Is it in Within ±20% range), if it meets the requirements... Otherwise, it decreases linearly to 0 according to the degree of deviation; The data stream is judged to be high-quality data and is directly input into the model. If Q(t) < 0.6, it needs to be smoothed before input.

[0052] The dynamic adjustment of the gating attention weights is achieved through a stage-adaptive gating unit, the core of which is to determine the attention weight matrix based on S(t). The formula is . The basic attention weight matrix (dimensions [16, 128], obtained from training) This is the stage correction matrix, for the dechlorination stage. (The third dimension corresponds to the HCl concentration characteristic, and is assigned the maximum correction amount), main pyrolysis stage (The 4th dimension corresponds to the carbon deposition thickness characteristic, with the largest correction amount), high-temperature pyrolysis stage (The last dimension corresponds to the heat transfer fluid pressure characteristic, with the largest correction amount); Attention output The softmax function is used to ensure that the sum of the weights is 1, thereby strengthening the key features.

[0053] Extreme condition identification and output are based on model output vectors Classification and regression, , (Dimensions [4, 128]) and (Initial value 0) represents the output layer parameters. The first three elements, after being activated by softmax, correspond to the probabilities of three extreme operating conditions (high PVC ratio, sudden change in heat carrier, and raw material fluctuation). The operating condition with the highest probability is the prediction type. The fourth element, after being activated by linearity, outputs the prediction occurrence time. (Unit: min); Predicted confidence level This is the product of the operating condition probability and the data quality score, ensuring that the confidence level reflects both the model's judgment and the data's reliability.

[0054] In this method, the operating condition prediction serves as the starting point of the five-stage architecture. The time-series prediction model provides basic time-series processing capabilities. The pyrolysis reaction stage identifier is used to distinguish different reaction scenarios such as dechlorination, main pyrolysis, and high-temperature cracking. The mechanism activation factor triggers weight enhancement based on the specific reaction characteristics related to PVC dechlorination and hydrogen free radicals. The data quality score is used to screen reliable input data. The key features of each stage are focused through staged gating weight adjustment. Finally, the prediction result includes the operating condition type, time, and confidence level.

[0055] The input data for this stage comes from the sensor acquisition system, and after data preprocessing, it is divided into time steps. The frequency of s is input into the LSTM model. The model is trained using the Adam optimizer with a learning rate of 0.001. The loss function is a weighted sum of cross-entropy loss (for work condition type classification) and mean squared error loss (for occurrence time regression), i.e. The model converged after 500 rounds of iterative training. In the output prediction results, when... When this happens, the parameter preloading of the weight adaptation module is directly triggered; when When this happens, consistency verification must be performed based on the prediction results of the previous three time steps before triggering subsequent modules; when At this time, the results are only used as a reference and no active response is triggered to ensure the reliability of the working condition prediction and the efficient collaboration of subsequent modules.

[0056] S3: Weight adaptation adopts a staged four-dimensional dynamic weight adaptive algorithm constrained by carbon deposition characteristics; a four-dimensional weight model is constructed based on parameters such as the correlation degree of pyrolysis mechanism and fingerprint spectrum similarity, and the weight is dynamically corrected by carbon deposition density and thermal conductivity; the model sets a constraint that the total weight of core mechanism data is not lower than the preset value, and the feature weight is adjusted by configuring a nonlinear correction function according to the pyrolysis stage.

[0057] The core of weight adaptation is to construct a four-dimensional weight evaluation matrix. The four evaluation indicators are: correlation degree of pyrolysis mechanism, etc. Fingerprint similarity Two-way calibration deviation Compatibility with physical properties of carbon deposits , where n is the total number of input features (16 dimensions, consistent with step S2), and the weight of each feature i is... The basic weights are calculated using a weighted average of four-dimensional indicators. The formula for the basic weights is as follows: In the formula , , , The inherent weight coefficients of the four-dimensional indicators are determined through the analytic hierarchy process to ensure that the mechanism-related indicators account for the highest proportion.

[0058] Correlation of pyrolysis mechanism The correlation between feature i and the core pyrolysis mechanism is calculated using grey relational analysis, with the reference sequence being an ideal feature sequence constructed based on mechanisms such as PVC dechlorination and carbon deposition. (m is the number of samples), the original sequence of feature i is ,but The correlation coefficient , The resolution factor is set to 0.5, a commonly used value in the industry. A higher value indicates a higher correlation, such as the characteristic of HCl concentration. Typically ≥0.85.

[0059] Fingerprint similarity The similarity between the current feature sequence and the baseline map sequence is calculated using the Dynamic Time Warping (DTW) algorithm. Let the time series sequence of the current feature i be... The corresponding baseline map sequence is Construct the distance matrix , of which elements DTW distance Calculated using dynamic programming: ,but , The maximum reference DTW distance for this feature (obtained from historical normal operating data, such as 5.2 for raw material composition features). The closer the value is to 1, the higher the similarity.

[0060] Two-way calibration deviation The degree of deviation of the quantized feature i after sensor-model bidirectional calibration is calculated using the following formula: In the formula These are the calibrated feature values. These are the true values ​​measured in the same period of the experiment (obtained through offline sampling and analysis). The smaller the value, the smaller the deviation. At that time, the feature weight will be additionally suppressed.

[0061] Compatibility of physical properties of carbon deposits The compatibility of associated feature i with the current carbon deposition state, and the core associated carbon deposition density. and thermal conductivity The formula is In the formula and The reference value for the physical properties of carbon deposits corresponding to feature i (such as the carbon deposit thickness feature). ), , , , (These are all extreme values ​​of the physical properties of pyrolytic carbon deposits, determined experimentally.) Take 0 at that time.

[0062] Dynamic correction based on the physical properties of carbon deposits is achieved through a correction factor. accomplish, In the formula , The average physical properties of carbon deposits are given, and the corrected weights are... ,when (During high-density carbon deposition) This will increase the weight of features such as carbon deposit thickness and heat flux.

[0063] The total weight constraint of core mechanism data is expressed through inequalities. Implementation, in which The core mechanism is associated with a set of features (such as HCl concentration, temperature, PVC content, etc., a total of 6 features). A preset weight threshold (0.6, ensuring the total weight of core mechanism features exceeds 60%) is used. If the constraint is not met, the weights are adjusted using the Lagrange multiplier method. Lagrange function , To adjust the step size (take 0.01), The Lagrange multiplier is used until the constraints are met.

[0064] The staged nonlinear correction function is designed in conjunction with the stage identifier S(t) of step S2. The correction function is used in the dechlorination stage (S(t)=1). Further strengthen the weighting of mechanism-related features; main pyrolysis stage ( )use Highlighting the correlation characteristics of carbon deposition; high-temperature pyrolysis stage ( )use It focuses on features with high graph similarity, and the final weights are adjusted accordingly. Ensure that the sum of the weights is 1.

[0065] In this step, the weights are adapted to the predicted working conditions. The construction parameters of the four-dimensional weight model include the correlation of pyrolysis mechanism, fingerprint spectrum similarity, bidirectional calibration deviation and carbon deposition physical properties. Real-time data of carbon deposition density and thermal conductivity are used to dynamically correct the weights of each feature. The total weight constraint of core mechanism data ensures the dominant position of mechanism-related data. The nonlinear correction function for different stages such as dechlorination and main pyrolysis realizes the staged adaptation of weights.

[0066] The input to this step includes the condition prediction results from step S2 (stage identifier S(t), condition confidence level). The real-time feature data collected by the sensor, the physical property detection data of carbon deposits, and the bidirectional calibration results are all used, and the time step for weight calculation is consistent with S2. When the confidence level of the operating condition prediction At that time, the weight update frequency is increased to once every 0.5 seconds, quickly adapting to extreme working conditions; when At this time, the average weight of the first 5 time steps is used to maintain weight stability. The final output weight vector It directly impacts subsequent data purification modules, enhancing the targeting of data processing through weighted feature inputs, such as the HCl concentration characteristics during the dechlorination stage. Typically ≥0.15, which is much higher than other non-core features.

[0067] S4: Perform bidirectional sensor calibration. This step uses a pyrolysis sensor-specific failure adaptation algorithm. For pyrolysis sensor-specific failure modes such as HCl corrosion drift and carbon deposit adhesion thickness measurement error, an adaptive correction coefficient coupled with carbon deposit characteristics is designed during the design phase. Based on the calibrated data, the coupling coefficient of the digital twin model is back-optimized by combining the correlation weight of the pyrolysis mechanism to form a bidirectional closed loop.

[0068] The core of the failure adaptation algorithm for pyrolysis sensors is to construct a failure mode assessment and two-end correction mechanism. First, failure features are extracted from the sensor output data to identify the HCl corrosion drift failure factor. And carbon deposit adhesion thickness measurement error factor This serves as the basis for calculating the adaptive correction coefficient. HCl corrosion drift failure factor. The correlation between sensor output deviation and HCl concentration is quantified using a model, and the formula is as follows: In the formula The corrosion drift coefficient (value 3.2 × 10⁻⁶) -6 ppm -1 ·h -1 (calibrated by accelerated aging test of HCl corrosion resistant sensor). The cumulative exposure amount (ppm·h) of HCl concentration before time t. For reference temperature, The current sensor operating temperature (K) is represented by this factor. The larger this factor is, the more significant the error caused by corrosion drift will be.

[0069] Carbon deposit adhesion thickness measurement error factor The formula is used to quantify measurement deviations caused by carbon deposits adhering to the sensor probe surface, especially for carbon deposit thickness sensors. In the formula Real-time carbon deposition density (g / cm³) 3 ), The carbon deposition adhesion coefficient (experimentally determined, reflecting the ability of carbon deposits to adhere to the sensor surface). The effective area of ​​the sensor probe. For sensor measurement accuracy, The unit is mm, which directly corresponds to the positive deviation in thickness measurement caused by carbon deposits.

[0070] Adaptive correction coefficients coupled with stage and carbon deposition characteristics Dynamic adjustments are made for different sensors (feature i), and the calculation incorporates stage identifiers. Failure factors and physical properties of carbon deposits, the formula is as follows: In the formula and As a failure factor weight, HCl concentration sensor , Temperature sensor , Carbon deposit thickness sensor , The stage correction function is determined by experiments on the correlation between sensor type and failure mode. Based on the values ​​taken in the pyrolysis stage, the dechlorination stage ( Main pyrolysis stage High-temperature pyrolysis stage Matching the severity of failure at different stages; carbon deposition characteristic correction function. , This represents the average density of carbon deposits, used to enhance error correction during high-density carbon deposit formation.

[0071] The calibration calculation of sensor data is divided into two steps: coarse calibration and fine calibration. The coarse calibration formula is: , The original output value of the sensor. This is the coarse calibration result. Fine calibration incorporates feedback from historical calibration data, using the following formula: In the formula The actual offline measurement value at time k. The length of the historical data window. This is the deviation attenuation coefficient, ensuring that recent calibration deviations have a greater impact on current results, after fine calibration. This is the final sensor calibration data.

[0072] The inverse optimization of the coupling coefficient of the digital twin model uses calibrated data as a baseline, with the goal of minimizing the deviation between the model output and the calibration data. The optimization target is the heat transfer coupling coefficient in the model. Reaction coupling coefficient and mass transfer coupling coefficient The objective function is optimized as follows: In the formula The correlation degree of the pyrolysis mechanism calculated in step S3 is used to assign greater optimization weight to the core mechanism features. is the output function of feature i in the digital twin model.

[0073] The minimum value of the objective function is found using the gradient descent method, yielding the update formula for the coupling coefficient: , , In the formula The learning rate is determined experimentally based on the model's convergence speed. The gradient is calculated using the chain rule. k represents each coupling coefficient. The partial derivatives of the model output with respect to the coupling coefficients are obtained analytically from the mechanism formula of the digital twin model. For example, the partial derivatives of the heat transfer coupling coefficients can be derived using Fourier's law.

[0074] The two-way closed loop is achieved through a cycle of "calibration data feedback - model parameter update - model-assisted calibration": after the sensor data is calibrated, it is input into the digital twin model, driving the model to output the carbon deposition state prediction result; the deviation between the prediction result and the calibration data is used as an optimization signal to update the model coupling coefficient in reverse; the updated model can output the theoretical output range of each sensor. When the new sensor raw data exceeds this range, the calibration algorithm is triggered to be executed first, forming a two-way collaborative optimization of sensor data and model parameters.

[0075] In this step, the two-way calibration of the sensor serves as a data reliability assurance step. First, for specific failure modes such as HCl corrosion and carbon deposit adhesion, adaptive correction coefficients are calculated based on the current reaction stage and the physical characteristics of carbon deposits to complete the sensor data calibration. Then, based on the calibrated reliable data, the coupling coefficient of the digital twin model is optimized in reverse by using the correlation weight of the pyrolysis mechanism, thereby achieving two-way correction of sensor data and model parameters.

[0076] The inputs for this step include the stage identifier S(t) of step S2 and the correlation of the pyrolysis mechanism in step S3. The system outputs real-time raw sensor data, carbon deposition physical property detection data, and historical offline calibration data. The calibration algorithm's execution frequency is synchronized with the sensor's sampling frequency (10Hz~20Hz, adjusted according to the pyrolysis stage). The output consists of two parts: firstly, calibrated data for each feature. The first step is to directly input the data purification module from step S5; the second is to update the coupling coefficient of the digital twin model. This is used to update the digital twin model and improve the accuracy of subsequent carbon deposition state predictions. When the failure factor of a certain sensor... or When this happens, the sensor will trigger a redundancy switching prompt to ensure the continuity of data acquisition.

[0077] S5: Data purification adopts a pyrolysis kinetics-multiphysics coupling algorithm, which integrates the sensitivity derivative of carbon deposit growth rate into information gain calculation and constructs a pyrolysis kinetics weighted model; combined with the time-varying physical properties of carbon deposit, a multiphysics residual formula coupling heat transfer coefficient-carbon deposit thickness-dechlorination reaction rate is established; and specific residual thresholds are set for different pyrolysis stages to remove pyrolysis-specific abnormal data such as incomplete dechlorination.

[0078] The core input for data purification is the calibrated data output from step S4. Covering temperature T(t) and HCl concentration Carbon deposit thickness To determine the 16-dimensional characteristics, including the heat transfer coefficient h(t), the sensitivity derivative of the carbon deposit growth rate must first be calculated to quantify the influence of key parameters on carbon deposit growth, providing a basis for information gain weighting. The carbon deposit growth rate can be referenced here from the expression in S1 above. In another example, to consider the inhibitory effect of hydrogen free radicals on carbon deposit growth, the Arrhenius equation can also be referenced, with the basic expression being: In the formula The pre-exponential factor (obtained by fitting carbon deposition growth experimental data). The activation energy for carbon deposition growth (experimentally determined). The gas constant is... The concentration of hydrogen free radicals is n, and n=0.6 is the reaction order (determined based on kinetic experiments of hydrogen free radicals inhibiting carbon deposition).

[0079] Sensitivity derivative of carbon deposition growth rate Used to measure characteristics The fluctuations in carbon deposition growth rate caused by minute changes are calculated using three key parameters: temperature, HCl concentration, and hydrogen radical concentration. The temperature sensitivity derivative is also considered. The larger the derivative, the more significant the effect of temperature fluctuations on carbon deposition; the sensitivity derivative to HCl concentration... ,in The promoting factor of HCl on coke growth (calibrated by HCl concentration gradient experiments); the sensitivity derivative to hydrogen radical concentration. The negative sign indicates that increased hydrogen radical concentration inhibits carbon deposition. The sensitivity derivatives of other characteristics are uniformly transformed through grey relational analysis, i.e. , The correlation degree of the pyrolysis mechanism calculated for step S3 is used to ensure that the weights of non-core features are reasonable.

[0080] Information gain calculation is used to evaluate the discriminative power of features. The traditional information gain formula is: ,in For category entropy, Let Y be the conditional entropy and Y be the data category (normal / abnormal). Incorporating the sensitivity derivative, a weighted information gain model of pyrolysis kinetics is constructed: In the formula The stage weighting coefficient represents the dechlorination stage. Main pyrolysis stage High-temperature pyrolysis stage To match the core needs of carbon deposit control at different stages, the carbon deposit growth is most active in the main pyrolysis stage, so it is given the highest weight.

[0081] A weighted model of pyrolysis kinetics is constructed based on weighted information gain, and preliminary feature enhancement is performed on the input data. The formula is as follows: The denominator is weighted and normalized, so that the enhanced data retains the original information while highlighting features that significantly affect carbon deposition growth, such as the carbon deposition thickness characteristics during the main pyrolysis stage. It is typically 1.8 to 2.2 times the original calibration data.

[0082] The construction of the multiphysics residual formula requires coupling three physical fields: heat transfer, reaction, and carbon deposition growth. The core correlation is the heat transfer coefficient h(t) and the carbon deposition thickness. rate of dechlorination First, establish the mechanistic relationships between the various physical fields: the heat transfer coefficient and the carbon deposition thickness satisfy... In the formula For real-time carbon deposit thermal conductivity, The reactor wall thickness; the dechlorination reaction rate follows the reaction formula in step S1, or as shown in [reference needed]. , The dechlorination rate constant is This represents the real-time concentration of PVC. This is the dechlorination activation energy.

[0083] Based on the above-mentioned mechanism, a multiphysics residual formula is constructed. This formula, through normalization, integrates the deviations of the three physical fields into a single residual value. The closer the value is to 0, the more the data conforms to the multi-physics coupling mechanism; the greater the deviation, the higher the probability of data anomalies.

[0084] The time-varying physical properties of carbon deposits are incorporated into the formula through a residual correction term, resulting in a final corrected residual of... , The residual is the average value of carbon deposit characteristics. When the carbon deposit density or thermal conductivity deviates significantly from the average value, the residual is amplified, thereby improving the sensitivity of abnormal data identification.

[0085] The specific residual thresholds for different pyrolysis stages were determined statistically from historical normal operating data, using the 3σ criterion (σ being the standard deviation of the residuals in normal data): Dechlorination stage ,in Therefore Main pyrolysis stage , Therefore High-temperature pyrolysis stage , Therefore The specific anomaly of incomplete dechlorination is reinforced through the synergistic assessment of HCl concentration and dechlorination reaction rate. and At that time, it was directly judged as an abnormality of incomplete dechlorination, among which The values ​​represent the maximum HCl concentrations for the corresponding stages (500 ppm for the dechlorination stage and 50 ppm for the main pyrolysis stage).

[0086] The final execution logic for data purification is: to perform weighted enhancement on the data. Calculate the corrected residual ,like And if the criteria for incomplete dechlorination are not met, the data is retained; if If the dechlorination condition is met, the data is discarded, and the data is supplemented using the weighted average of the first three time steps. The supplementation formula is as follows: , This is the attenuation coefficient, ensuring that the supplementary data closely reflects the true trend.

[0087] In this step, data purification uses calibrated data as input. The sensitivity derivatives of carbon deposition growth rate to key parameters such as temperature and HCl concentration are incorporated into information gain calculation to construct a weighted model. A multiphysics residual formula is established by combining time-varying data of carbon deposition density and thermal conductivity. Specific residual thresholds are set for different stages such as dechlorination and main pyrolysis. When the data residual exceeds the corresponding threshold, it is judged as an anomaly. In particular, pyrolysis-specific anomalies such as incomplete dechlorination are accurately eliminated.

[0088] The inputs for this step also include the stage identifier S(t) of step S2 and the correlation of the pyrolysis mechanism in step S3. and real-time carbon deposit physical properties data ( The data processing frequency is synchronized with the sensor sampling frequency (10Hz~20Hz). The output is the purified data. (Retained data is used directly as output, and discarded data is replaced with complete values.) This data will serve as the core input for the consistency verification of the mechanism in step S6. Its reliability directly determines the accuracy of subsequent model evolution and carbon deposition prediction. After purification, the abnormality rate of the data usually drops from 8%~12% after calibration to below 2%, and the accuracy of identifying abnormal data of incomplete dechlorination can reach more than 95%.

[0089] S6: Perform mechanism consistency verification. This step adopts a fingerprint spectrum verification algorithm specific to the pyrolysis of mixed plastics; construct a pyrolysis fingerprint spectrum library for PVC / PE / PP with multiple ratios, and expand the spectrum features to a four-dimensional vector of temperature-gas concentration-carbon deposition characteristics-reaction stage; adjust the attention weight according to the pyrolysis stage, optimize the similarity calculation of the dynamic time warping algorithm, and provide a spectrum self-evolution mechanism with mechanism constraints and incremental data.

[0090] The core input for mechanism consistency verification is the purified data output from step S5. Its primary task is to construct a dedicated fingerprint library for the pyrolysis of PVC / PE / PP in multiple ratios. The library is based on experimental data of mixed plastic pyrolysis, covering 12 typical ratios (PVC accounting for 5%~50%, and PE and PP accounting for 1:1 complementary ratios). Each ratio corresponds to 3 sets of time-series data from parallel experiments, with sampling duration covering the complete pyrolysis cycle (dechlorination-main pyrolysis-high temperature pyrolysis), and a sampling frequency of 10Hz.

[0091] The construction of the four-dimensional feature vector requires standardization and fusion of four types of features: temperature, gas concentration, carbon deposition characteristics, and reaction stage. Let the four-dimensional feature vector of the m-th ratio and the k-th group of experiments at time t be: Temperature standardization , and These are the minimum and maximum temperatures for the pyrolysis of the m-th formulation (determined experimentally, e.g., when PVC accounts for 20%). , Gas concentration standardization , To determine the maximum concentration (experimental statistical value) of the three characteristic gases under the corresponding ratio, multi-gas information is integrated by averaging.

[0092] Standardization of carbon deposition characteristics , and To determine the maximum carbon deposit thickness and density (experimentally determined) for the corresponding formulation, thickness is assigned a higher weight to match the core requirements for carbon deposit control; reaction stage coding. The dechlorination stage uses one-hot encoding to expand into a three-dimensional vector. Main pyrolysis stage High-temperature pyrolysis stage Finally, the four-dimensional feature vector is extended into a seven-dimensional vector by concatenation for subsequent calculations.

[0093] Benchmark fingerprints for each formulation The value is obtained by taking the average of the eigenvectors from three parallel experiments, as shown in the formula. Simultaneously calculate the characteristic standard deviation vector for each ratio. This is used for bias assessment in subsequent similarity calculations. The standard deviation of other dimensions is calculated in a similar way.

[0094] The core of adjusting attention weights according to the pyrolysis stage is the adaptive attention vector construction stage. The weight values ​​were determined through feature importance analysis experiments, satisfying the following conditions: Dechlorination stage ( The concentration of the gas (especially HCl) is the core characteristic, therefore The main pyrolysis stage is dominated by carbon deposition characteristics. Temperature has a significant impact during the high-temperature pyrolysis stage. The contribution of core features at each stage is enhanced by attention weighting.

[0095] The optimization of the Dynamic Time Warping (DTW) algorithm is reflected in the incorporation of attention weights and mechanistic bias constraints into the distance calculation. Let the feature vector sequence of the current purified data be... The baseline spectral sequence for matching ratios is The optimized DTW distance matrix elements are In the formula For mechanism constraint coefficients, As the mechanism deviation factor, when Falling (Within the allowable deviation range of the mechanism) ,otherwise By penalizing deviations from the data, the reliability of similarity calculation is improved.

[0096] DTW similarity calculation uses normalization processing, and the formula is as follows: , For the optimized DTW distance, For the first The maximum DTW distance for this mix ratio (obtained from statistical data of abnormal operating conditions under this mix ratio, such as when PVC accounts for 20%) ), , The mechanism is consistent. If the mechanism is determined to be abnormal, and it falls somewhere in between, further verification with subsequent incremental data is required.

[0097] The graph self-evolution mechanism is driven by both mechanistic constraints and incremental data. The trigger threshold for the incremental data pool is... Each data point (approximately corresponding to 50 seconds of sampling data) triggers self-evolution when the data volume in the pool reaches a threshold. First, the feature mean vector of the incremental data is calculated. With covariance matrix The KL divergence was used to determine the distribution difference between the incremental data and the baseline map. , For incremental data feature distribution, As the baseline spectral feature distribution, when When the distributions are similar, only the standard deviation of the baseline spectrum is updated. , This is the original baseline data volume. The standard deviation is the incremental data.

[0098] when When there are significant differences in distribution, the baseline spectrum needs to be updated in conjunction with mechanistic constraints. The update formula is as follows: , The weighting coefficients are for historical data. Mechanism constraint factor, , For the average similarity of incremental data, when hour (Complete acceptance), when hour (Reject updates) to ensure that the updated spectrum conforms to the pyrolysis mechanism. At the same time, if the incremental data corresponds to a new PVC / PE / PP ratio (with a deviation of ≥10% from the existing ratio), a benchmark spectrum for that ratio will be added to the spectrum library to achieve the expansion and evolution of the spectrum library.

[0099] In this step, the mechanism consistency verification takes the purified data and first constructs a pyrolysis-specific fingerprint spectrum library containing different mixing ratios of PVC, PE, and PP. The spectrum features are expanded into a four-dimensional vector composed of temperature, gas concentration, carbon deposition characteristics, and reaction stage. The attention weights of each feature are adjusted according to the current pyrolysis stage. The similarity between the data and the benchmark spectrum is calculated through an optimized dynamic time warping algorithm. When the incremental data reaches a set scale, the spectrum is updated automatically by combining mechanism constraints.

[0100] The inputs for this step also include the stage identifier S(t) from step S2 and real-time proportioning data of the mixed plastic raw materials. The data processing frequency is synchronized with the previous steps (10Hz~20Hz). The output consists of two parts: first, the mechanism consistency judgment result (consistent / abnormal / to be verified). Consistent data is directly input into the model evolution module of step S7, abnormal data triggers an alert and is fed back to step S4 for recalibration, and data to be verified is temporarily stored in the incremental data pool; second, the updated fingerprint spectrum library is used to improve the accuracy of subsequent verification. After this step, the mechanism consistency rate of the data entering the model evolution module can reach over 98%, ensuring the reliability of model evolution.

[0101] S7: The model evolution adopts a digital twin self-evolution algorithm that is adaptive to the pyrolysis working condition; the Arrhenius equation for carbon deposition growth and the carbon deposition density constraint are embedded into the reinforcement learning reward function; evolution is triggered when the raw material composition fluctuates or the carbon deposition rate changes abruptly, and the model parameters are self-evolved and dynamically adjusted by the collaborative driving of prediction error and multi-physics residual.

[0102] The core input to model evolution is the mechanistic consistency data output from step S6. The core of the digital twin model is the fusion of the carbon deposition growth prediction sub-model and the multiphysics coupling sub-model. The mechanism of the carbon deposition growth prediction sub-model can be referred to the expression in S1. To consider the inhibitory effect of hydrogen free radicals on carbon deposition growth, the subsequent strategy can also be referred to. Based on the Arrhenius equation, the initial parameters are obtained by fitting historical experimental data. In the formula Pre-exponential factor, For carbon deposition growth activation energy, The reaction order is... The gas constant is... For real-time temperature, The concentration of hydrogen free radicals. This is a predicted value for the carbon deposition growth rate.

[0103] In the reinforcement learning framework, the agent is the parameter adjustment module of the digital twin model, the environment is the real-time working condition of the pyrolysis reaction, and the action space is the core parameters of the model. and multiphysics coupling coefficient The adjustment amount, the state space is (Temperature, HCl concentration, carbon deposit density, carbon deposit growth rate, reaction stage), reward function It is the core of realizing the embedding of physical constraints and the guidance of evolutionary direction, and its construction integrates the three goals of carbon deposition growth mechanism, prediction accuracy and physical constraints.

[0104] The basic form of the reward function is ,in The reward is for prediction accuracy and is used to measure the deviation between the model's predicted value and the actual value. To ensure the mechanism aligns with the reward, the Arrhenius equation for carbon deposition growth is relevant. Physical constraints are used to ensure that carbon deposition density remains within a reasonable range.

[0105] Prediction accuracy reward In the formula The prediction error penalty coefficient (determined by experiments based on model accuracy requirements). This represents the true value of the carbon deposition growth rate after purification in step S5. The multiphysics correction residuals calculated in step S5. This is the residual penalty coefficient. When the prediction error is less than 0.01 mm / h and the residual is less than 0.1, Approaching 1, the maximum reward is reached.

[0106] Mechanism aligns with reward In the formula This is a penalty coefficient for mechanistic bias; this reward term ensures that the model predictions always follow the Arrhenius mechanism of carbon deposition growth, and that the predicted values ​​perfectly match the mechanistic relationship. If the mechanism is deviated from (e.g., the ratio is greater than 1.2 or less than 0.8 due to parameter drift), the reward value will quickly drop below 0.

[0107] Physical constraint reward In the formula The carbon deposition density predicted by the digital twin model. This represents the physical upper limit of carbon deposition density (as determined experimentally). This is the penalty coefficient for mild constraints. To severely constrain the penalty coefficient, a strong penalty is applied when the carbon deposition density exceeds the upper limit, thus preventing the model output from violating physical laws.

[0108] The evolution triggering mechanism is determined by two thresholds: the coefficient of variation of the raw material components and the coefficient of change of the carbon deposition rate. The coefficient of variation of the raw material components follows the definition of step S1. In the formula The standard deviation of the PVC, PE, and PP component contents within the 5 minutes prior to time t. Trigger threshold corresponding to average content ,when The raw material composition was determined to be fluctuating beyond the standard.

[0109] Coefficient of change in carbon deposition rate In the formula For time window, The threshold is set to the actual carbon deposition growth rate from the previous window. ,when A sudden change in the carbon deposition rate is identified. When any threshold condition is met, the model evolution process is triggered, and the current operating condition characteristics are recorded as the evolution baseline.

[0110] The parameter self-evolution employs a gradient descent method driven by both prediction error and multi-physics residuals, with the optimization objective being to maximize the cumulative reward. The core parameter update formula is as follows: , where gradient

[0111] ,gradient

[0112] ,gradient

[0113] In the formula The learning rate (determined experimentally based on the model's convergence speed). , The calculation logic for the reward gradient of other parameters is similar, ensuring that the parameter update direction conforms to the carbon deposition growth mechanism.

[0114] Multiphysics coupling coefficient ( The updated fusion of calibration data deviation in step S4 is calculated using the following formula: In the formula To calibrate the bias weights, The feature prediction values ​​output by the model. The data from step S4, after calibration, enables the synergistic optimization of carbon deposition growth prediction and multi-physics coupling.

[0115] The dynamic adjustment of parameter boundaries is based on the raw material composition and reaction stage under the current operating conditions, and the boundary range is determined using statistical learning methods. , ,in , , The baseline pre-exponential factor for the m-th raw material ratio (based on experimental data statistics). The boundary calculation logic is similar. When the coefficient of variation of raw material components... At that time, the boundary range was expanded to This improves the model's adaptability to extreme working conditions.

[0116] The evolution termination condition is the cumulative reward increment. Alternatively, after 500 iterations, the current parameters will be used as the new model baseline parameters and stored in the parameter history database. At the same time, the corresponding working condition characteristics will be recorded to provide a reference for parameter initialization for subsequent similar working conditions.

[0117] In this step, the model evolution takes data that has passed mechanistic verification as input. Physical constraints such as the Arrhenius equation for carbon deposition growth and the upper limit of carbon deposition density are embedded into the reward function of reinforcement learning. The evolutionary process is triggered when a sudden change in the coefficient of variation of raw material components or the carbon deposition growth rate reaches a set threshold. Through the synergistic effect of model prediction error and multi-physics residuals, the model's k0 and E are driven. a Parameters such as n are updated automatically, and the parameter boundary range is dynamically adjusted according to changes in working conditions.

[0118] The input for this step also includes the stage identifier for step S2. Calibration data after step S4 The process incorporates real-time raw material proportioning data. The execution frequency of the evolution process is related to fluctuations in operating conditions. Under stable operating conditions, routine optimization is triggered every 30 minutes, while under fluctuating operating conditions, emergency evolution is triggered in real time. The output is an updated digital twin model (including optimized parameters and parameter boundaries). This model will be directly used for predicting the carbon deposition state in step S8. After evolution, the prediction error of the carbon deposition growth rate of the model can be reduced from the initial 8%~12% to below 3%, and the adaptation response time to raw material fluctuations is ≤2s.

[0119] S8: Perform carbon deposition state prediction. This step adopts a pyrolysis staged multi-scale fusion algorithm; dynamically adjust the feature extraction kernel size and loss function weight according to the pyrolysis stage; use an improved feature extraction model to capture data fluctuations in the short term; and combine the working condition prediction results with the digital twin mechanism model in the long term to achieve multi-objective prediction of carbon deposition thickness, density and growth rate.

[0120] The core inputs for predicting carbon deposition state include the parameters of the digital twin model evolved from the S7 step and the mechanistic consistency data from the S6 step. and the S2 step working condition prediction results (extreme working condition types) Time of occurrence Confidence level of prediction The prediction target is carbon deposit thickness. Carbon deposit density Carbon deposit growth rate The short-term forecast timescale Long-term forecast timescale The prediction frequency is synchronized with the data acquisition frequency (10Hz).

[0121] The core of staged multi-scale feature extraction is to dynamically adjust the convolution kernel parameters based on the stage identifier S(t) of step S2. The convolution kernel design follows the principle of "stage feature matching": in the dechlorination stage (S(t)=1), the pyrolysis reaction is intense and the data fluctuation frequency is high. A small-sized convolution kernel is used to capture high-frequency details. The kernel size is set to... convolution stride Feel the wild (dilation1=1 represents the void ratio); main pyrolysis stage ( Carbon deposit growth is stable, and the data shows a low-to-medium frequency trend. A medium-sized convolutional kernel is used to balance detail and global information. (Kernel size...) Step length Feel the wild High-temperature pyrolysis stage ( The reaction tends to be gradual, with small data fluctuations. A large-size convolutional kernel is used to capture the global trend. (Kernel size...) Step length Feel the wild (The receptive field is expanded by dilation3=2).

[0122] The improved feature extraction model adopts a "multi-scale convolution + attention fusion" structure, with the input being a temporal feature matrix. ( For time step, (For feature dimensions), firstly, features at different scales are extracted through three parallel convolutional branches: Branch 1 (small kernel) outputs... Branch 2 (core) output Branch 3 (Big Core) Output Where C=32 is the number of convolution channels, and the convolution kernel weights are initialized using a He normal distribution.

[0123] Feature fusion employs a stage-adaptive attention mechanism, with attention weights... satisfy Dechlorination stage , , Main pyrolysis stage High-temperature pyrolysis stage fusion features After passing through the BatchNorm layer and the ReLU activation function, the data is input into the subsequent prediction module.

[0124] The short-term prediction module is based on an improved feature extraction model, and incorporates a bidirectional LSTM (Bi-LSTM) to capture temporal dependencies. The Bi-LSTM hidden layer dimension is set to 128, and the forward LSTM output... Output to LSTM After splicing, the short-term prediction results are output through a fully connected layer: ,in This is the weight matrix. For bias vectors, This is the short-term prediction vector.

[0125] The long-term forecasting module employs a fusion strategy of "mechanism model + data-driven approach." First, it calculates the baseline forecast value using an evolved digital twin mechanism model. ,in ( (Predicted carbon deposition growth rate for step S7). ( (This refers to the density temperature coefficient).

[0126] The baseline prediction value is corrected based on the working condition prediction results of step S2. The correction formula is as follows: ,in The operating condition influence coefficient, This is a correction matrix for operating conditions, specifically for high PVC ratio operating conditions. Sudden change in heat transfer fluid conditions Raw material fluctuation conditions Long-term forecast accuracy can be improved through operating condition correction.

[0127] The multi-objective loss function employs a staged weighted design, comprehensively considering the differences in importance of each prediction objective at different stages. The loss function formula is as follows: ,in For the stage weight coefficients, satisfying The weighting coefficients for each stage are as follows: Dechlorination stage (Focusing on growth rate prediction); Main pyrolysis stage (Focusing on thickness prediction); High-temperature pyrolysis stage (Focusing on density prediction).

[0128] The loss component calculation uses a hybrid loss form to balance bias and robustness: , and The calculation methods are similar, among which These are the MSE weighting coefficients. Mean square error, The mean absolute error, This is the actual carbon deposit thickness data after purification in step S5.

[0129] The reliability of the prediction results is verified by coupling prediction bias with confidence level, and the bias coefficient is used to determine the reliability of the prediction results. ,when When the prediction result is "highly reliable", when... or When the result is "medium reliability", it needs to be smoothed by combining the results of the previous three predictions; when... or At this point, the system is considered "low reliability," triggering the S7 step model to re-evolve.

[0130] In this step, the carbon deposition state prediction follows the output of the evolved digital twin model. The kernel size of the multi-scale feature extraction is dynamically adjusted according to the current pyrolysis stage. Small kernels are used to extract features in the dechlorination stage, while large kernels are used in the main pyrolysis stage. At the same time, the weights of each objective in the loss function are adjusted. Short-term prediction captures subtle data fluctuations through the improved feature extraction model, while long-term prediction combines the working condition prediction results from step S2 with the digital twin mechanism model to output the predicted results of carbon deposition thickness, density, and growth rate.

[0131] The switching between the two prediction architectures in this step is determined by the confidence level of the operating condition prediction in step S2. Triggered, when At that time, a multi-scale convolution + Bi-LSTM architecture is adopted, utilizing the long-term dependency capture capability of Bi-LSTM to adapt to the abrupt changes in carbon deposition under extreme working conditions; when At this time, a multi-scale convolutional + GRU architecture is adopted to improve the prediction efficiency under normal conditions by leveraging the lightweight characteristics of GRU. The core parameters of the two architectures have a mapping relationship: the 128-dimensional hidden layer of Bi-LSTM corresponds to the 64-dimensional hidden layer of GRU, and the stage attention weights of Bi-LSTM... Stage coding vectors of GRU satisfy This ensures consistency in parameter output. The prediction results for both architectures must be constrained by the parameter boundary conditions output from the S7 step model evolution, i.e. Predicted values ​​that exceed the boundary need to be corrected using a truncation function: The corrected prediction results serve as the core input for step S9.

[0132] The output of this step is a graded and reliable carbon deposition state prediction report, including predicted values, deviation coefficients, and reliability levels at different time scales. High-reliability prediction results are directly used to generate carbon deposition control strategies for pyrolysis reactors, medium-reliability results require manual verification before application, and low-reliability results trigger early warnings and initiate the model optimization process. Field measurements have verified that the short-term predicted carbon deposition thickness error is ≤2%, and the long-term prediction error is ≤8%, effectively supporting early prevention and control of carbon deposition.

[0133] S9: The decision feedback adopts a multi-objective coking control decision algorithm specific to the pyrolysis reactor; combined with reactor-specific constraints such as coating tolerance temperature, it integrates six-dimensional indicators such as carbon deposit thickness and fingerprint spectrum similarity to construct a coupled early warning function; it divides the early warning level into five levels, balances the coking effect, energy consumption and product yield, and generates a step-by-step coking control strategy such as operating condition adjustment and light / deep coking.

[0134] The core input to the decision feedback is the highly reliable carbon deposit state prediction result output from step S8. Fingerprint similarity in step S6 Calibration deviation in step S4 Carbon deposition characteristics data for step S5 ( ) and data quality assessment indicators At the same time, reactor-specific constraints must be included, including the coating's temperature tolerance. (Determined experimentally based on the high-temperature resistant coating material of the reactor inner wall), upper limit of product yield loss. (Production process requirements) and upper limit of coke removal energy consumption (Equipment power constraints).

[0135] First, the six-dimensional indicators are standardized to eliminate dimensional differences and form a unified evaluation benchmark. Carbon deposit thickness indicator. , The maximum safe limit for carbon buildup thickness (as specified in the equipment maintenance manual). The higher the value, the higher the risk of carbon buildup; carbon buildup growth rate index , The rate safety threshold (experimentally determined). Fingerprint similarity index , The similarity calculated in step S6, The larger the value, the more serious the deviation from the mechanism.

[0136] Calibration Deviation Index , This represents the average value of the calibration deviations for each feature in step S4. Carbon deposition characteristics index , This represents the extreme value of carbon deposition characteristics. Data quality indicators , To provide a comprehensive score for data quality, , The larger the value, the lower the data reliability.

[0137] The coupled early warning function is constructed using a combination of weighted fusion and constraint penalties, as shown in the formula: ,in The weights of the six-dimensional indicators were determined using the analytic hierarchy process (AHP) combined with expert scoring: Ensure that the core indicators of carbon deposit status account for more than 50%.

[0138] The constraint penalty term P(t) is used to quantify the degree of violation of reactor-specific constraints, and the formula is as follows: ,in The current reactor temperature, Product yield loss calculated based on the current state of carbon deposition ( When temperature or yield loss exceeds the limit, the penalty term increases rapidly, enhancing the sensitivity of the warning function to constraints.

[0139] The five-level warning level is determined by the warning function value. The threshold is determined by classifying and combining actual production conditions with risk assessment: Level 1 Warning (Safety) The risk of carbon buildup is extremely low and no intervention is needed; Level 2 alert (pay attention). The risk is low, but monitoring needs to be strengthened; Level 3 warning (mild). There are certain risks involved, and operating conditions need to be adjusted; Level 4 warning (moderate). High risk, requires mild descaling; Level 5 warning (severe). The risk is extremely high, and deep desiccation is required.

[0140] The generation of the stepped focus control strategy is based on multi-objective optimization, with the objective function being: ,in For clearing the scorch ( , (for strategy execution cycle) Energy consumption for cleaning coke (unit: kW·h). For product yield loss, The target weight is used to balance the desiccant removal effect and economy.

[0141] The specific strategies for each level are as follows: Level 1 warning implements a "monitoring strategy," maintaining a data acquisition frequency of only 10Hz without adjusting operating conditions; Level 2 warning implements an "enhanced monitoring strategy," increasing the acquisition frequency to 20Hz and outputting a status report every 15 minutes; Level 3 warning implements an "operating condition adjustment strategy," adjusting the heat transfer fluid flow rate. and raw material feed rate Achieve focus control The baseline flow rate and velocity are used.

[0142] Level IV warning implements a "mild coking removal strategy," employing low-pressure steam coking removal with low pressure. Clearing time At the same time, the temperature of the heat transfer fluid is reduced to After descorching, through Rapidly restore operating conditions; Level 5 warning implements a "deep coking strategy," employing a combination of high-pressure nitrogen and steam for coking removal, with high coking pressure... Clearing time Before decoking, feeding must be stopped, and after decoking, a heating process should be initiated. (t is the heating time in minutes) to restore to the working temperature.

[0143] Feedback on the strategy execution results is achieved through "state assessment - parameter update," which updates the measured values ​​of the carbon deposition state after strategy execution. Compare the values ​​with the predicted values ​​and calculate the correction factor. and will The strategy execution record is fed back to the working condition prediction module in step S2 to correct the prediction bias of the LSTM model. The formula is as follows: This completes the closed-loop optimization of "prediction-decision-feedback".

[0144] Multi-objective optimization objective function The solution is obtained using a constrained nonlinear programming algorithm, with the following constraints:

[0145] In the formula The upper and lower limits are determined by the reactor equipment manual. The constraints are incorporated into the objective function using the Lagrange multiplier method to construct an augmented objective function: ,in Let be the relaxation function for the constraint inequality. It is a Lagrange multiplier, and its range is 1 / 2. The stricter the constraints, the better. The larger the value, the more important it is to verify the optimal solution through strategy simulation: Use the digital twin model from step S7 to simulate the carbon deposition state after strategy execution. If the simulation results satisfy... and If the output is positive, then the policy execution instruction is output; otherwise, the weight coefficients are adjusted. Solve again.

[0146] The effectiveness of the strategy after implementation is evaluated using incremental revenue calculation: ,in These represent the product yields before and after strategy implementation. For the revenue per unit of product, when When the strategy is deemed effective, its parameters are stored in the strategy library for rapid matching of similar operating conditions in the future; when When the time comes, the parameters of the S2 step condition prediction model are updated to optimize the threshold settings and improve the economy of subsequent strategies.

[0147] In this step, the decision feedback serves as the closed-loop output of the five-stage architecture. It takes the carbon deposition state prediction result as the core input, combines specific constraints such as reactor coating tolerance temperature and product yield loss limit, and integrates six dimensions of indicators, including carbon deposition thickness, growth rate, fingerprint spectrum similarity, calibration deviation, carbon deposition characteristics, and data quality, to construct a coupled early warning function. The early warning results are divided into five levels, and a tiered coking control strategy is generated according to different early warning levels to achieve a multi-objective balance between coking removal effect, equipment energy consumption, and product yield. The execution results are then fed back to the operating condition prediction stage in step S2 to complete the closed-loop optimization.

[0148] The output of this step is the coking control strategy execution command and closed-loop feedback data. The execution command directly controls the reactor actuator through the PLC system, and the feedback data is used for model optimization in step S2. Industrial trials have verified that after adopting this strategy, the number of reactor shutdowns due to excessive carbon buildup is reduced by 80%, coking removal energy consumption is reduced by 35%, and the product yield is stabilized at over 92%, achieving a balance between the economy and reliability of carbon buildup control.

[0149] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0150] The core requirement for carbon deposition control in pyrolysis reactors is to solve the problems of unpredictable carbon deposition status and difficulty in adapting descaling strategies caused by fluctuations in the composition of mixed plastic raw materials. The achievement of its technical effect depends on the synergy of multi-level technical means of "data end-link optimization - dynamic model evolution - precise decision-making closed loop". Each step forms an organic whole through data flow and parameter feedback, ultimately achieving accurate early warning and efficient control of carbon deposition.

[0151] The technical logic begins with the five-stage closed-loop algorithm architecture built by S1, consisting of "condition prediction, weight adaptation, data purification, model evolution, and decision feedback." This architecture, by embedding the temperature-HCl concentration correlation mechanism and staged feature weight allocation in the PVC dechlorination stage, defines the core principle of "mechanism constraint + data-driven" for subsequent steps. Fluctuations in the composition of mixed plastic raw materials are quantified using the coefficient of variation (CVcomp). A rapid response mechanism triggered when CVcomp ≥ 10% enables the entire system to proactively perceive changes in raw material properties. This avoids the limitations of a single data processing mode at the architectural level and provides directional guidance for the targeted execution of subsequent steps.

[0152] The S2 operating condition prediction algorithm, employing a stage-aware, gated attention fusion approach, is a technical implementation of the closed-loop architecture's "operating condition prediction" module. It captures temporal data features using an LSTM model, combines this with the pyrolysis reaction stage identifier S(t) to accurately distinguish reaction scenarios, and then quantifies reaction activity through PVC dechlorination and hydrogen radical activating factors, along with data quality scoring to filter reliable inputs. This combination of "temporal processing + stage identification + mechanism activation" strengthens key features such as the coefficient of variation of raw material components, accurately outputting extreme operating condition types, occurrence times, and confidence levels. This provides stage-based and operating condition-based parameter basis for S3 weight adaptation, ensuring that subsequent weight allocation does not deviate from the actual reaction scenario.

[0153] The S3 algorithm, a four-dimensional dynamic weighted adaptive algorithm constraining carbon deposition characteristics, builds upon the operating condition prediction results of S2. It combines four-dimensional indicators such as pyrolysis mechanism correlation and fingerprint spectrum similarity with real-time data on carbon deposition density and thermal conductivity. Through dynamic correction factors and total weight constraints (≥60%) of core mechanism data, it achieves phased optimization of feature weights. The design of strengthening the weight of mechanism-related features in the dechlorination stage and highlighting the correlation features of carbon deposition characteristics in the main pyrolysis stage enables the weight vector to accurately match the carbon deposition control requirements of each stage. This provides differentiated feature priority guidance for S4 sensor calibration and S5 data purification, ensuring that core mechanism data plays a dominant role in subsequent processing.

[0154] The S4 sensor bidirectional calibration targets specific failure modes such as HCl corrosion drift and carbon deposit adhesion thickness measurement errors. It combines the stage markers of S2 with the correlation between the pyrolysis mechanism of S3 to calculate adaptive correction coefficients that couple stage and carbon deposit characteristics. This bidirectional closed-loop approach of "targeted failure mode correction + calibration data-driven model optimization" not only improves the reliability of sensor data but also optimizes the coupling coefficients of the digital twin model through calibration data feedback. This provides high-quality input data for S5 data purification and, in turn, reinforces the mechanistic consistency of the digital twin model, forming a bidirectional optimization of "data-model".

[0155] The S5 data purification algorithm employs a pyrolysis kinetics-multiphysics coupling algorithm. This algorithm integrates the sensitivity derivative of the carbon deposit growth rate into the information gain calculation, constructing a weighted model to highlight features significantly affecting carbon deposit growth. Then, through a multiphysics residual formula coupling heat transfer coefficient, carbon deposit thickness, and dechlorination reaction rate, and combined with specific residual thresholds for each stage, outlier data is eliminated. This step uses the calibration data from S4 as input and the weights from S3 as feature priority criteria. Through "kinetic weighting + multiphysics residual verification," it effectively eliminates pyrolysis-specific outlier data such as incomplete dechlorination, providing a data foundation with high mechanistic consistency for the S6 mechanism consistency verification. This ensures the reliability of subsequent model evolution from a data perspective.

[0156] The S6 mechanism consistency verification constructs a pyrolysis fingerprint database for PVC / PE / PP with multiple ratios, expanding the features into a four-dimensional vector of temperature, gas concentration, carbon deposition characteristics, and reaction stage. It then optimizes the DTW algorithm's similarity calculation using staged attention weights. Its core value lies in using the data purified from S5 as input, and verifying the mechanistic compliance of the data through "benchmark spectrum matching + mechanism constraints + incremental data self-evolution," ensuring that the data entering the S7 model evolution exhibits consistency in pyrolysis reaction patterns. The spectrum self-evolution mechanism allows the system to adapt to new raw material ratios and reaction conditions, avoiding verification deviations caused by the solidification of benchmark spectra, and providing dynamically updated mechanism references for model evolution.

[0157] The S7 model evolution embeds the Arrhenius equation for carbon deposition growth and carbon deposition density constraints into a reinforcement learning reward function. Using data consistent with the S6 mechanism as input, parameter self-evolution is triggered when there are fluctuations in raw material composition (CVcomp ≥ 10%) or sudden changes in carbon deposition rate (≥ 30%). This approach of "mechanism constraint embedded in reward function + dual-threshold triggered evolution" ensures that model parameter updates always follow the physical laws of carbon deposition growth. Simultaneously, through the synergistic driving of prediction error and multi-physics residuals, the digital twin model can dynamically adapt to changes in raw materials and operating conditions, providing high-precision model support for S8 carbon deposition state prediction and solving the prediction bias problem caused by fixed parameters in traditional models.

[0158] The S8 carbon deposition state prediction employs a staged multi-scale fusion algorithm, with the evolved model from S7 as its core. It dynamically adjusts the feature extraction kernel size and loss function weights according to the pyrolysis stage—small kernels capture high-frequency fluctuations during the dechlorination stage, medium kernels balance details and the overall situation during the main pyrolysis stage, and large kernels capture trends during the high-temperature pyrolysis stage. Short-term prediction captures data fluctuations by improving the feature extraction model, while long-term prediction combines the S2 condition prediction results to correct the mechanism model output. This approach of "stage-adaptive feature extraction + long- and short-term prediction fusion" achieves accurate prediction of multiple objectives, including carbon deposition thickness, density, and growth rate. This provides quantitative evidence of carbon deposition state for S9 decision feedback, clarifying the timing and objectives of coke control strategies.

[0159] As the terminal link of the closed-loop architecture, the S9 decision feedback integrates the prediction results of S8 with six-dimensional indicators to construct a coupled early warning function. Combined with specific constraints such as the reactor coating's temperature tolerance, it divides the warning levels into five levels, generating a tiered coking control strategy. The enhanced monitoring of warning levels one to two, the adjustment of operating conditions for warning level three, and the gradient coking design for warning levels four to five avoid energy waste and product yield loss caused by excessive coking, while also preventing equipment risks caused by excessive carbon buildup. The strategy execution results are fed back to the S2 operating condition prediction module through correction coefficients, completing the entire closed loop of "prediction-decision-feedback-optimization," enabling the entire system to achieve continuous iterative adaptive capabilities.

[0160] In summary, each step progresses through a logical chain of "architectural criteria - predictive direction - weighted priority - calibration to improve quality - purification to ensure reliability - checksum mechanism - evolutionary model strengthening - predictive target - decision-making strategy." The technical means of each step are based on the output of the preceding steps and provide support for subsequent steps. Ultimately, through closed-loop iteration, three core technical effects are achieved: First, the accuracy of carbon deposition status judgment is improved through deep coupling of mechanism and data; second, the system's adaptability to raw material fluctuations and reaction mutations is enhanced through dynamic adaptation to stages and operating conditions; and third, the balance between the economy and reliability of carbon deposition control is achieved through a step-by-step coking control strategy and closed-loop optimization, providing full-chain technical support for the stable operation of the pyrolysis reactor.

[0161] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for early warning and decoking control of carbon buildup in a pyrolysis reactor based on multi-source data fusion, characterized in that, Includes the following steps: Real-time data collection of temperature, HCl concentration, carbon deposit thickness, carbon deposit density, and raw material composition within the pyrolysis reactor; Based on the temperature, HCl concentration and raw material composition data, the current pyrolysis reaction stage is identified, and the pyrolysis reaction stage includes at least a dechlorination stage and a main pyrolysis stage. Based on the identified pyrolysis reaction stage, a preset feature weight allocation strategy corresponding to that stage is invoked to perform weighted fusion of the collected multi-data to obtain weighted feature data; The weighted feature data is input into the digital twin model to obtain the predicted value of the carbon deposition state; the digital twin model integrates the carbon deposition growth kinetic equation and the reactor heat transfer equation. Based on the predicted carbon buildup status, generate an early warning signal or a decoking control command; Based on the execution result of the desiccant removal control command, the parameters of the feature weight allocation strategy or the digital twin model are corrected. The determination of the feature weight allocation strategy includes: A four-dimensional weighted evaluation matrix is ​​constructed, wherein the four dimensions include: correlation degree of pyrolysis mechanism, similarity of fingerprint spectrum, bidirectional calibration deviation of sensor, and adaptability of carbon deposition physical properties; The dynamic correction factor is calculated based on the real-time carbon deposit density and thermal conductivity to correct the basic weights; Set constraints to ensure that the total weight of features associated with the core pyrolysis mechanism is not lower than a preset threshold. Based on the current pyrolysis reaction stage, different nonlinear correction functions are selected to adjust the corrected weights in stages. The correction function used in the dechlorination stage enhances the characteristics with high correlation to the mechanism, while the correction function used in the main pyrolysis stage enhances the characteristics with high correlation to carbon deposition.

2. The method according to claim 1, characterized in that, The identification of the current pyrolysis reaction stage includes: The Long Short-Term Memory (LSTM) network was used as the time-series processing unit, and the input was a time-series feature vector containing temperature, HCl concentration, and raw material components. The pyrolysis reaction stage identifier is embedded in the input layer of the LSTM in the form of a unique thermal encoding, and the pyrolysis reaction stage identifier is determined based on the real-time temperature and HCl concentration threshold. The activating factors of PVC dechlorination mechanism and hydrogen radical mechanism were calculated, and the activating factors of PVC dechlorination mechanism were calculated based on the dechlorination reaction kinetic equation and real-time HCl concentration. A data quality score is introduced, which is calculated based on data integrity, noise level, and consistency with dechlorination reaction kinetics; The gating attention weight matrix is ​​dynamically adjusted according to the pyrolysis reaction stage identifier to enhance the importance of the raw material component variation coefficient and the corresponding characteristics of high activation factors. The raw material component variation coefficient is pre-calculated based on the raw material component data. Outputs the type of extreme working condition, the predicted occurrence time, and the prediction confidence level that combines the working condition probability and data quality score.

3. The method according to claim 1, characterized in that, Before inputting the weighted feature data into the digital twin model, a data purification step is also included: Calculate the sensitivity derivatives of the carbon deposition growth rate to temperature, HCl concentration, and hydrogen radical concentration; The sensitivity derivative is incorporated into the calculation of feature information gain to construct a pyrolysis kinetic weighted model, thereby enhancing the input features; A multiphysics residual formula coupling heat transfer coefficient, carbon deposit thickness and dechlorination reaction rate was established, and the data residuals were calculated. Select the corresponding residual threshold according to the current pyrolysis reaction stage. If the data residual exceeds the threshold or meets the judgment condition of incomplete dechlorination, the data is marked as abnormal and removed and supplemented.

4. The method according to claim 1, characterized in that, After data acquisition and before weight allocation, a two-way sensor calibration step is also included: For HCl concentration sensors, the corrosion drift failure factor driven by cumulative HCl exposure and temperature was calculated; For carbon deposit thickness sensors, the thickness measurement error factor driven by carbon deposit density, adhesion coefficient and time is calculated; Based on the failure factors, the current pyrolysis reaction stage, and the physical characteristics of carbon deposits, calculate the adaptive sensor data correction coefficients. Using the calibrated data and combining the correlation weights of the pyrolysis mechanism, the heat transfer, reaction, and mass transfer coupling coefficients in the digital twin model are optimized in reverse using the gradient descent method.

5. The method according to claim 1, characterized in that, It also includes a mechanism consistency verification step: A pyrolysis fingerprint library containing different mixing ratios of PVC, PE and PP is constructed. The spectral features in the pyrolysis fingerprint library include standardized temperature, gas concentration, carbon deposition characteristics and encoded reaction stages. Based on the current pyrolysis reaction stage, different attention weights are assigned to temperature, gas concentration, carbon deposition characteristics, and reaction stage characteristics. A dynamic time warping algorithm that integrates the attention weights and mechanism bias constraints is used to calculate the similarity between real-time data and the benchmark fingerprint. When the incremental data reaches a scale threshold and differs significantly from the baseline spectrum distribution, the fingerprint spectrum library is updated under mechanistic constraints.

6. The method according to claim 1, characterized in that, The parameters of the digital twin model are self-evolving, including: The Arrhenius equation for carbon deposition growth is embedded as a mechanistic constraint in the reinforcement learning reward function. The reward function is embedded with a penalty term that the carbon deposition density does not exceed the physical upper limit. When the coefficient of variation of the raw material components exceeds the first threshold, or the coefficient of variation of the carbon deposition growth rate exceeds the second threshold, the model evolution process is triggered. With the goal of maximizing cumulative reward, the gradient descent method driven by prediction error and multiphysics residual is used to update the carbon deposition growth kinetic parameters and multiphysics coupling coefficients in the digital twin model. The allowable boundary range of the kinetic parameters is dynamically adjusted according to the degree of fluctuation in the raw material composition.

7. The method according to claim 1, characterized in that, The process of obtaining the predicted carbon deposit state includes: Based on the current pyrolysis reaction stage, convolution kernels of different sizes are dynamically selected for multi-scale feature extraction. Small-sized convolution kernels are used in the dechlorination stage, medium-sized convolution kernels are used in the main pyrolysis stage, and large-sized convolution kernels are used in the high-temperature pyrolysis stage. Stage-adaptive weighted fusion of extracted multi-scale features; For short-term prediction, the fused features are input into a recurrent neural network for time-series prediction; For long-term forecasts, corrections are made by combining the baseline forecast value output by the digital twin model with the operating condition prediction results; Based on the current pyrolysis reaction stage, different weights are assigned to the prediction loss functions for carbon deposit thickness, density, and growth rate.

8. The method according to claim 1, characterized in that, The generation of the early warning signal or the clearing control command includes: A coupled early warning function is constructed, whose inputs include six standardized indicators: carbon deposition thickness, growth rate, fingerprint spectrum similarity, calibration deviation, carbon deposition characteristics, and data quality. Constraint penalty terms for reactor coating tolerance temperature and product yield loss are also introduced. Five warning levels are determined based on the value of the coupled warning function; For different warning levels, a tiered coking control strategy is generated, ranging from enhanced monitoring, operating condition adjustment, light coking removal to deep coking removal. The generation of the strategy is based on multi-objective optimization of coking removal effect, energy consumption and product yield. The measured value of the carbon deposit state after the strategy is implemented is compared with the predicted value, and the resulting correction coefficient is fed back into the model of the pyrolysis reaction stage identification step.

9. A multi-source data fusion early warning and decoking control system for carbon buildup in a pyrolysis reactor, characterized in that, include: A data acquisition sensor array resistant to HCl corrosion is used to collect data on temperature, HCl concentration, carbon deposit thickness, and density inside the reactor. An edge computing device is communicatively connected to the data acquisition sensor group. The edge computing device is equipped with an FPGA acceleration chip and a high-temperature heat dissipation unit, and is configured to perform the method as described in any one of claims 1 to 8. A cloud server, which is communicatively connected to the edge computing device, is used to store and update the pyrolysis fingerprint database and digital twin model parameters; The coke-clearing actuator is communicatively connected to the edge computing device and is used to receive and execute coke-clearing control commands.

Images