Anthracene oil processing automation control method and system based on machine learning

Abstract

The invention discloses an anthracene oil processing automatic control method and system based on machine learning, and the method comprises the steps: building a DRL deep reinforcement learning model, taking a process quality index and an energy consumption target as award functions, dynamically adjusting the distillation temperature gradient, the extraction agent ratio and the reaction time, obtaining an initial control instruction, and carrying out the automatic control of the anthracene oil processing. Performing real-time control based on the initial control instruction, collecting multi-modal data from a production device, preprocessing the multi-modal data, performing noise reduction processing by using wavelet transform, performing feature fusion on the data by using a CNN convolutional neural network and RNN recurrent neural network combined architecture, and extracting process state implicit features; inputting the process state implicit features into a model to predict product characteristics and potential risks at the next moment; and feeding back the product characteristics and the potential risk to the DRL deep reinforcement learning model, and producing a target control instruction. The automatic control efficiency can be improved.

Description

Technical Field

[0001] This invention relates to the field of coal chemical automation technology, and in particular to an anthracene oil processing automation control method and system based on machine learning. Background Technology

[0002] Anthracene oil is an important fraction in coal tar processing, with a complex composition containing high-value-added components such as anthracene, phenanthrene, and carbazole. The refining process of anthracene oil is a typical nonlinear, large-time-lag, and multivariate coupled complex industrial process. Its product quality is significantly affected by factors such as raw material fluctuations, operating condition changes, and equipment coking. Traditional PID controllers, with their fixed parameters, cannot adapt to slow time-varying characteristics such as raw material composition fluctuations and catalyst activity decay, leading to a decline in control quality. The production process has significant time lags, and traditional feedback control, being a reactive measure, cannot adjust operating parameters in advance to suppress disturbances. This results in low efficiency in the automated control of anthracene oil processing. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing an automated control method and system for anthracene oil processing based on machine learning.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned automated control method for anthracene oil processing based on machine learning: A deep reinforcement learning model (DRL) is constructed, with process quality indicators and energy consumption targets as reward functions. By combining simulated annealing algorithm with Q-learning strategy, the distillation temperature gradient, extractant ratio and reaction time are dynamically adjusted to obtain initial control commands. Real-time control is performed based on the initial control command. Multimodal data from the production unit is collected. After preprocessing the multimodal data, wavelet transform is used for noise reduction. The data is then fused using a joint architecture of CNN convolutional neural network and RNN recurrent neural network to extract the implicit features of the process state. The implicit features of the process state are input into the Transformer model to predict the product characteristics and potential risks at the next time step; The product characteristics and potential risks are fed back to the DRL deep reinforcement learning model to generate target control commands.

[0005] Furthermore, in the aforementioned machine learning-based automated control method for anthracene oil processing, the construction of a DRL deep reinforcement learning model uses process quality indicators and energy consumption targets as reward functions. By combining simulated annealing algorithm with a Q-learning strategy, the distillation temperature gradient, extractant ratio, and reaction time are dynamically adjusted to obtain initial control commands, including: Obtain core process quality indicators, including at least the purity, product yield, and impurity content of the target components such as anthracene, phenanthrene, and carbazole; The energy consumption control targets should include at least the steam consumption in the distillation process, the solvent circulation energy consumption in the extraction process, and the heating power of the reaction vessel. Define the range of controllable process parameters, including distillation temperature gradient, extractant ratio, and reaction time.

[0006] Furthermore, in the aforementioned machine learning-based automated control method for anthracene oil processing, the construction of a DRL deep reinforcement learning model uses process quality indicators and energy consumption targets as reward functions. By combining simulated annealing algorithm with a Q-learning strategy, the distillation temperature gradient, extractant ratio, and reaction time are dynamically adjusted to obtain initial control commands, including: Define the state space, including current process parameters, real-time quality indicators, and real-time energy consumption data; define the action space, including the adjustment amount of distillation temperature gradient, the adjustment amount of extractant ratio, and the adjustment amount of reaction time. Construct a reward function: add 10 points for every 1% improvement in quality indicators and deduct 15 points for every 1% improvement below the target; add 8 points for every 5% improvement in energy consumption and deduct 12 points for every 5% improvement above the target; and use the overall score as the immediate reward. An initial Q-table is constructed based on historical process data to determine the initial optimal action in each state. During the Q-learning iteration process, high-value actions are prioritized, while low-value actions are accepted with a certain probability. The acceptance probability is gradually reduced as the number of iterations increases.

[0007] Furthermore, in the aforementioned machine learning-based automated control method for anthracene oil processing, the real-time control based on the initial control command involves collecting multimodal data from the production unit, preprocessing the multimodal data, performing noise reduction using wavelet transform, and using a joint architecture of CNN convolutional neural network and RNN recurrent neural network to perform feature fusion on the data and extract implicit features of the process state, including: Drive the production unit according to the initial control command, and collect multimodal data from the production unit, including at least physical sensor data, chemical detection data and equipment status data; Sensor outliers in multimodal data are removed, missing values ​​are filled in by interpolation, and data with different sampling frequencies are unified into a 1Hz time series after unifying the scale of data with different dimensions. Based on the db4 wavelet, the preprocessed data is decomposed into 3-5 levels of wavelet decomposition to separate high-frequency noise components from low-frequency effective signals, thus obtaining processed multimodal data.

[0008] Furthermore, in the aforementioned machine learning-based automated control method for anthracene oil processing, the real-time control based on the initial control command involves collecting multimodal data from the production unit, preprocessing the multimodal data, performing noise reduction using wavelet transform, and using a joint architecture of CNN convolutional neural network and RNN recurrent neural network to perform feature fusion on the data and extract implicit features of the process state, including: Near-infrared spectral data is converted into a wavelength intensity matrix using a CNN convolutional neural network, and spatial features of spectral peak positions and peak values ​​are extracted through two layers of convolution and pooling. The time-series data of temperature and pressure are input into the LSTM network to capture the trend of parameter changes over time and obtain time features; Spatial and temporal features are concatenated through a fully connected layer and then processed by an activation function to extract the hidden features of the process state.

[0009] Furthermore, in the aforementioned machine learning-based automated control method for anthracene oil processing, the step of inputting the implicit features of the process state into the Transformer model to predict the product characteristics and potential risks at the next moment includes: The implicit features of the process state are arranged in time series and added with position encoding as input to the Transformer encoder. The encoder captures the long-range dependencies between features through a 6-layer self-attention mechanism, and the decoder outputs the product characteristics and potential risks at the next moment.

[0010] Furthermore, in the aforementioned machine learning-based automated control method for anthracene oil processing, the step of feeding back the product characteristics and potential risks to the DRL deep reinforcement learning model to generate target control commands includes: By inputting product characteristics and potential risks into the DRL model as new states, updating the state space, recalculating reward values ​​based on the new states, updating the Q-table through Q-learning, and adjusting action selection probabilities using simulated annealing strategies, the target control command is obtained.

[0011] Furthermore, in the machine learning-based automated control system for anthracene oil processing, the automated control system for anthracene oil processing includes the following modules: The control command generation module is used to build a DRL deep reinforcement learning model. It uses process quality indicators and energy consumption targets as reward functions, and dynamically adjusts the distillation temperature gradient, extractant ratio and reaction time by combining simulated annealing algorithm with Q-learning strategy to obtain initial control commands. The hidden feature extraction module is used to perform real-time control based on the initial control command, collect multimodal data from the production unit, preprocess the multimodal data and then use wavelet transform for noise reduction, and use a joint architecture of CNN convolutional neural network and RNN recurrent neural network to perform feature fusion on the data and extract the hidden features of the process state. The product characteristic prediction module is used to input the implicit features of the process state into the Transformer model to predict the product characteristics and potential risks at the next moment. The control command optimization module is used to feed back the product characteristics and potential risks to the DRL deep reinforcement learning model to generate target control commands.

[0012] Furthermore, in the machine learning-based automated control system for anthracene oil processing, the product characteristic prediction module includes the following sub-modules: The prediction submodule is used to arrange the implicit features of the process state in time series and add position encoding as input to the Transformer encoder. The encoder captures the long-range dependencies between features through a 6-layer self-attention mechanism, and the decoder outputs the product characteristics and potential risks at the next time step.

[0013] Furthermore, in the machine learning-based automated control system for anthracene oil processing, the control command optimization module includes the following sub-modules: The optimization submodule is used to input the product characteristics and potential risks as new states into the DRL model, update the state space, recalculate the reward value based on the new state, update the Q-table through Q-learning, and adjust the action selection probability in combination with the simulated annealing strategy to obtain the target control command.

[0014] Its beneficial effects lie in combining traditional time-series data with unstructured data such as visual information through multimodal deep learning, achieving a more comprehensive and accurate perception of the production status, which is impossible with traditional methods. Based on a high-precision digital twin predictor, this invention achieves a leap from post-event feedback to pre-event prediction and optimization. It can adjust operations in advance, smoothly respond to disturbances, and significantly reduce product quality fluctuations. The system can automatically adapt to slow time-varying disturbances such as raw material changes and catalyst deactivation through an online learning mechanism, enabling the control system to have autonomous evolution capabilities and reducing manual maintenance costs. Deep neural networks are naturally adept at handling nonlinear and multivariate coupled problems, and this characteristic is used to achieve effective decoupling and collaborative optimization control of the complex process of anthracene oil processing. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of the automated control method for anthracene oil processing based on machine learning in this invention. Figure 2 This is a schematic diagram of the second embodiment of the automated control method for anthracene oil processing based on machine learning in this invention. Figure 3 This is a schematic diagram of the first embodiment of the automated control system for anthracene oil processing based on machine learning in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, a machine learning-based automated control method for anthracene oil processing includes the following steps: Step 101: Construct a DRL deep reinforcement learning model, using process quality indicators and energy consumption targets as reward functions. By combining simulated annealing algorithm with Q-learning strategy, the distillation temperature gradient, extractant ratio and reaction time are dynamically adjusted to obtain initial control commands. Specifically, this embodiment obtains core process quality indicators, including at least the purity, product yield, and impurity content of the target components anthracene, phenanthrene, and carbazole; obtains energy consumption control targets, including at least steam consumption in the distillation stage, solvent circulation energy consumption in the extraction stage, and reactor heating power; defines the range of controllable process parameters, including distillation temperature gradient, extractant ratio, and reaction time. A state space is defined, including current process parameters, real-time quality indicators, and real-time energy consumption data; an action space is defined, including adjustments to the distillation temperature gradient, extractant ratio, and reaction time; a reward function is constructed, adding 10 points for every 1% exceeding the target quality indicator and deducting 15 points for every 1% below; adding 8 points for every 5% below the target energy consumption and deducting 12 points for every 5% exceeding; the comprehensive score serves as the immediate reward; an initial Q-table is constructed based on historical process data to determine the initial optimal action for each state. During the Q-learning iteration process, high-value actions are prioritized, while low-value actions are accepted with a certain probability, gradually decreasing the acceptance probability as the number of iterations increases.

[0020] In constructing a deep reinforcement learning (DRL) model to generate initial control commands for anthracene oil processing, the primary task is to clarify the control objectives and parameter boundaries, and to build the core architecture of the model. From the perspective of the process's essence, core quality indicators must accurately target the key product characteristics of anthracene oil processing: anthracene purity must be ≥95% to meet the needs of high-end applications such as dye intermediates; the purity of associated components such as phenanthrene and carbazole must match downstream chemical raw material standards; the total yield of target components ≥80% is directly related to production economics; and the content of impurities such as quinoline insolubles ≤0.5% is the bottom line to ensure product stability. Energy consumption control must consider the characteristics of each stage; steam consumption ≤5t / h in the distillation stage can avoid heat waste; solvent recycling energy consumption ≤200kW・h / t in the extraction stage can reduce solvent recovery costs; and the heating power of the reaction vessel ≤150kW balances heating efficiency and equipment load. The definition of controllable parameter ranges needs to be combined with the material properties and equipment capabilities: the distillation temperature gradient is divided into the initial distillation section of 180-220℃ and the rectification section of 220-280℃, which not only matches the separation requirements of anthracene (boiling point 340℃) and phenanthrene (boiling point 340℃), but also allows for fine adjustment through gradient intervals of 5-10℃; the extractant ratio, with N-methylpyrrolidone to anthracene oil mass ratio of 1:1 to 3:1, needs to balance extraction efficiency and solvent cost; the reaction time of 30-90 min ensures that the components are fully separated and avoids over-reaction. Based on this, the DRL model architecture needs to accurately map the process state: the state space integrates real-time operating parameters such as temperature, ratio, and time, online detection of component purity and other quality data, and energy consumption values ​​collected by sensors to comprehensively reflect the current operating conditions of the system; the action space sets adjustment ranges of ±1-5℃ for temperature, ±0.1-0.5 for ratio, and ±5-10min for time, ensuring both adjustment flexibility and avoiding drastic system fluctuations; the reward function follows the principle of "quality first, energy consumption optimization", adding 10 points for every 1% exceeding the target in quality indicators and deducting 15 points for every 1% below, strengthening the rigid constraints on quality; adding 8 points for every 5% below the target in energy consumption and deducting 12 points for every 5% exceeding, incentivizing the exploration of energy-saving potential, and guiding the model to learn towards the optimal equilibrium state through comprehensive scoring.

[0021] To improve the robustness and optimization efficiency of the model strategy, simulated annealing and Q-learning strategies need to be integrated, and initial control commands should be generated based on raw material characteristics. During the initialization phase of the Q-learning strategy, it is necessary to deeply mine stable production records from the past three months, extract the "state-action-result" correlation patterns under different raw material conditions, and construct an initial Q-table and a state-action value table to provide reliable prior experience for the model and reduce ineffective exploration. The introduction of the simulated annealing algorithm specifically addresses the local optimum problem in complex processes: in the Q-learning iteration, low-Q-value actions are initially accepted with a higher probability, such as attempting to deviate from the conventional temperature gradient, helping the model break free from the limitations of historical experience and explore potential better solutions; as the iteration progresses, the acceptance probability is gradually reduced, simulating the "cooling" process, allowing the model to focus on high-value actions and achieve a dynamic balance between "exploration and utilization," which is particularly important for anthracene oil processing, as fluctuations in raw material composition may cause historically optimal strategies to fail. When generating the initial control commands, the initial characteristics of the anthracene oil feedstock must be input. Density and viscosity reflect flowability, and the initial component content indicates the separation difficulty. Through 1000-5000 iterations of training, the model adapts to different feedstock conditions: low-density, low-viscosity feedstocks may require a slightly higher distillation temperature to accelerate separation, while feedstocks with high initial anthracene content can have their extractant ratio appropriately reduced to save costs. After training, the output initial commands must specify the temperature settings for each distillation stage, such as 260℃ at the bottom and 220℃ at the top, a precise extractant ratio (e.g., 1.8:1), and the reactor stirring time (e.g., 65 minutes). This provides a baseline for subsequent real-time control while reserving dynamic adjustment space to address uncertainties in production.

[0022] Step 102: Perform real-time control based on the initial control command, collect multimodal data from the production unit, preprocess the multimodal data and then use wavelet transform for noise reduction, use a joint architecture of CNN convolutional neural network and RNN recurrent neural network to perform feature fusion on the data and extract the implicit features of the process state. Specifically, in this embodiment, the production unit is driven according to the initial control command, and multimodal data from the production unit is collected, including at least physical sensor data, chemical detection data, and equipment status data. Sensor outliers in the multimodal data are removed, missing values ​​are filled using interpolation, and data from different sampling frequencies are unified into a 1Hz time series after standardizing the scale of data with different dimensions. Based on the db4 wavelet, the preprocessed data is decomposed into 3-5 layers of wavelet decomposition to separate high-frequency noise components from low-frequency effective signals, obtaining processed multimodal data. Near-infrared spectral data is converted into a wavelength intensity matrix using a CNN convolutional neural network, and spatial features of spectral peak positions and peak values ​​are extracted through two layers of convolution and pooling. Temporal data of temperature and pressure are input into an LSTM network to capture the trend of parameter changes over time, obtaining temporal features. Spatial and temporal features are concatenated through a fully connected layer and processed by an activation function to extract implicit features of the process state.

[0023] After starting the anthracene oil processing unit based on the initial control commands, real-time data acquisition and preprocessing form the basis for process status perception. Multi-dimensional data capture and systematic processing are necessary to ensure the integrity and reliability of the information. During operation, multi-modal data acquisition must cover three dimensions: physical, chemical, and equipment status. At the physical level, the temperature of each section of the distillation column is acquired via thermocouples at a frequency of 1Hz, accurately capturing subtle fluctuations in the temperature gradient, such as a temperature change of ±2℃ within one minute. The pressure inside the column is recorded by a pressure transmitter at 0.5Hz, balancing the amount of data with the gradual pressure change. The stirring speed of the extraction vessel is monitored by an encoder at 1Hz, reflecting the mixing intensity in real time. Material flow rate is acquired by an electromagnetic flowmeter at 1Hz, ensuring feed / discharge stability. At the chemical level, online near-infrared spectroscopy is used to detect product components at a frequency of 0.1Hz. Although the sampling frequency is low, purity can be accurately inverted through characteristic peaks, such as the characteristic absorption of anthracene at 1690nm. The pH value is monitored at 0.5Hz to monitor the acidity and alkalinity of the extraction phase, preventing extractant failure due to acid-base imbalance. In the equipment status data, the motor current acquisition at 1Hz can detect load abnormalities in a timely manner. For example, a sudden increase in current may indicate that the agitator is stuck. The valve opening record at 0.5Hz reflects the execution accuracy of the flow regulation.

[0024] Data preprocessing needs to specifically address data quality issues in industrial environments: During the cleaning phase, obvious outliers such as temperature spikes exceeding 20°C or negative flow rates, often caused by sensor malfunctions or electromagnetic interference, are directly discarded. Missing values ​​with a single duration of ≤5 seconds are filled using linear interpolation. For example, if a temperature sensor is temporarily offline, the intermediate value can be fitted based on stable data from the preceding and following 10 seconds to ensure temporal continuity. In the standardization phase, temperatures (180-280°C) are normalized to the 0-1 range to eliminate absolute numerical differences. Fluctuating data such as flow rates are standardized to a mean of 0 and a variance of 1, allowing parameters with different dimensions, such as temperature and current, to work synergistically in the model. Data alignment is achieved through time axis calibration. Low-frequency data such as pressure and pH (0.5Hz) are interpolated to 1Hz, and high-frequency redundant data are averaged, ultimately forming a time series with a unified time scale, laying a consistent foundation for subsequent analysis.

[0025] The preprocessed data requires further noise reduction and feature fusion to extract deeper information that truly reflects the process status. Wavelet transform denoising addresses the non-stationary nature of industrial data, selecting the db4 wavelet due to its excellent ability to capture abrupt changes. A 3-5 level decomposition is performed: lower-level decomposition separates high-frequency noise from sensor mechanical vibrations, such as random fluctuations of ±0.5℃ in temperature signals; higher-level decomposition preserves the low-frequency trends of the process itself, such as the slow increase in distillation temperature with increasing feed rate. High-frequency coefficients are processed using a soft thresholding method, with the threshold adaptively adjusted based on noise energy. For example, the threshold is increased when the noise variance exceeds 10%. The reconstructed data can remove over 90% of random noise while fully preserving key process inflection points, such as the anomalous signal of a sudden 0.3 drop in the pH value of the extraction phase. In the feature fusion stage, a joint CNN and RNN architecture achieves deep integration of multi-dimensional information: The CNN converts the one-dimensional wavelength-intensity sequence of the near-infrared spectrum into a two-dimensional matrix, and extracts spatial features such as peak position and half-peak width of characteristic peaks at 1690nm (anthracene) and 1750nm (phenanthrene) through two layers of 3×3 convolutional kernels, accurately reflecting component differences; the RNN uses a two-layer LSTM network with 64 hidden units to model time-series data such as temperature and pressure, capturing trend features such as "the temperature at the top of the distillation column increases by 1°C every 5 minutes," and relating them to the dynamic changes in the reaction process. Finally, the two types of features are spliced ​​together through a fully connected layer, and after ReLU activation, a 128-dimensional process state implicit feature is generated. This feature includes the influence of raw material density fluctuations, integrates signals of equipment load changes, and also implicitly contains stage information of the reaction process, providing high-value input for subsequent product characteristic prediction and risk warning.

[0026] Step 103: Input the implicit features of the process state into the Transformer model to predict the product characteristics and potential risks at the next time step; Specifically, in this embodiment, the implicit features of the process state are arranged in time sequence and position encoding is added as the input of the Transformer encoder. The encoder captures the long-range dependencies between features through a 6-layer self-attention mechanism, and the decoder outputs the product characteristics and potential risks at the next moment.

[0027] When predicting product characteristics and risks based on the Transformer model, the scientific nature of the input construction is the core prerequisite for ensuring prediction accuracy, while the synergistic effect of the encoder and decoder enables accurate prediction of product characteristics. Considering the 1-5 minute lag in the influence of process parameters on product characteristics during anthracene oil processing, a 10×128 time series matrix needs to be constructed using 128-dimensional hidden features from the past 10 minutes. This time span covers key reaction cycles, such as the 5-8 minute mixing reaction of the extractant and anthracene oil, while avoiding data redundancy that could reduce model computational efficiency. To avoid the model confusing the temporal order of features, such as the drastically different effects on product purity caused by "heating before adding the extractant" versus "adding the extractant before heating," sine-cosine positional encoding needs to be added to the matrix. Different frequency encoding values ​​mark the temporal position of features at each moment, enabling the model to accurately identify the causal relationship of "temperature gradient increase preceding component change." In the product characteristic prediction stage, a 6-layer self-attention encoder with 8 attention heads plays a crucial role. Each attention head can focus on a type of key feature; for example, head 1 focuses on temperature time series, and head 2 focuses on spectral features. By calculating the attention weights between features, long-range dependencies are captured. For instance, when the temperature in the middle section of the distillation column rises from 240℃ to 250℃, the model can identify a strong correlation between this change and the increase in anthracene purity from 94% to 95.5% three minutes later. The decoder, based on the global features output by the encoder, combines a greedy decoding strategy to generate the prediction results for the next time step, one minute later. This prediction not only covers the purity of target components such as anthracene and phenanthrene (prediction error ≤1.2%), but also includes key indicators such as product yield (error ≤3%) and density (error ≤0.02 g / cm³), providing a quantitative basis for subsequent process adjustments.

[0028] The self-attention mechanism of the Transformer model is particularly advantageous in predicting potential risks, enabling early warnings of various hidden dangers by accurately locating risk-related features. Risks in anthracene oil processing are often caused by the coordinated anomalies of multiple parameters. For example, coking in a distillation column originates from local temperatures exceeding 280°C and pressures above 0.15 MPa for more than 2 minutes. Traditional single-indicator warnings are prone to misjudgment, while the Transformer can lock in core related features through attention weight allocation. For instance, when a sudden temperature rise of 15°C is detected and coking characteristic peaks appear in the near-infrared spectrum, with enhanced absorption at 1800 nm, the model will allocate high attention weights to temperature time-series features and spectral features, combining historical coking case data to calculate the risk probability. When the probability is ≥70%, a coking warning is triggered. At this time, the control system can reduce the heating power in advance to prevent coking from adhering to the column wall. To address the risk of extractant emulsification, the model focuses on the synergistic characteristics of pH fluctuations, deviations from the 6.5-7.5 range, and abnormal stirring speeds. It uses attention weights to pinpoint the temporal correlation between these two factors. For example, if a sudden drop in pH of 0.8 is followed by a sudden increase in stirring current of 10A, a demulsifier addition warning is triggered when the risk probability is ≥60%. For the risk of substandard product purity, the model integrates multiple features such as raw material component fluctuations and temperature gradient deviations. When the predicted probability of purity below 95% is ≥50%, an early warning is issued to adjust the extractant ratio. This risk prediction mode based on multi-feature correlation increases the warning lead time from 30 seconds to 2 minutes compared to traditional methods, significantly improving process stability.

[0029] Step 104: Feed the product characteristics and potential risks back to the DRL deep reinforcement learning model to generate target control instructions.

[0030] Specifically, in this embodiment, the product characteristics and potential risks are used as new state inputs to the DRL model to update the state space. The reward value is recalculated based on the new state, the Q-table is updated through Q-learning, and the action selection probability is adjusted by combining the simulated annealing strategy to obtain the target control command.

[0031] In the feedback optimization phase, the accurate input of the prediction results and the dynamic strategy adjustment of the DRL model are the core of achieving closed-loop optimization of the process. It is necessary to use the logical chain of "future state perception - reward mechanism adaptation - action probability adjustment" to enable the model to address the problems exposed by the prediction. The prediction data output in step 3 is not a simple input, but needs to be integrated into a "future-oriented state space" that the DRL model can interpret: in addition to retaining basic operating parameters such as real-time temperature and extractant ratio, a quality deviation index of "predicted anthracene purity 92%, lower than the 95% target value" and a risk level parameter of "coking risk 80%, exceeding the 70% warning line" are added to form a two-dimensional state description of "current operating conditions + future hidden dangers". This design allows the DRL model to break away from the limitation of "adjusting only based on current data" and formulate strategies in advance for potential problems one minute later.

[0032] The strategy adjustment phase requires reconstructing and guiding model behavior through a reward mechanism: First, the reward value is recalculated based on the new state, using the three-dimensional evaluation system of "quality-energy consumption-risk". In the quality dimension, 15 points are deducted for every 1% below the target. For example, if the purity is 92% lower than the target by 3%, 45 points are deducted directly. In the risk dimension, a new "deduction rule for risk probability exceeding the warning line" is added. For example, if the coking risk is 80% higher than the 70% warning line by 10%, an additional 20 points are deducted. If the energy consumption prediction is normal, such as steam consumption of 4.8t / h ≤ 5t / h target, no additional points are deducted. The final comprehensive reward value is significantly lower than the target threshold, forcing the model to adjust its actions. Subsequently, the Q-table was updated using Q-learning. For the state of "current temperature gradient 260℃+ coking risk 80%", the Q value of the action "reducing distillation temperature" was increased by 30%, while the Q value of actions unrelated to risk reduction, such as "increasing the extractant ratio", was decreased. When combined with the simulated annealing strategy, the iteration was in the "medium temperature stage". The acceptance probability of low-value actions, such as continuing to increase the temperature, was reduced from 20% to 5%. High-value actions such as "reducing the temperature of the rectification section by 3℃" and "reducing the heating power" were prioritized to ensure that the strategy adjustment directly targets purity improvement and risk control.

[0033] After strategy adjustments, the generation of target control instructions must balance optimization effectiveness and equipment safety. Through an iterative optimization-feasibility verification-closed-loop execution process, precise and actionable instructions are output. Considering the thermal inertia in anthracene oil processing, such as the 3-5 minute stabilization and mass transfer delay required after temperature adjustment in the distillation column, and the 2-minute impact on component separation caused by changes in the extractant ratio, the model needs 500-1000 iterations: the first 500 iterations focus on risk management, observing changes in coking risk by making small temperature adjustments, such as decreasing the temperature by 1°C each time. When the risk drops below 60%, the next 500 iterations shift to purity optimization, gradually increasing the extractant ratio by 0.1% each time and extending the reaction time by 2 minutes each time, ultimately finding a balance between "controllable risk and quality compliance".

[0034] The output instructions must include specific executable correction parameters: To address the risk of coking, the rectification section temperature is lowered by 3°C from 260°C to 257°C, thus avoiding the coking-sensitive range above 280°C without affecting the separation efficiency of anthracene; to address insufficient purity, the extractant ratio is increased from 1.8:1 to 2.0:1 to enhance the solubility of N-methylpyrrolidone in anthracene and improve extraction efficiency; to address incomplete separation, the reaction time is extended from 65 min to 70 min to make the mass transfer process between anthracene and phenanthrene and carbazole more thorough. After the instruction is generated, it must pass three layers of feasibility verification: First, equipment safety threshold verification, ensuring that the distillation temperature is 257℃≤300℃, the maximum withstand temperature of the heating element, the extractant ratio is 2.0:1≤3:1, and the maximum supply capacity of the solvent storage tank; second, process logic verification, confirming that the combination of "lowering temperature + increasing ratio + extending time" will not cause new problems, such as extending the time will not lead to energy consumption exceeding the target; third, execution accuracy verification, converting the temperature adjustment into a 4-20mA current signal that the PLC can recognize, corresponding to a reduction in heating power from 140kW to 125kW, ensuring that the equipment can respond accurately. After the final instruction is issued to the PLC control system, the execution time and parameter baseline values ​​are recorded simultaneously, providing initial data for the next round of "data acquisition-feature extraction-prediction feedback", forming a closed loop of "prediction-optimization-execution-re-prediction", allowing anthracene oil processing to remain stable within the optimal range of "anthracene purity ≥95% + coking risk ≤60% + energy consumption ≤5t / h". Its beneficial effects lie in combining traditional time-series data with unstructured data such as visual information through multimodal deep learning, achieving a more comprehensive and accurate perception of the production status, which is impossible with traditional methods. Based on a high-precision digital twin predictor, this invention achieves a leap from post-event feedback to pre-event prediction and optimization. It can adjust operations in advance, smoothly respond to disturbances, and significantly reduce product quality fluctuations. The system can automatically adapt to slow time-varying disturbances such as raw material changes and catalyst deactivation through an online learning mechanism, enabling the control system to have autonomous evolution capabilities, reducing manual maintenance costs, and improving automatic control efficiency. Deep neural networks are naturally adept at handling nonlinear and multivariate coupled problems, and this characteristic is used to achieve effective decoupling and collaborative optimization control of the complex process of anthracene oil processing.

[0035] Please see Figure 2 In the automated control method for anthracene oil processing based on machine learning, a deep reinforcement learning (DRL) model is constructed. Using process quality indicators and energy consumption targets as reward functions, and combining simulated annealing algorithm with a Q-learning strategy, the distillation temperature gradient, extractant ratio, and reaction time are dynamically adjusted to obtain the initial control commands. The process includes the following steps: Step 201: Define the state space, including the current process parameters, real-time quality indicators and real-time energy consumption data; define the action space, including the adjustment amount of the distillation temperature gradient, the adjustment amount of the extractant ratio and the adjustment amount of the reaction time. Step 202: Construct a reward function. For each 1% improvement in quality indicators above the target, add 10 points and deduct 15 points for each 1% improvement below the target. For each 5% improvement in energy consumption below the target, add 8 points and deduct 12 points for each 5% improvement above the target. The overall score is used as the immediate reward. Step 203: Construct an initial Q-table based on historical process data, determine the initial optimal action in each state, prioritize high-value actions during the Q-learning iteration process, and accept low-value actions with a certain probability, gradually reducing the acceptance probability as the number of iterations increases.

[0036] The above describes embodiments of the machine learning-based automated control method for anthracene oil processing according to the present invention. Please refer to [link / reference]. Figure 3 In the machine learning-based automated control system for anthracene oil processing, the automated control system for anthracene oil processing includes the following modules: The control command generation module is used to build a DRL deep reinforcement learning model. It uses process quality indicators and energy consumption targets as reward functions, and dynamically adjusts the distillation temperature gradient, extractant ratio and reaction time by combining simulated annealing algorithm with Q-learning strategy to obtain initial control commands. The hidden feature extraction module is used for real-time control based on initial control commands, collecting multimodal data from the production unit, preprocessing the multimodal data and then using wavelet transform for noise reduction, and using a joint architecture of CNN convolutional neural network and RNN recurrent neural network to perform feature fusion on the data and extract the hidden features of the process status. The product characteristic prediction module is used to input the implicit features of the process state into the Transformer model to predict the product characteristics and potential risks at the next moment. The control command optimization module is used to feed back product characteristics and potential risks to the DRL deep reinforcement learning model to generate target control commands.

[0037] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for anthracene oil processing automation control based on machine learning, characterized by, The anthracene oil processing automatic control method comprises the following steps: A DRL deep reinforcement learning model is constructed, process quality indicators and energy consumption targets are used as reward functions, simulated annealing algorithm and Q-learning strategy are combined, distillation temperature gradient, extractant ratio and reaction time are dynamically adjusted, and initial control instructions are obtained; Real-time control is performed based on the initial control instructions, multi-modal data from the production device is collected, the multi-modal data is preprocessed, then wavelet transform is used for noise reduction processing, CNN convolutional neural network and RNN recurrent neural network joint architecture are used for feature fusion of the data, and process state implicit features are extracted; The process state implicit features are input into a Transformer model to predict product characteristics and potential risks at the next time; The product characteristics and potential risks are fed back to the DRL deep reinforcement learning model, and production target control instructions are obtained.

2. The machine learning based automatic control method of anthracene oil processing according to claim 1, wherein, The DRL deep reinforcement learning model is constructed, process quality indicators and energy consumption targets are used as reward functions, simulated annealing algorithm and Q-learning strategy are combined, distillation temperature gradient, extractant ratio and reaction time are dynamically adjusted, and initial control instructions are obtained, comprising: Obtain core process quality indicators, including at least the purity of anthracene, phenanthrene and carbazole target components, product yield and impurity content; Obtain energy consumption control targets, including at least distillation link steam consumption, extraction link solvent circulation energy consumption and reaction kettle heating power; Define the controllable process parameter range, including distillation temperature gradient, extractant ratio and reaction time.

3. The machine learning based automatic control method of anthracene oil processing according to claim 1, wherein, The DRL deep reinforcement learning model is constructed, process quality indicators and energy consumption targets are used as reward functions, simulated annealing algorithm and Q-learning strategy are combined, distillation temperature gradient, extractant ratio and reaction time are dynamically adjusted, and initial control instructions are obtained, comprising: Define the state space, including the current process parameters, real-time quality indicators and real-time energy consumption data, define the action space, including the adjustment amount of distillation temperature gradient, the adjustment amount of extractant ratio and the adjustment amount of reaction time; Construct a reward function, add 10 points for each 1% over target for quality indicators, deduct 15 points for each 1% below target, add 8 points for each 5% below target for energy consumption, deduct 12 points for each 5% over target, and the comprehensive score is used as the immediate reward; Based on the historical process data, an initial Q table is constructed to determine the initial optimal action under each state, in the Q-learning iteration process, high-value actions are preferentially selected, and low-value actions are accepted with a certain probability, and the acceptance probability is gradually reduced with the increase of the iteration number.

4. The machine learning based automatic control method of anthracene oil processing according to claim 1, wherein, Based on the initial control instructions, real-time control is performed, multi-modal data from the production device is collected, the multi-modal data is preprocessed, then wavelet transform is used for noise reduction processing, CNN convolutional neural network and RNN recurrent neural network joint architecture are used for feature fusion of the data, and process state implicit features are extracted, comprising: Drive the production device according to the initial control instructions, collect multi-modal data from the production device, including at least physical sensor data, chemical detection data and equipment state data; The sensor outliers in the multi-modal data are removed, the missing values are filled by interpolation, the data of different dimensions are unified in scale, and the data of different sampling frequencies are unified into 1Hz time sequence; The preprocessed data is decomposed by 3-5 layers of wavelet based on db4 wavelet, the high-frequency noise component is separated from the low-frequency effective signal, and the processed multi-modal data is obtained.

5. The machine learning based automatic control method of anthracene oil processing as claimed in claim 1, wherein, The real-time control based on the initial control instruction is performed, multi-modal data from the production device is collected, the multi-modal data is preprocessed, and wavelet transform is used for noise reduction processing, CNN convolutional neural network and RNN recurrent neural network joint architecture are used for feature fusion of the data, and process state implicit features are extracted, including: The near-infrared spectrum data is converted into a wavelength intensity matrix by using the CNN convolutional neural network, and the spatial features of the spectral peak position and peak value are extracted by 2 layers of convolution and pooling; The time sequence data of temperature and pressure are input into the LSTM network to capture the change trend of the parameters over time, and the time features are obtained; The spatial features and time features are spliced through the full connection layer, and the process state implicit features are extracted after the activation function processing.

6. The machine learning based automatic control method of anthracene oil processing according to claim 1, wherein, The process state implicit features are input into the Transformer model to predict the product characteristics and potential risks at the next moment, including: The process state implicit features are arranged in time sequence, and the position encoding is added as the input of the Transformer encoder, the encoder captures the long-range dependence between features through 6 layers of self-attention mechanism, and the decoder outputs the product characteristics and potential risks at the next moment.

7. The machine learning based automatic control method of anthracene oil processing according to claim 1, wherein, The product characteristics and potential risks are fed back to the DRL deep reinforcement learning model, and the target control instruction is generated, including: The product characteristics and potential risks are input into the DRL model as new states, the state space is updated, the reward value is recalculated based on the new state, the Q table is updated through Q-learning, the action selection probability is adjusted combined with the simulated annealing strategy, and the target control instruction is obtained.

8. A machine learning based anthracene oil processing automation control system characterized by, The anthracene oil processing automatic control system includes the following modules: The control instruction generation module is used to construct the DRL deep reinforcement learning model, take the process quality index and energy consumption target as the reward function, dynamically adjust the distillation temperature gradient, extractant ratio and reaction time combined with the simulated annealing algorithm and Q-learning strategy, and obtain the initial control instruction; The implicit feature extraction module is used to perform real-time control based on the initial control instruction, collect multi-modal data from the production device, pre-process the multi-modal data, and use wavelet transform for noise reduction processing, and use CNN convolutional neural network and RNN recurrent neural network joint architecture to fuse the features of the data and extract process state implicit features. The product characteristic prediction module is used to input the process state implicit features into the Transformer model to predict the product characteristics and potential risks at the next moment. The control instruction optimization module is used to feed back the product characteristics and potential risks to the DRL deep reinforcement learning model to generate a target control instruction.

9. The machine learning based automatic control system for anthracene oil processing according to claim 8, wherein, The product characteristic prediction module includes the following sub-modules: The prediction submodule is configured to arrange the process state implicit features in time sequence, add position encoding as input of a transformer encoder, and capture long-range dependence among the features through a 6-layer self-attention mechanism of the encoder, and output product properties and potential risks at the next moment through a decoder.

10. The machine learning based automatic control system for anthracene oil processing according to claim 8, wherein, The control instruction optimization module comprises the following submodules: The optimization submodule is configured to input the product properties and potential risks into a DRL model as new states, update a state space, recalculate a reward value based on the new states, update a Q table through Q-learning, adjust an action selection probability by combining a simulated annealing strategy, and obtain a target control instruction.

Images