Cracking furnace intelligent optimization system control method and system architecture

Abstract

The invention discloses a cracking furnace intelligent optimization system control method and system architecture, and belongs to the cracking furnace application field, and the cracking furnace intelligent optimization system control method comprises the following steps: S1, collecting cracking furnace operation time sequence characteristic data and flame combustion video data in real time through a sensor and a high temperature camera; s2, preprocessing the collected multi-modal data, S3, integrating the preprocessed multi-modal data by using federal learning, and constructing a cross-modal causal logic model; s4, designing a deep reinforcement learning framework, introducing the dynamic prediction model of the cracking furnace, and reasoning the recommendation range of each parameter of the optimized operation of the cracking furnace; s5, dynamic process data and the environment-friendly emission prediction model are combined, a data-driven MPC algorithm framework is adopted, and dynamic optimization control over the cracking furnace is achieved; and S6, the operator performs adjustment or selects emergency operation according to the recommended parameters and strategies to realize man-machine cooperative control. Finally, the dynamic optimization control of the cracking furnace can be realized, and the energy consumption is effectively reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of cracking furnace applications, in particular to a cracking furnace intelligent optimization system control method and system architecture. BACKGROUND

[0002] Cracking furnace is the core equipment of petrochemical, chemical and other industries. Its pain points mainly lie in insufficient energy management capability, leading to fuel waste; weak pollution emission treatment level, difficult to meet environmental protection standards, affecting the green image and international competitiveness of enterprises; and low intelligent level, difficult to cope with complex working conditions. The specific performance is as follows: (1) The fan scheduling relies on experience and lacks dynamic optimization, resulting in high annual fuel cost of a single cracking furnace; energy efficiency analysis is missing, and there is a lack of big data platform support, which cannot predict energy consumption trend or identify energy saving potential points.

[0003] (2) Low combustion efficiency, flame combustion state depends on manual visual judgment, and insufficient combustion phenomenon is frequent; parameter adjustment lags, affecting product quality consistency.

[0004] (3) Data island is serious, DCS system, video monitoring and environmental protection monitoring data are stored separately, and lack of fusion analysis; low control precision, traditional PID control algorithm is difficult to cope with complex working conditions; lack of prediction ability, unable to predict fuel and pressure changes in advance, leading to insufficient combustion, affecting the stability of cracking furnace temperature.

[0005] (4) The monitoring means is backward, and the treatment relies on manual adjustment, lacks combustion process optimization, the operation risk is big, the key parameter adjustment depends on skilled workers, and the personnel flow leads to the process stability decline.

[0006] In view of the above problems, the present application designs and manufactures a cracking furnace intelligent optimization system control method and system architecture to overcome the above defects. SUMMARY

[0007] For the problems existing in the prior art, the cracking furnace intelligent optimization system control method and system architecture provided by the present application can realize dynamic optimization control of the cracking furnace, effectively reduce energy consumption and pollution emission, and improve fuel utilization.

[0008] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows: a cracking furnace intelligent optimization system control method, comprising the following steps: S1. Real-time acquisition of time sequence characteristic data and flame combustion video data of cracking furnace operation through sensors and high temperature cameras deployed at key nodes of cracking furnace; S2. Preprocessing the collected multi-modal data, including reading, sampling, cleaning, and dimensionality reduction of the time-series feature data, and using deep learning to extract flame state features from the video data to identify the flame burning state and convert it into time-series feature data; S3. Integrating the preprocessed multi-modal data using federated learning, combining the relevance and complementarity of the time-series feature data and the flame video data of the cracking furnace, designing a loss function or regularization term for joint training, and constructing a cross-modal causal logic model; S4. Designing a deep reinforcement learning framework, combining the cross-modal causal logic model with a data-driven model, introducing a cracking furnace dynamic prediction model, accelerating the learning speed of the environment, and improving the estimation accuracy of the value function and actions, thereby learning the causal relationships between various factors in the cracking furnace operation process and reasoning out the recommended parameter range for the optimal operation of the cracking furnace; S5. Combining dynamic process data and environmental emission prediction models, using a data-driven MPC algorithm framework to calculate the adjustment amount of the fan frequency and fuel pressure parameters in real time, and achieving dynamic optimization control of the cracking furnace; S6. The operator adjusts or selects emergency operations according to the recommended parameters and strategies to achieve human-machine collaborative control.

[0009] Preferably, in S3, the flame video data is input as an image, and a deep neural network architecture with a shared CNN backbone and multi-task heads (SharedBackbone + Multi-Task Heads) is used to achieve multi-target joint learning of the flame burning state through a joint training mechanism.

[0010] Preferably, the loss is dynamically balanced by weighting different task losses, and the loss function is: where s is the flame state, i.e., the color of the flame, the brightness of the flame, and the dynamics of the flame, is the weight coefficient of the s-th state task, is the true label encoding of the s-th state, is the probability that the model predicts that the sample belongs to the s-th state, is the cross-entropy loss core term, which measures the difference between the predicted probability and the true label; c represents the classification of the flame, i.e., whether there is a flame, is the weight coefficient of the binary classification task, used to balance the importance of it and the flame state classification task, is the true label of the presence or absence of a flame, is the probability that the model predicts "there is a flame", is the binary classification cross-entropy loss.

[0011] Preferably, the data collected by the sensors and high-temperature cameras is transmitted to the central control system through Modbus or OPC protocol.

[0012] Preferably, the central control system can present multi-source data on a unified platform. And / or, it can store full-volume operation data and playback. And / or, it is provided with a safety threshold that automatically triggers an alarm when the parameters exceed the range.

[0013] Preferably, in S4, the simulation data output by the dynamic prediction model of the cracking furnace jointly drive the Critic and Actor network parameter updates together with the real multi-modal data.

[0014] Preferably, in S5, a parameterized model predictive control (PMPC) is introduced to reduce the computational complexity of model predictive control (MPC).

[0015] Preferably, the parameterized model predictive control (PMPC) is integrated with the reinforcement learning (RL) agent to improve the ability and flexibility to cope with variable or unknown environments and disturbances.

[0016] An intelligent optimization system architecture for a cracking furnace, which is composed of a perception layer, a network layer, a platform layer, and an application layer. The perception layer collects temperature, pressure, flow, material, and image data in real time during the production process through sensors and high-temperature cameras. The network layer is used for real-time detection of production site data transmitted to the data platform. The platform layer provides data storage and analysis, and performs data processing and machine learning modeling, dynamically adjusts the production process through data analysis models, and provides the best process parameters and operation suggestions for operators. The application layer provides various intelligent application services for users, and displays real-time data and optimization suggestions in the production process through a visual interface.

[0017] Preferably, the multi-source data collected and captured by the sensors and high-temperature cameras are saved in a MySQL database. The high-temperature camera stores the photos taken in Minio, the algorithm execution service reads the pictures collected by the high-temperature camera from Minio, performs flame state recognition, stores the processed photos in Minio again, and finally displays the real-time picture results after processing in the visual interface.

[0018] The advantages of the present application are: 1. The present application applies federated learning and causal reasoning to the multi-variable strong coupling system of a cracking furnace. By integrating real-time DCS time series data and flame combustion video features, the quantitative indicators provided by the DCS time series data, and the spatial and dynamic details supplemented by the flame combustion video data are used to construct a complete working condition cognition and a cross-modal causal logic model to accurately analyze the system response law under complex working conditions.

[0019] 2. The present application proposes a three-layer collaborative optimization architecture of "federated learning-model predictive control-reinforcement learning adaptive", generates key parameter optimization intervals based on dynamic data in real time, realizes second-level closed-loop control using a data-driven MPC algorithm, and combines a reinforcement learning continuous iteration control strategy to improve the stability of combustion efficiency.

[0020] 3. When the fuel pressure fluctuates and the oxygen content changes, the reinforcement learning (RL) agent automatically adjusts the parameters, so that the parameterized model predictive control (PMPC) always outputs the optimal control strategy, ensuring stable furnace temperature, efficient combustion, and avoiding frequent manual intervention.

[0021] 4. The present application combines machine intelligence and human experience, allowing operators to manually fine-tune recommended parameters or switch to manual control in emergency situations. This combination of "machine intelligence + human experience" avoids the risks of pure automatic control under extreme conditions and compensates for the efficiency shortcomings of manual operation. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 A multi-modal federated learning diagram of the present application; Figure 2 A model and data jointly driven deep reinforcement learning algorithm diagram of the present application; Figure 3 A model predictive control diagram based on data learning of the present application; Figure 4 A general flowchart of the present application. DETAILED DESCRIPTION

[0023] To facilitate understanding by those skilled in the art, the present application will be further described below with reference to the accompanying drawings.

[0024] As shown in Figure 1 , Figure 4 , a cracking furnace intelligent optimization system control method includes the following steps: S1. Through the sensors and high-temperature cameras deployed at the key nodes of the cracking furnace, real-time collection of time sequence characteristic data and flame combustion video data of the cracking furnace operation is performed, and the deployment positions of the specific sensors can be, for example, the furnace chamber, the fuel pipe network, the fan inlet, etc., and real-time collection of time sequence characteristic data such as temperature, pressure, oxygen content, fuel flow, etc. is performed, while the high-temperature camera shoots the flame combustion video to obtain video data such as flame shape, brightness, color, jitter frequency, etc. S2. Since the data is derived from different modalities, the collected multi-modal data needs to be pre-processed, including reading, sampling, cleaning and dimensionality reduction of the time sequence characteristic data, and the video data is relatively special and needs to be processed separately. Therefore, deep learning is used for flame state feature extraction on the video data, the flame combustion state is recognized and converted into time sequence characteristic data, and the time sequence characteristic data of the flame combustion state can describe the dynamic changes of the flame combustion state and be used to judge the sufficient combustion degree of the cracking furnace system. S3. The pre-processed multi-modal data is integrated by using federated learning, the correlation and complementarity of the cracking furnace time sequence characteristic data and the flame video data are combined, a loss function or a regularization term is designed for joint training, and a cross-modal causal logic model is constructed, that is, for example, the causal relationship between factors such as fuel composition changes, material fluctuations and cracking parameters is found out, and for example, the causal relationship between temperature data and flame video is found out, and the two together describe the cracking furnace working condition, for example, temperature fluctuations are highly correlated with flame brightness changes, and video data can assist in explaining the reasons for the abnormality of time sequence characteristic data. The present application applies federated learning and causal reasoning to the multi-variable strong coupling system of the cracking furnace, integrates the DCS time sequence data and the flame combustion video features in real time, specifically provides quantitative indicators through the DCS time sequence data, supplements the space and dynamic details through the flame combustion video data, and jointly constructs a complete working condition cognition, constructs a cross-modal causal logic model, and accurately analyzes the system response law under complex working conditions.

[0025] S4. A deep reinforcement learning framework is designed, the cross-modal causal logic model and the data-driven model are combined, a cracking furnace dynamic prediction model is introduced, the learning speed of the environment is accelerated, the estimation accuracy of the value function and the action is improved, the causal relationship between various factors in the cracking furnace operation process is learned, and the recommended range of each parameter for the optimal operation of the cracking furnace is reasoned out; S5. Combined with the dynamic process data and the environmental emission prediction model, a data-driven MPC algorithm framework is used, the adjustment amount of the fan frequency and the fuel pressure parameter is calculated in real time according to the current data and the future prediction, the combustion stability is ensured, and the dynamic optimization control of the cracking furnace is realized; S6. The operator adjusts or selects the emergency operation according to the recommended parameters and strategies, and realizes the man-machine collaborative control.

[0026] By proposing a three-layer collaborative optimization architecture of "federal learning-model predictive control-reinforcement learning adaptation", the key parameter optimization interval is generated in real time based on dynamic data, the data-driven MPC algorithm is used to realize the second-level closed-loop control, and the reinforcement learning continuous iteration control strategy is combined to improve the stability of the combustion efficiency.

[0027] The flame video data in S3 of the application takes an image as input, adopts a deep neural network architecture of shared CNN backbone and multi-task head (SharedBackbone + Multi-Task Heads), realizes multi-target joint learning of flame burning state through joint training mechanism, and aims to accurately detect the flame burning state.

[0028] The shared backbone network (Backbone) is the core of multi-target learning, adopts a pre-trained CNN (convolutional neural network) as the basic structure, extracts general hierarchical features of the image through convolution and attention mechanism, and provides the subsequent multi-task head for sharing, and the output is a high-dimensional feature map, which not only retains the rich details of the flame (such as different height levels and dynamic forms of the flame), but also avoids repeated extremes through sharing, balances the calculation efficiency and feature expression ability, and is the "cornerstone" of multi-task collaborative learning, providing unified and high-quality feature input for the subsequent task head.

[0029] The multi-task feature layer (Neck) is a "bridge" between the backbone and the task head, which fuses the high-dimensional features output by the backbone through the FPN (feature pyramid network) structure, adjusts the channels (different tasks have different feature requirements, such as flame color recognition paying more attention to color channels and dynamic recognition focusing on motion-related channels, adjusting the channels to adapt the features to the tasks) or spatial enhancement (highlighting the key positions of the flame in the image space, such as the dynamic change area of the flame edge, enhancing the feature discrimination), so that a single feature can adapt to different task requirements, break the barrier of feature use between tasks, and improve the collaborative performance of multi-task, such as the flame quantity statistics and color recognition tasks, which can be more efficiently coordinated through the adjusted features.

[0030] Task-Specific Heads are independent output modules designed for different tasks, respectively classifying the collected flame image features, including the number of flames, the color of flames, the brightness of flames, and the dynamics of flames. A fully connected layer is used as the output "terminal", which can perform nonlinear transformation and integration of the previously extracted features, and output the classification results. Each head "exerts its own strength" based on the shared backbone features, and then the results are fused (for example, fewer flames, darker color, unstable dynamics, and comprehensive judgment of abnormal combustion state), to determine the heating furnace combustion state and ensure high-precision parallel prediction of multiple tasks without interference and collaborative output.

[0031] Different tasks (flame presence, color / brightness / dynamic state) have different losses, and the model needs to be dynamically balanced for overall optimization. The loss weighting dynamically balances the loss of different tasks, and the loss function is: The first half is the loss calculation related to the flame state (color, brightness, dynamics, etc.), and the second half is the loss calculation of the flame presence classification. Among them, s is the flame state, i.e., the color of the flame, the brightness of the flame, and the dynamics of the flame, is the weight coefficient of the s-th state task, is the real label encoding of the s-th state, is the probability that the model predicts that the sample belongs to the s-th state, is the cross-entropy loss core item, which measures the difference between the predicted probability and the real label; c represents the classification of the flame, i.e., whether there is a flame, is the weight coefficient of the binary classification task, used to balance its importance with the flame state classification task, is the real label of the presence of the flame, is the probability that the model predicts "flame present", is the binary classification cross-entropy loss.

[0032] The present application transmits the data collected by the sensors and high-temperature cameras to the central control system through Modbus and OPC protocols, which can ensure the reliability and integrity of the data, and avoid monitoring failure caused by network delay or packet loss. The central control system can present multi-source data on a unified platform to form a visual dashboard, which is convenient for operators to view the global state in real time; it can store full-quantity operation data and playback, which is convenient for management personnel to analyze and optimize the production process; at the same time, a safety threshold is provided, which automatically triggers an alarm when the parameter exceeds the range, preventing abnormal conditions from affecting production.

[0033] The combination of the cracking furnace state-action pair (s, a) has hundreds of millions of combinations, it is impossible to experience learning of each case during the training process, and the accuracy of the Q value estimation of the state-action pair and the action selection are also affected by the cumulative (s, a) experience and the number of experiences. In order to help the Critic network and the Actor network speed up the learning speed of the environment and improve the estimation accuracy of the actual value function Q value and the selected action, the simulation data output by the cracking furnace dynamic prediction model in S4 of the present application and the real multi-modal data jointly drive the Critic and Actor network parameter update, forming a "prediction-verification-optimization" closed loop, and accelerating the convergence to the global optimal strategy.

[0034] Specifically as Figure 2 As shown in the figure, from the cracking furnace, the observation (o, sensor data, flame video data), weight (w, which can be understood as state importance or resource allocation), state (s, key parameters of system operation such as temperature, pressure sequence), action (a, actuator output such as valve opening, equipment start-stop instruction), reward (r, index for measuring decision-making advantage and disadvantage such as energy consumption reduction and output improvement), and next state (s_{k+1}) are collected to form an experience tuple. The above experience is stored for subsequent algorithm training, similar to an "experience replay pool", breaking the data correlation and stabilizing the training process.

[0035] The double-Actor structure in the figure includes an Online Actor (online policy network) and a Target Actor (target policy network). The former learns and outputs actions in real time, and the latter updates with a delay to avoid training shocks. The Online Actor outputs actions based on the current state s through a policy, and also introduces noise to increase exploration, allowing the agent to try more decisions. At the same time, experiences are sampled from the memory cache, combined with the value evaluation feedback from the Critic network, and the Actor network parameters are optimized. The goal is to make the actions output by the policy obtain higher rewards.

[0036] The double-Critic structure in the figure also has an Online Critic (online value network) and a Target Critic (target value network). The Online Critic evaluates the value of the current (s, a) combination in real time, and the Target Critic provides a relatively stable target value. Then the value is calculated, the action sequence output by the Actor is integrated, the expected value of the current action is calculated, and the Critic network is updated to make the value evaluation more accurate.

[0037] The overall logic in the figure is to collect industrial environment experience → store in cache → Actor output exploratory action → Critic evaluate action value → bidirectional optimization strategy and value network → iterative improvement of decision-making ability in industrial scenarios.

[0038] The ultimate goal of the cracking furnace control is to stabilize the temperature of the cracking furnace even if the input material changes, so that the input material can be continuously heated at a stable temperature. The cracking furnace temperature stabilization control process is affected by two important factors: 1) whether the oxygen content is sufficient and stable, and sufficient and stable oxygen content ensures that the flame burns fully, ensuring that the heat energy generated per unit of fuel is maximized; 2) whether the fuel pressure is stable, stable pressure can ensure the stability of the input fuel adjustment, improve the effect of cracking furnace temperature control.

[0039] Due to the uncertainty of the prediction of nitrogen oxides emissions in the cracking furnace, it is difficult to model, in addition, the fuel pressure is affected by factors such as equipment linkage, pipeline aging, valve wear, etc., showing dynamic changes, making it difficult for existing control methods to adapt to these complex changes. The traditional control algorithm, model predictive control (MPC), establishes a dynamic model of the system, and optimizes the control strategy for a certain period of time in the future, which is the mainstream method of industrial control. However, the strong nonlinearity and multivariable coupling characteristics of the cracking furnace system result in a high dimensionality of the optimization variables of the MPC model (such as oxygen content, fuel pressure, valve opening), which requires a large amount of online calculation, making it difficult to meet the real-time requirements.

[0040] Parameterized model predictive control (PMPC) is introduced in S5 to reduce the computational complexity of model predictive control (MPC). Specifically, it represents some variables in MPC (such as control weights, constraint boundaries) as adjustable parameters, reducing the number of variables for online optimization and significantly reducing the computational complexity, allowing the algorithm to run quickly on industrial controllers (such as PLCs) to meet the real-time requirements of cracking furnace temperature control. At the same time, parameterization ensures the model prediction ability of MPC, ensuring control accuracy.

[0041] Parameterized model predictive control (PMPC) is combined with reinforcement learning (RL) agents to improve the ability and flexibility to respond to changing or unknown environments and disturbances. The RL agent takes the cracking furnace operating state (such as temperature, oxygen content, NOx concentration) as input, and through continuous interaction with the environment (i.e. the cracking furnace system), it learns the optimal PMPC parameter adjustment strategy. For example, when the oxygen content fluctuation is detected to cause the temperature to drop, the RL agent automatically adjusts the parameters of the PMPC frequency regulation of the fan to speed up the recovery of the oxygen content. Moreover, the RL agent has online learning ability and can adapt to nitrogen oxide prediction bias, fuel pressure fluctuations and other uncertainties in real time. For example, when the fuel pipe network pressure suddenly changes, the RL agent explores a new pressure regulation strategy through trial and error, gradually optimizing the parameters to quickly restore the system to stability.

[0042] As shown in Figure 3 The reinforcement learning agent serves as the "decision layer" and outputs a parameter vector to guide the parameterized components of the PMPC to adjust. The PMPC module defines the boundaries and optimization objectives of the control through constraints and objective functions. The optimizer solves the "simplified" optimization problem online based on the parameters to generate control variables. The PMPC module also includes a prediction model (to predict future states) and a parameterized control law (to quickly calculate control actions), making the optimization more efficient. The control variables act on the cracking furnace, and the state of the cracking furnace is fed back to the PMPC and the RL agent, forming a closed-loop control.

[0043] The PMPC of the present application is responsible for short-term model-based prediction and control, providing stable basic regulation. The RL agent dynamically adjusts the PMPC parameters from the perspective of long-term optimization, balancing control performance and computational cost. The combination of the two achieves the dual goals of "precise control + adaptive optimization". Compared with traditional MPC, this scheme can not only maintain accurate control of key parameters such as temperature and oxygen content, but also significantly reduce computational resource consumption, while having the adaptability to respond to complex working condition changes, providing an efficient solution for intelligent control of the cracking furnace.

[0044] The present application also provides a cracking furnace intelligent optimization system architecture, which is composed of a perception layer, a network layer, a platform layer and an application layer; The perception layer collects temperature, pressure, flow, material and image data in real time through sensors and high-temperature cameras, and collects raw data of the cracking furnace operation in real time like "five senses"; The network layer is used for real-time detection of production site data transmission to the data platform, and transmits data like "nerves"; The platform layer provides data storage and analysis, and performs data processing and machine learning modeling. Through the data analysis model, the production process is dynamically adjusted to provide the best process parameters and operation suggestions for the operator, and the data is processed and decisions are made like "brain"; The application layer provides various intelligent application services for users, and displays real-time data and optimization suggestions in the production process through a visual interface, like "display" to show the decisions, making it easy to operate and manage.

[0045] The sensor and high-temperature camera collect and capture the multi-source data collected by the operation and save them to the MySQL database. The high-temperature camera stores the photos taken in Minio, the algorithm execution service reads the pictures collected by the high-temperature camera from Minio, performs flame state recognition, stores the processed photos in Minio again, and finally displays the real-time picture results after processing in the visual interface.

[0046] It should be understood that the use of the term "example" anywhere in this description is not intended to limit the application's scope. Rather, the term "example" is used to illustrate certain examples and aspects of the application. It is not intended to indicate that the application is limited in scope to the examples given. Furthermore, it is intended that every means within the scope of the example be considered to be within the scope of the application. Moreover, it is intended that every combination of means within the scope of the example be considered to be within the scope of the application.

Claims

1. A cracking furnace intelligent optimization system control method, characterized in that: The following steps are involved: S1. Sensors and high-temperature cameras deployed at key nodes of the cracking furnace collect real-time time series characteristic data and flame combustion video data of the cracking furnace operation; S2. Preprocess the collected multimodal data, including reading, sampling, cleaning, and dimensionality reduction of time series feature data, and extracting flame state features from video data using deep learning to identify flame combustion states and convert them into time series feature data. S3. Use federated learning to integrate preprocessed multimodal data. By combining the correlation and complementarity between the cracking furnace time series feature data and the flame video data, a loss function or regularization term is designed for joint training to construct a cross-modal causal logic model. S4. Design a deep reinforcement learning framework that combines a cross-modal causal logic model with a data-driven model. This framework incorporates a dynamic prediction model for the cracking furnace. This accelerates learning of the environment and improves the accuracy of estimating value functions and actions. This framework then learns the causal relationships between various factors during the cracking furnace's operation and infers the recommended parameter ranges for optimal operation. S5. Combining dynamic process data with environmental emission prediction models, a data-driven MPC algorithm framework is used to calculate adjustments to fan frequency and fuel pressure parameters in real time, achieving dynamic optimization control of the cracking furnace. S6. The operator makes adjustments or selects emergency operations based on the recommended parameters and strategies to achieve human-machine collaborative control.

2. The method for controlling a cracking furnace intelligent optimization system according to claim 1, wherein: The flame video data in S3 uses images as input and adopts a deep neural network architecture with a shared CNN backbone and multi-task heads (Shared Backbone + Multi-TaskHeads). Through a joint training mechanism, it achieves multi-objective joint learning of flame combustion states.

3. The method for controlling a cracking furnace intelligent optimization system according to claim 1, wherein: The loss function is used to dynamically balance the losses of different tasks by weighting the losses: Among them, s is the flame state, that is, the color of the flame, the brightness of the flame, and the dynamics of the flame, is the weight coefficient of the s-th state task, is the true label encoding for the s-th class state, The model predicts the probability that the sample belongs to the s-th category state, It is the core term of cross entropy loss, which measures the difference between the predicted probability and the true label; c represents the classification of flame, that is, whether there is flame, is the weight coefficient of the binary classification task, which is used to balance its importance with the flame state classification task. A true label for the presence or absence of flames, The model predicts the probability of "flame", is the binary cross entropy loss.

4. The method for controlling a cracking furnace intelligent optimization system according to claim 1, wherein: The data collected by sensors and high-temperature cameras are transmitted to the central control system through Modbus and OPC protocols.

5. A cracking furnace intelligent optimization system control method according to claim 4, characterized in that: The central control system can present multi-source data on a unified platform; and / or, the ability to store full run data and playback; And / or, a safety threshold is set to automatically trigger an alarm when the parameter exceeds the range.

6. The method for controlling a cracking furnace intelligent optimization system according to claim 1, wherein: The simulated data output by the cracking furnace dynamic prediction model in S4 and the real multimodal data jointly drive the update of the critic and actor network parameters.

7. A cracking furnace intelligent optimization system control method according to claim 6, characterized in that: S5 introduces parameterized model predictive control (PMPC) to reduce the computational complexity of model predictive control (MPC).

8. The method for controlling a cracking furnace intelligent optimization system according to claim 1, wherein: The integration of parameterized model predictive control (PMPC) and reinforcement learning (RL) agents improves the ability and flexibility to cope with changing or unknown environments and disturbances.

9. A cracking furnace intelligent optimization system architecture, characterized in that: It consists of the perception layer, network layer, platform layer and application layer; The perception layer collects temperature, pressure, flow, material and image data in real time during the production process through sensors and high-temperature cameras; The network layer is used to transmit production site data to the data platform for real-time detection; The platform layer provides data storage and analysis, and performs data processing and machine learning modeling. It dynamically adjusts the production process through data analysis models and provides operators with optimal process parameters and operation suggestions. The application layer provides users with various intelligent application services and displays real-time data and optimization suggestions during the production process through a visual interface.

10. The intelligent optimization system architecture of a cracking furnace according to claim 9, characterized in that: The multi-source data collected by the sensor and high temperature camera acquisition and crawling operations are saved in a MySQL database; The high-temperature camera stores the captured photos in Minio. The algorithm execution service reads the images captured by the high-temperature camera from Minio, identifies the flame status of the furnace, stores the processed photos in Minio again, and finally displays the processed real-time image results in the visualization interface.

Images