A chemical fertilizer production line equipment control system and method

Abstract

The present application relates to the technical field of automatic control of chemical fertilizer production, in particular to a chemical fertilizer production line equipment control system and method. The system adopts a three-layer architecture: a multi-modal perception and execution layer is used to collect the physical and chemical properties of materials and the equipment state in real time; an edge intelligent control layer is responsible for data preprocessing and local closed-loop control; a cloud digital twin and global scheduling layer optimizes the global strategy through the construction of a three-dimensional digital model and a multi-agent reinforcement learning model. The present application aligns the timestamps of perception data, predicts the evolution of digital twins in real time, and cooperatively controls the logic of multiple agents, and then sends the optimal action vector to the edge side for smooth execution. This scheme solves the problems of strong process parameter coupling, manual adjustment lag and high energy consumption in chemical fertilizer production, realizes precise centralized control of the whole process, improves the product qualification rate, and effectively reduces the energy efficiency and fault shutdown risk of the whole plant.

Description

Technical Field

[0001] This invention belongs to the field of automated control technology for fertilizer production, specifically relating to equipment control systems and methods for fertilizer production lines. Background Technology

[0002] Fertilizer production is a complex, continuous industrial process involving chemical reactions, physical granulation, and thermodynamic mass and heat transfer, exhibiting high nonlinearity, strong coupling, and hysteresis. Traditional fertilizer production line control primarily relies on field instruments and distributed control systems (DCS), with process parameters set manually based on experience. However, the following technical bottlenecks exist in actual production: First, the perception dimension is singular and lagging. Currently, key indicators such as granulation rate and raw material nutrient fluctuations still mainly rely on manual sampling and testing. The data collection frequency is low and the feedback cycle is long, which causes control decisions to often lag behind production fluctuations, easily leading to large-scale rework or scrap.

[0003] Second, there is a lack of global coordinated control. The batching, granulation, drying, and cooling processes in a fertilizer production line are interconnected, with strong physical links. Existing control methods mostly involve independent control of each process, which makes it difficult to cope with the chain reaction caused by slight adjustments to upstream parameters. This makes it difficult to achieve optimal overall system efficiency, and the parameter optimization process across the entire plant is extremely slow when raw material composition fluctuates.

[0004] Third, there is a disconnect between equipment maintenance and production control. Sudden failures of large rotating equipment such as dryers and elevators often lead to unplanned shutdowns. Existing monitoring systems mostly rely on threshold alarms and lack predictive maintenance schemes based on operational status characteristics, making it difficult to achieve predictive degradation protection while ensuring production continuity. Summary of the Invention

[0005] To overcome the above-mentioned technical problems, the present invention provides a control system and method for fertilizer production line equipment.

[0006] The present invention adopts the following technical solution: A centralized control system for the entire process of a fertilizer production line, the system adopting a three-layer architecture in physical topology, including: The multimodal sensing and execution layer is deployed at the end of each physical device in the fertilizer production line. It is used to collect real-time data on the operating status of the equipment and the physical and chemical properties of the materials, and to receive instructions from the upper layer to drive the physical devices to perform actions. The edge intelligent control layer consists of multiple industrial edge computing nodes and high-speed Ethernet switches distributed in various production workshops. It is used to process the raw data uploaded by the multimodal sensing and execution layers, and independently maintain the stable operation of the current section's equipment by using a locally preset expert rule base and adaptive proportional integral derivative control algorithm when there is network latency or interruption in the cloud. The cloud-based digital twin and global scheduling layer is deployed in a cloud server cluster and includes a data lake, a digital twin engine, a multi-agent reinforcement learning model, and a global production scheduling center. It is used to aggregate data from the entire plant and reconstruct a three-dimensional digital model that maps to the physical production line in virtual space. The multi-agent reinforcement learning model outputs the globally optimal strategy, and the global production scheduling center converts the strategy into specific equipment control settings and sends them to the edge intelligent control layer.

[0007] Preferably, the multimodal sensing and execution layer includes an online near-infrared spectroscopy analyzer deployed in the raw material batching area, a multi-acoustic array sensor and an industrial microphone arranged around the outside of the granulator cylinder, and a triaxial high-frequency vibration sensor deployed at the bearing of a large rotating device; the multimodal sensing and execution layer also includes a frequency converter, a high-precision mass flow meter, and a pneumatic regulating ball valve for receiving instructions and executing actions.

[0008] Preferably, the edge computing node has a built-in field-programmable gate array and a graphics processing unit for timestamping and Kalman filtering noise reduction of the raw data collected by the sensor; the edge computing node is also used to extract features from the collected acoustic and vibration unstructured data using a built-in lightweight convolutional neural network model, and to convert the acoustic spectrum into a structured feature vector characterizing the pelletizing rate.

[0009] Preferably, the cloud-based digital twin and the digital twin engine in the global scheduling layer have built-in fertilizer granulation kinetics model and drying thermodynamic mass and heat transfer model, which can perform high-frequency iterative calculations using real-time data of the physical line as boundary conditions; the multi-agent reinforcement learning model abstracts the production line into granulation agent, drying agent and batching agent, and uses the lowest energy consumption of the whole plant and the highest rate of high-quality finished products as the comprehensive reward function for trial and error training of collaborative adjustment strategies.

[0010] Preferably, the edge nodes of the granulation section in the edge intelligent control layer are used to fuse in real time the pelletizing rate features extracted by the acoustic array, the material flow rate of the feed belt, and the ambient humidity data, and receive commands from the cloud to perform closed-loop fine-tuning that combines feedforward and feedback, such as the operating frequency of the granulator drum motor, the opening degree of the steam injection regulating valve, and the rotation speed of the granulation spray pump.

[0011] This invention also discloses a centralized control method for the entire process of fertilizer production line equipment based on the above system, comprising the following steps: The synchronous sensing and edge preprocessing steps of global multimodal production data involve real-time collection of multimodal sensing data, microsecond-level timestamp labeling and dimensionality reduction structuring, and then uploading to the cloud data lake. The cloud-based digital twin state mapping and full-process evolution prediction steps update the model parameters of each machine in the virtual fertilizer production line in real time, predict the production process in the future and output multi-dimensional prediction curves. Based on the global parameter optimization step of multi-agent deep reinforcement learning, when the trend of product quality deviation or equipment energy consumption exceeding the standard is predicted in the future, the multi-agent deep reinforcement learning model is activated to play games and cooperate, and outputs the continuous equipment action space vector that achieves the optimal comprehensive benefits of the whole plant. Multi-level machine collaborative control and edge-side smooth execution steps resolve the optimal continuous device motion space vector into control setpoints and send them to the edge computing nodes. The edge computing nodes then generate smooth transition curves to guide the physical devices to transition to the new target working point.

[0012] Preferably, in the synchronous perception and edge preprocessing step of the global multimodal production data, the edge computing node uses a precise time protocol to perform time series alignment on all data streams, performs dimensionality reduction processing on unstructured data to remove redundant information, and uploads the structured equipment status matrix, process parameter matrix, and environmental parameter matrix to the cloud data lake in an encrypted manner through a message queue telemetry transmission protocol.

[0013] Preferably, in the cloud-based digital twin state mapping and full-process evolution prediction step, the cloud-based digital twin engine uses high-order partial differential equations to solve the mass conservation and energy conservation states of materials in the entire production line, and extrapolates the production process over the next three hours at ultra-real-time speed, outputting multi-dimensional prediction curves including energy consumption trends at each stage, bottleneck equipment warnings, and the nutrient content of the final fertilizer product.

[0014] Preferably, in the global parameter optimization step based on multi-agent deep reinforcement learning, each process agent takes the current state matrix of all equipment in the plant as the initial state of the environment, and scores the action combination by taking into account the overall plant benefit evaluation function that takes into account maximizing output, minimizing energy consumption per ton of fertilizer, and minimizing nutrient deviation. Finally, the optimal action vector is output, which includes the amount of fine adjustment of the batching ratio, the correction curve of the liquid phase injection amount of the granulator, the set value of the temperature gradient of the dryer, and the dynamic curve of the negative pressure of the induced draft fan.

[0015] Preferably, the method further includes predictive maintenance and abnormal safety degradation control steps for equipment operation status: the edge computing node continuously monitors the high-frequency vibration data of the large rotating machinery, and uses a long short-term memory neural network to analyze the drift characteristics of the bearing vibration spectrum in real time to predict the remaining service life and potential failure points; when it is determined that there is an imminent failure risk in the critical machine, the edge computing node takes over the control of the equipment and triggers a local safety degradation sequence, which specifically includes automatically reducing the front-end feed rate, slowing down the overall operating speed, and interlocking to shut down relevant dangerous gas and heat sources.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention achieves transparent perception and proactive prediction throughout the entire production process. By introducing near-infrared spectroscopy analysis and acoustic array sensing, coupled with the high-frequency iteration of a digital twin engine, the system can calculate future operating states at ultra-real-time speeds. Compared to traditional solutions, this invention can predict product moisture and nutrient deviations in advance, transforming "post-event adjustments" into "pre-event predictions," significantly improving the yield of superior-grade products.

[0017] It reduces unit energy consumption and optimizes resource allocation. Utilizing a multi-agent deep reinforcement learning model, the system can overcome the limitations of human experience and find the optimal solution for overall plant energy efficiency among thousands of parameter combinations. Through deep coupling control of processes such as batching, granulation, and drying, it can effectively reduce fuel consumption and electricity waste, thereby lowering the energy consumption per ton of standard coal in fertilizer production.

[0018] This enhances the resilience and safety of the production line. The architecture, combining edge computing and cloud collaboration, ensures the inherent safety of the production line even under abnormal conditions such as network fluctuations. Simultaneously, the predictive maintenance mechanism based on LSTM neural networks can identify early signs of equipment degradation and execute automatic degradation control, effectively mitigating the significant economic losses caused by sudden downtime.

[0019] It enhances the adaptability to complex raw materials. The dynamic adaptive batching system can automatically adjust the proportioning instructions based on the real-time fluctuations in the composition of raw materials, solving the problem of unstable product quality caused by uneven nutrient content in different batches of raw materials, and reducing the cost of production dependence on high-standard raw materials. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0021] This embodiment provides a centralized control system and method for the entire process of fertilizer production line equipment based on digital twins and multi-agent collaboration. This technical solution aims to solve the technical problems of isolated equipment control at each stage in traditional fertilizer production lines, lag in parameter adjustment when facing raw material fluctuations, and the inability to globally optimize energy and quality across the entire plant.

[0022] Part 1: Control System Architecture Examples The fertilizer production line equipment control system in this embodiment adopts a three-layer architecture in terms of physical topology, which is a collaborative deployment of edge computing and cloud computing. Specifically, it includes a multimodal perception and execution layer, an edge intelligent control layer, and a cloud digital twin and global scheduling layer.

[0023] I. Multimodal Perception and Execution Layer The multimodal sensing and execution layer is deployed at each physical device end of the fertilizer production line. It is responsible for collecting real-time data on the operating status of the equipment and the physical and chemical properties of the materials, and receiving instructions from the upper layer to drive the physical devices to perform actions.

[0024] This layer not only includes conventional temperature sensors, pressure transmitters, flow meters, and weighing sensors, but also creatively introduces a high-frequency multimodal sensing array. In the raw material batching area, a near-infrared spectroscopy online analyzer is deployed for real-time, non-destructive testing of the initial nitrogen, phosphorus, and potassium content and real-time moisture content of various mixed raw materials. Outside the granulator cylinder, the system surrounds multiple acoustic array sensors and industrial microphones to capture the acoustic spectrum characteristics of material tumbling and collision during granulation. At the bearings of large rotating equipment such as dryers and coolers, triaxial high-frequency vibration sensors are deployed to extract microscopic features of equipment degradation.

[0025] In terms of actuators, all conveyor belts, granulator main shafts, fans, and water pumps throughout the entire line are controlled by frequency converters. The liquid addition pipeline is equipped with a high-precision mass flow meter and a pneumatic regulating ball valve to achieve microsecond-level response and precise dosing of material flow.

[0026] II. Edge Intelligent Control Layer The edge intelligent control layer consists of multiple industrial edge computing nodes distributed across various production workshops and high-speed Ethernet switches. Each edge computing node corresponds to a specific process segment, such as edge nodes for the batching segment, granulation segment, drying and cooling segment, and screening and coating segment.

[0027] Edge computing nodes, equipped with field-programmable gate arrays (FPGAs) and high-performance graphics processing units, are primarily responsible for processing the massive and high-frequency raw data uploaded from the multimodal sensing layer. The edge nodes first perform timestamp synchronization and Kalman filtering noise reduction on the sensor data. For unstructured data such as acoustics and vibration, the edge computing nodes utilize a built-in lightweight convolutional neural network model for feature extraction, transforming complex acoustic spectra into structured feature vectors representing the pelletizing rate.

[0028] In addition, the edge intelligent control layer undertakes the task of real-time closed-loop control of key equipment. When there is a delay or interruption in the cloud network, the edge nodes can independently maintain the stable operation of the equipment in the current section by using the locally preset expert rule base and adaptive proportional-integral-derivative control algorithm, thus ensuring the inherent safety of the production process.

[0029] III. Cloud-based Digital Twin and Global Scheduling Layer The cloud-based digital twin and global scheduling layer is the core brain of the entire fertilizer production line. It is deployed in the enterprise's private cloud server cluster and includes a data lake, a digital twin engine, a multi-agent reinforcement learning model, and a global production scheduling center.

[0030] The data lake is responsible for aggregating historical production data, equipment ledgers, environmental temperature and humidity data, and real-time data uploaded from edge nodes across the entire plant.

[0031] The digital twin engine utilizes multiphysics simulation software and historical operational data to reconstruct a three-dimensional digital model in virtual space that perfectly maps to the physical production line. This digital twin not only includes the geometric features of the equipment but also incorporates a fertilizer granulation kinetics model and a drying thermodynamics mass and heat transfer model. Using real-time data from the physical line as boundary conditions, the digital twin performs high-frequency iterations, enabling ultra-real-time calculations of the production results of current process parameters over a future period, such as predicting the strength and moisture content of the finished granules ten minutes later.

[0032] The multi-agent reinforcement learning model abstracts the entire production line into multiple cooperating agents. The granulation agent, drying agent, and batching agent undergo millions of trial-and-error training iterations in a virtual environment provided by a digital twin. The model uses the lowest overall plant energy consumption and the highest yield of high-quality finished products as the comprehensive reward function, learning the optimal multi-equipment collaborative adjustment strategy under various complex raw material fluctuations and environmental disturbances.

[0033] The global production scheduling center then transforms the globally optimal strategy output by the reinforcement learning model into specific equipment control settings, which are then distributed to various edge computing nodes via the Industrial Internet of Things (IIoT) protocol.

[0034] Part Two: Implementation Examples of Collaborative Control of Core Equipment This embodiment focuses on the technological solution of centralized control of machines in fertilizer production lines with strong coupling relationships.

[0035] I. Dynamic Adaptive Batching and Mixing Control Traditional batching systems operate with fixed formulas. In this system, the edge nodes of the batching section receive real-time raw material nutrient and moisture data from a near-infrared spectrometer. The cloud-based global scheduling center dynamically fine-tunes the given weight ratio of the feeding belt scale every ten minutes based on the national standard requirements of the finished product and the real-time detected deviations in raw material nutrients, using a nonlinear programming algorithm.

[0036] In terms of mixer control, the edge node assesses the mixing uniformity of materials by monitoring the real-time operating current and power factor of the mixer's main motor. Only when the system determines that the mixing uniformity meets the standard will it control the pneumatic discharge valve to open, effectively avoiding energy waste caused by mixing dead zones or over-mixing due to changes in raw material viscosity.

[0037] II. Depth Control of Granulators with Multivariable Coupling Granulation is the core process in fertilizer production. This solution treats the rotary drum granulator, liquid ammonia injection pump, steam regulating valve, and return conveyor belt as a deeply multivariable coupled system for centralized control.

[0038] The granulation section's edge nodes integrate real-time data from acoustic array-extracted pelleting rate features, material flow rate from the feed belt, and ambient humidity. Based on these state variables, the cloud-based granulation AI outputs three core control commands: First, the drum motor's operating frequency, used to adjust the material's drop height within the drum. Second, the opening degree of the steam injection regulating valve, used to control system heat input and material liquid phase quantity. Third, the rotational speed of the granulation spray pump.

[0039] The system employs a combined feedforward and feedback strategy during control. When the moisture content of the material from the batching section suddenly increases, the system triggers feedforward adjustment. Before the material reaches the granulator, the steam regulating valve is proportionally reduced and the drum speed is increased to prevent material from agglomerating and sticking to the granulator walls. Simultaneously, closed-loop fine-tuning is performed using pelleting quality data fed back from acoustic sensors to stabilize the pelleting pass rate at an extremely high level.

[0040] III. Predictive Feedforward Control of Drying and Cooling Sections The dryer and cooler exhibit significant thermal inertia, resulting in substantial lag in parameter adjustments. This technical solution employs a model predictive control algorithm based on a digital twin model in the drying and cooling section.

[0041] The cloud-based digital twin engine predicts the moisture content of the material at the dryer's outlet based on the current temperature, moisture content, and flow rate of the material discharged from the granulator, combined with the current fuel supply to the hot blast stove and the negative pressure of the induced draft fan. If the predicted outlet moisture content will exceed the upper limit of the mass, the system will not blindly and immediately increase the combustion load of the hot blast stove significantly. Instead, it will calculate an optimal operating trajectory through multi-agent collaboration.

[0042] The system first fine-tunes the rotational speed of the dryer's main shaft to extend the residence time of the material inside the drum. Simultaneously, it smoothly increases the frequency of the induced draft fan to increase the air volume, and finally, it slowly increases the gas input to the hot air furnace. This predictive control, with centralized coordination of multiple devices, not only avoids drastic fluctuations in hot air temperature and significantly reduces the consumption of natural gas or coal, but also prevents fertilizer granules from surface cracking due to rapid heating, thus improving the granules' crushing resistance.

[0043] Part Three: Examples of Centralized Control Methods for Fertilizer Plant Equipment Based on the above system architecture, this embodiment provides a centralized control method for the entire process of fertilizer production line equipment. This method achieves seamless integration and machine collaborative control through the following core steps.

[0044] Step 1: Synchronous sensing and edge preprocessing of global multimodal production data High-frequency multimodal sensing arrays deployed on various equipment on the production line collect operating parameters in real time. After receiving these parameters, the edge computing nodes first use a precise time protocol to timestamp all data streams at the microsecond level, ensuring absolute alignment of data from different work sections in the time series.

[0045] Subsequently, the edge computing nodes perform dimensionality reduction processing on the unstructured data, remove redundant information, and encrypt and upload the structured equipment status matrix, process parameter matrix, and environmental parameter matrix to the cloud data lake via message queue telemetry transmission protocol.

[0046] Step Two: Cloud-based Digital Twin State Mapping and Full-Process Evolution Prediction The cloud-based digital twin engine reads the latest perception matrix from the data lake and updates the model parameters of each machine in the virtual fertilizer production line in real time. The engine uses high-order partial differential equations to solve for the mass and energy conservation states of materials throughout the entire production line.

[0047] Based on the current state, the digital twin can project the production process over the next three hours at a speed a hundred times faster than actual production, outputting multi-dimensional prediction curves including energy consumption trends at each stage, early warnings of bottleneck equipment, and the nutrient content of the final fertilizer product.

[0048] Step 3: Global parameter optimization based on multi-agent deep reinforcement learning When the digital twin predicts a future trend of product quality deviation or equipment energy consumption exceeding the standard, the system automatically activates the multi-agent deep reinforcement learning model.

[0049] The model uses the current state matrix of all equipment in the plant as the initial state of the environment. Under this environment, various process agents engage in game theory and cooperation, proposing parameter adjustment actions for the equipment under their jurisdiction. The model scores these action combinations using a pre-defined plant-wide comprehensive benefit evaluation function, which considers maximizing output, minimizing energy consumption per ton of fertilizer, and minimizing nutrient deviation.

[0050] After rapid iterative calculations on the cloud server, the model outputs a set of continuous equipment motion space vectors that can achieve the best overall plant efficiency. These vectors include the fine-tuning amount of the batching ratio, the correction curve of the liquid phase injection amount of the granulator, the set value of the temperature gradient of the dryer, and the dynamic curve of the negative pressure of the induced draft fan.

[0051] Step 4: Multi-level machine collaborative control and smooth execution at the edge. The global production scheduling center resolves the above optimal motion space vector into control setpoints for machines in different work sections and sends them to the corresponding edge computing nodes.

[0052] To prevent sudden changes in control commands from impacting the mechanical structure of physical equipment, the edge computing node initiates its local trajectory planning algorithm upon receiving a new setpoint. The edge node generates a smooth transition curve, guiding the inverters and regulating valves of each device to gradually transition to the new target operating point with a certain acceleration or rate of change.

[0053] Step 5: Predictive maintenance and degraded safety control for equipment operating status While performing normal process control, edge computing nodes continuously monitor high-frequency vibration data of large rotating machinery. Pre-installed long short-term memory neural networks at the edge analyze the drift characteristics of the bearing vibration spectrum in real time, predicting the remaining service life and potential failure points of the equipment.

[0054] If the system determines that a critical machine, such as a bucket elevator or cooling fan, has an imminent risk of failure, the edge node will immediately take over control of the equipment. Simultaneously with a cloud-based alert, the system will trigger a local safety degradation sequence. This sequence will automatically reduce the feed rate at the front end, slow down the overall operating speed, and interlock to shut down relevant hazardous gas and heat sources. This ensures the equipment is safely shut down for maintenance under controlled conditions with minimal material buildup, minimizing economic losses caused by abnormal downtime.

[0055] Although embodiments of the present invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to the above embodiments without departing from the principles and spirit of the present invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A full-process centralized control system for a chemical fertilizer production line equipment, characterized in that, The system adopts a three-layer architecture in terms of physical topology, including: The multimodal sensing and execution layer is deployed at the end of each physical device in the fertilizer production line. It is used to collect real-time data on the operating status of the equipment and the physical and chemical properties of the materials, and to receive instructions from the upper layer to drive the physical devices to perform actions. The edge intelligent control layer consists of multiple industrial edge computing nodes and high-speed Ethernet switches distributed in various production workshops. It is used to process the raw data uploaded by the multimodal sensing and execution layers, and independently maintain the stable operation of the current section's equipment by using a locally preset expert rule base and adaptive proportional integral derivative control algorithm when there is network latency or interruption in the cloud. The cloud-based digital twin and global scheduling layer is deployed in a cloud server cluster and includes a data lake, a digital twin engine, a multi-agent reinforcement learning model, and a global production scheduling center. It is used to aggregate data from the entire plant and reconstruct a three-dimensional digital model that maps to the physical production line in virtual space. The multi-agent reinforcement learning model outputs the globally optimal strategy, and the global production scheduling center converts the strategy into specific equipment control settings and sends them to the edge intelligent control layer.

2. The control system of claim 1, wherein, The multimodal sensing and execution layer includes an online near-infrared spectroscopy analyzer deployed in the raw material batching area, a multi-acoustic array sensor and industrial microphone arranged around the outside of the granulator cylinder, and a triaxial high-frequency vibration sensor deployed at the bearing of a large rotating device; the multimodal sensing and execution layer also includes a frequency converter, a high-precision mass flow meter, and a pneumatic regulating ball valve for receiving commands and executing actions.

3. The control system of claim 1, wherein, The edge computing node has a built-in field-programmable gate array and a graphics processing unit for timestamping and Kalman filtering noise reduction of the raw data collected by the sensor. The edge computing node is also used to extract features from the collected acoustic and vibration unstructured data using a built-in lightweight convolutional neural network model, and to transform the acoustic spectrum into a structured feature vector characterizing the pelletizing rate.

4. The control system of claim 1, wherein, The cloud-based digital twin and global scheduling layer's digital twin engine incorporates a fertilizer granulation kinetics model and a drying thermodynamics mass and heat transfer model, enabling high-frequency iterative calculations using real-time data from the physical line as boundary conditions. The multi-agent reinforcement learning model abstracts the production line into granulation agents, drying agents, and batching agents, using the lowest overall plant energy consumption and the highest finished product premium rate as the comprehensive reward function for trial-and-error training of collaborative adjustment strategies.

5. The control system according to claim 1, characterized in that, The edge nodes of the granulation section in the edge intelligent control layer are used to fuse in real time the pelletizing rate characteristics extracted by the acoustic array, the material flow rate of the feed belt and the ambient humidity data, and receive commands from the cloud to adjust the granulator drum motor operating frequency, steam injection regulating valve opening and granulation spray pump speed in a closed-loop fine-tuning manner that combines feedforward and feedback.

6. A method for centralized control of the entire process of fertilizer production line equipment based on the system described in any one of claims 1 to 5, characterized in that, Includes the following steps: The synchronous sensing and edge preprocessing steps of global multimodal production data involve real-time collection of multimodal sensing data, microsecond-level timestamp labeling and dimensionality reduction structuring, and then uploading to the cloud data lake. The cloud-based digital twin state mapping and full-process evolution prediction steps update the model parameters of each machine in the virtual fertilizer production line in real time, predict the production process in the future and output multi-dimensional prediction curves. Based on the global parameter optimization step of multi-agent deep reinforcement learning, when the trend of product quality deviation or equipment energy consumption exceeding the standard is predicted in the future, the multi-agent deep reinforcement learning model is activated to play games and cooperate, and outputs the continuous equipment action space vector that achieves the optimal comprehensive benefits of the whole plant. Multi-level machine collaborative control and edge-side smooth execution steps resolve the optimal continuous device motion space vector into control setpoints and send them to the edge computing nodes. The edge computing nodes then generate smooth transition curves to guide the physical devices to transition to the new target working point.

7. The control method according to claim 6, characterized in that, In the synchronous sensing and edge preprocessing steps of the global multimodal production data, the edge computing nodes use a precise time protocol to perform time series alignment on all data streams, perform dimensionality reduction processing on unstructured data to remove redundant information, and encrypt and upload the structured equipment status matrix, process parameter matrix, and environmental parameter matrix to the cloud data lake through a message queue telemetry transmission protocol.

8. The control method according to claim 6, characterized in that, In the cloud-based digital twin state mapping and full-process evolution prediction steps, the cloud-based digital twin engine uses high-order partial differential equations to solve the mass and energy conservation states of materials in the entire production line, and extrapolates the production process over the next three hours at ultra-real-time speed, outputting multi-dimensional prediction curves including energy consumption trends at each stage, bottleneck equipment warnings, and the nutrient content of the final fertilizer product.

9. The control method according to claim 6, characterized in that, In the global parameter optimization step based on multi-agent deep reinforcement learning, each process agent takes the current state matrix of all equipment in the plant as the initial state of the environment. The action combination is scored by the plant-wide comprehensive benefit evaluation function that takes into account maximizing output, minimizing energy consumption per ton of fertilizer, and minimizing nutrient deviation. Finally, the optimal action vector is output, which includes the fine adjustment amount of the batching ratio, the correction curve of the liquid phase injection amount of the granulator, the set value of the temperature gradient of the dryer, and the dynamic curve of the negative pressure of the induced draft fan.

10. The control method according to claim 6, characterized in that, The method also includes predictive maintenance and abnormal safety degradation control steps for equipment operation status: edge computing nodes continuously monitor high-frequency vibration data of large rotating machinery, and use long short-term memory neural networks to analyze the drift characteristics of bearing vibration spectrum in real time to predict remaining service life and potential failure points; when it is determined that there is an imminent failure risk in a critical machine, the edge computing node takes over the control of the equipment and triggers a local safety degradation sequence, which specifically includes automatically reducing the front-end feed rate, slowing down the overall operating speed, and interlocking to shut down relevant dangerous gas and heat sources.

Images