Pneumatic conveying system for powder materials and control method thereof

Abstract

The application provides a kind of pneumatic conveying system for powder material and its control method, it is related to industrial automation control technical field, the method includes collecting the monitoring data of pneumatic conveying system, based on VMD-improved AGNGO-Transformer-GRU hybrid model, the key state parameters of pneumatic conveying system are predicted, based on the improved proximal policy optimization PPO algorithm, an intelligent decision controller is constructed, and the optimal control instruction is generated according to the state of pneumatic conveying system.The application realizes the accurate prediction of solid-gas ratio and pipe blocking risk based on the hybrid model, changes the traditional passive treatment into active early warning, effectively prevents pipe blocking accidents;The improved PPO algorithm makes the system autonomously optimize the best balance point of energy consumption and efficiency through the reward function of multi-objective optimization;Online adaptive learning mechanism makes the system continuously adapt to material characteristics and equipment state changes, greatly reduces the dependence on artificial experience.

Description

Technical Field

[0001] This invention relates to the field of industrial automation control technology, and in particular to a pneumatic conveying system for powder materials and its control method. Background Technology

[0002] As the global automotive industry transforms towards electrification and intelligentization, the demand for automated and precision material handling systems in vehicle manufacturing and upstream component production is becoming increasingly urgent. In the automotive manufacturing process, the handling of powder materials permeates several key stages: on the one hand, in tire manufacturing, large quantities of powder materials such as carbon black, silica, and zinc oxide are used as rubber reinforcing agents and additives. These materials are precisely delivered from the auxiliary equipment system to the internal mixer via pneumatic conveying systems. On the other hand, in the power battery manufacturing stage, the positive and negative electrode materials for lithium-ion batteries, such as lithium iron phosphate, lithium carbonate, ternary precursors, and conductive carbon black, are all fine powders, which also rely on pneumatic conveying to transfer from the raw material silo to the electrode coating process. Therefore, pneumatic conveying systems have become an indispensable key piece of equipment in modern automotive manufacturing workshops.

[0003] However, existing pneumatic conveying control methods applied in the automotive manufacturing field still have significant technical bottlenecks: The complex properties of materials lead to poor conveying stability: Whether it's carbon black in tire manufacturing or nano-sized powders in battery materials, both are characterized by small particle size, poor flowability, easy moisture absorption and agglomeration, and strong abrasiveness. Carbon black, as a lightweight powder with a high specific surface area, easily adsorbs and deposits on the inner wall of pipes, causing blockages. Furthermore, lithium battery positive and negative electrode materials are extremely sensitive to magnetic foreign objects; once deposited, they can be washed away by airflow, introducing metallic impurities such as Fe, Cr, and Ni, contaminating high-purity materials and directly affecting battery safety and cycle life. Traditional conveying systems mostly rely on PID control based on fixed parameters or manual experience adjustments, making it difficult to adapt to subtle changes in the characteristics of different batches of materials in real time, leading to frequent pipe blockages and forced production line shutdowns.

[0004] The passive response mode leads to increased energy consumption and wear: existing control methods lack the ability to deeply perceive the internal state of the delivery pipeline and cannot grasp the core process parameter of "solid-to-gas ratio" in real time. To eliminate the risk of pipe blockage, the system is often forced to adopt a conservative delivery mode of "high gas velocity and low concentration". This extensive operation not only causes the fan to be in a high-energy-consumption state for a long time, but also aggravates pipeline wear and shortens equipment life, which is seriously inconsistent with the current automotive industry's requirements for "dual carbon" goals and lean production.

[0005] Insufficient automation hinders production cycle improvement: Although industrial automation is widely used in automobile manufacturing, in the specific segment of powder conveying, most systems remain at a semi-automatic level of "data acquisition + manual intervention." Adjustments to control parameters heavily rely on engineers' experience, failing to incorporate the risk of pipe blockage as prior knowledge into the control system. Once upstream material formulations are adjusted or equipment ages, the system requires lengthy trial runs to regain stability, severely restricting the production efficiency of complete vehicles and components.

[0006] In summary, traditional pneumatic conveying systems for powders are no longer sufficient to meet the stringent requirements of high continuity, high purity, low energy consumption, and adaptability in the demanding working conditions of automobile manufacturing (including tire and battery manufacturing). Therefore, it is essential to design a pneumatic conveying system and its control method for powder materials. Summary of the Invention

[0007] In order to overcome the shortcomings of the prior art, the purpose of this invention is to provide a pneumatic conveying system for powder materials and its control method.

[0008] To achieve the above objectives, the present invention provides the following solution: This invention provides a pneumatic conveying system for powder materials, comprising: a physical execution unit, a sensor network unit, and an intelligent control unit; The physical execution unit includes an air supply device, a material supply device, and a gas-solid separation device connected in sequence via a conveying pipeline; The sensor network unit is deployed at key nodes of the physical execution unit to collect system operating parameters in real time. The intelligent control unit is communicatively connected to the sensor network unit and the physical execution unit. The intelligent control unit includes: The data acquisition and processing module is used to receive and process data from the sensor network unit; The critical state parameter prediction module is used to predict the critical state parameters of the pneumatic conveying system based on the processed data using a VMD-improved AGNGO-Transformer-GRU hybrid model. The intelligent decision-making module is used to generate optimal control commands based on the system status.

[0009] Preferably, the sensor network unit includes: A pressure sensor installed at the outlet of the gas supply device is used to monitor the outlet pressure of the blower. A pressure sensor is installed at the inlet of the conveying pipe after the feeding device to monitor the inlet pressure of the pipe; A pressure sensor is installed at the outlet of the delivery pipeline before the gas-solid separation device to monitor the pipeline outlet pressure; A gas flow meter installed at the outlet of the gas supply device is used to monitor the volumetric flow rate of air. A current sensor connected to the air supply unit is used to monitor the operating current of the fan; A frequency acquisition device connected to the gas supply unit is used to monitor the operating frequency of the fan; A speed sensor connected to the feeding device is used to monitor the operating speed of the feeder.

[0010] Preferably, the intelligent decision-making module includes: State, action, and reward / penalty construction unit, used to construct state vectors, action vectors, and reward / penalty functions; The strategy network unit uses the near-end strategy optimization PPO algorithm to output the fan frequency adjustment and feeder speed adjustment. Value network unit, used to evaluate the state value function.

[0011] The present invention also provides a control method for a pneumatic conveying system for powder materials, applied to the above-mentioned pneumatic conveying system for powder materials, comprising: Step 1: Collect monitoring data from the pneumatic conveying system; Step 2: Based on the VMD-improved AGNGO-Transformer-GRU hybrid model, predict the key state parameters of the pneumatic conveying system; Step 3: Based on the improved near-end strategy optimization PPO algorithm, construct an intelligent decision controller to generate optimal control commands according to the state of the pneumatic conveying system.

[0012] Preferably, in step 1, the monitoring data of the pneumatic conveying system is collected, specifically as follows: Based on the sensor network unit, the monitoring data of the pneumatic conveying system is collected to generate a multi-sensor time series data vector, including the blower outlet source pressure, the conveying pipeline inlet pressure, the conveying pipeline outlet pressure, the total system pressure difference, the air volume flow rate, the blower operating frequency, the blower operating current, and the feeder operating speed.

[0013] Preferably, in step 2, the key state parameters of the pneumatic conveying system are predicted based on the VMD-improved AGNGO-Transformer-GRU hybrid model, specifically as follows: Construct a key state parameter prediction model based on the VMD-improved AGNGO-Transformer-GRU hybrid model; The key state parameter prediction model is trained based on a pre-set dataset; The multi-sensor time series data vector is input into the key state parameter prediction model to obtain the key state parameters, which include the predicted solid-gas ratio and the pipeline deposition state identification probability value.

[0014] Preferably, in step 3, an intelligent decision controller is constructed based on the improved near-end strategy optimization PPO algorithm to generate optimal control commands according to the state of the pneumatic conveying system, specifically as follows: Define the state space for reinforcement learning, whose state vector includes the fan outlet source pressure, the conveying pipeline inlet pressure, the conveying pipeline outlet pressure, the total system pressure difference, the air volume flow rate, the fan operating frequency, the fan operating current, the feeder operating speed, the solid-to-gas ratio prediction value, and the pipeline deposition state identification probability value. Define the dynamic space of reinforcement learning, whose dynamic vectors include the fan frequency adjustment and the feeder speed adjustment; Define the reward / punishment function for reinforcement learning, where the reward function R is... t At time t, it is defined as: R t =α*Efficiency t -β*Energy t -γ*Risk t -δ*Action t (1) In the formula, Efficiency t Energy t To incur energy consumption penalties, Risk t Action to punish the risk of pipe blockage t To control the smoothness penalty of actions, α, β, γ, and δ are the main weighting coefficients; A policy network and a value network based on the Proximal Policy Optimization (PPO) algorithm are constructed. Through training, the agent learns a control policy that maximizes the cumulative reward. The policy network outputs the mean of two action vectors and an independent adaptive log standard deviation to construct a Gaussian distribution of actions, achieving a balance between exploration and exploitation. The value network outputs a scalar value to represent the long-term expected reward of the current state.

[0015] Preferably, the method further includes: Step 4: Establish a closed-loop control process, cyclically execute state perception, intelligent decision-making, instruction execution and limit protection, and experience data storage, establish an online adaptive learning mechanism, and periodically or triggerively use real experience data collected during system operation to fine-tune the policy network and value network.

[0016] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention provides a pneumatic conveying system for powder materials and a control method thereof. The method includes collecting monitoring data of the pneumatic conveying system, predicting key state parameters of the pneumatic conveying system based on a VMD-improved AGNGO-Transformer-GRU hybrid model, constructing an intelligent decision controller based on an improved near-end strategy optimization PPO algorithm, and generating optimal control commands according to the state of the pneumatic conveying system. This invention, by constructing a closed-loop control system integrating mechanism perception and intelligent decision-making, has produced several significant beneficial effects: First, based on the VMD-improved AGNGO-Transformer-GRU hybrid model, it achieves accurate prediction of the solid-gas ratio and pipe blockage risk, enabling the system to have a deep perception capability of its internal state, transforming traditional post-fault handling into risk prevention, and fundamentally avoiding pipe blockage accidents; Second, the improved PPO algorithm, through multi-objective optimization reward function design, enables the system to autonomously find the optimal balance between energy consumption and efficiency, achieving energy-saving effects; Third, through an online adaptive learning mechanism, the system can continuously adapt to slow dynamics such as changes in material properties and equipment aging, maintaining optimal performance in long-term operation and significantly reducing reliance on human experience; Finally, the entire system forms a complete intelligent closed loop from state perception and intelligent diagnosis to optimization decision-making, realizing a leap from "automation" to "intelligence" in the pneumatic conveying process, improving system reliability and economy, and providing a feasible technical path for the digital transformation and upgrading of industrial production. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a diagram of the overall structure of Transformer-GRU. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The purpose of this invention is to provide a pneumatic conveying system and its control method for powdered materials. Based on a hybrid model, it achieves accurate prediction of the solid-to-gas ratio and the risk of pipe blockage, transforming traditional passive processing into proactive early warning and effectively preventing pipe blockage accidents. The improved PPO algorithm uses a multi-objective optimized reward function to enable the system to autonomously find the optimal balance between energy consumption and efficiency. The online adaptive learning mechanism enables the system to continuously adapt to changes in material characteristics and equipment status, significantly reducing reliance on human experience.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] This invention provides a pneumatic conveying system for powder materials, comprising: a physical execution unit, a sensor network unit, and an intelligent control unit; The physical execution unit includes an air supply device, a material supply device, and a gas-solid separation device connected in sequence via a conveying pipeline; The sensor network unit is deployed at key nodes of the physical execution unit to collect system operating parameters in real time. The intelligent control unit is communicatively connected to the sensor network unit and the physical execution unit. The intelligent control unit includes: The data acquisition and processing module is used to receive and process data from the sensor network unit; The critical state parameter prediction module is used to predict the critical state parameters of the pneumatic conveying system based on the processed data using a VMD-improved AGNGO-Transformer-GRU hybrid model. The intelligent decision-making module is used to generate optimal control commands based on the system status.

[0023] The sensor network unit includes: A pressure sensor installed at the outlet of the gas supply device is used to monitor the outlet pressure of the blower. A pressure sensor is installed at the inlet of the conveying pipe after the feeding device to monitor the inlet pressure of the pipe; A pressure sensor is installed at the outlet of the delivery pipeline before the gas-solid separation device to monitor the pipeline outlet pressure; A gas flow meter installed at the outlet of the gas supply device is used to monitor the volumetric flow rate of air. A current sensor connected to the air supply unit is used to monitor the operating current of the fan; A frequency acquisition device connected to the gas supply unit is used to monitor the operating frequency of the fan; A speed sensor connected to the feeding device is used to monitor the operating speed of the feeder.

[0024] The intelligent decision-making module includes: State, action, and reward / penalty construction unit, used to construct state vectors, action vectors, and reward / penalty functions; The strategy network unit uses the near-end strategy optimization PPO algorithm to output the fan frequency adjustment and feeder speed adjustment. Value network unit, used to evaluate the state value function.

[0025] The air supply device in the physical execution unit is a variable frequency driven Roots blower, and the feeding device is a speed-controlled driven rotary feeder; the actuator drive module controls the blower operating frequency by adjusting the output frequency of the frequency converter, and controls the feeder operating speed by adjusting the output of the driver.

[0026] The system monitoring and learning module supports seamless switching between manual control mode, automatic control mode and intelligent control mode.

[0027] The intelligent control unit is deployed on an industrial control computer and communicates with the sensor network unit and physical execution unit via industrial Ethernet.

[0028] like Figure 1 As shown, the present invention also provides a control method for a pneumatic conveying system for powder materials, applied to the aforementioned pneumatic conveying system for powder materials, comprising: Step 1: Collect monitoring data from the pneumatic conveying system; Step 2: Based on the VMD-improved AGNGO-Transformer-GRU hybrid model, predict the key state parameters of the pneumatic conveying system; Step 3: Based on the improved near-end strategy optimization PPO algorithm, construct an intelligent decision controller to generate optimal control commands according to the state of the pneumatic conveying system.

[0029] In step 1, monitoring data of the pneumatic conveying system is collected, specifically as follows: Monitoring data from the pneumatic conveying system is collected using a sensor network unit, generating a multi-sensor time-series data vector, including the fan outlet source pressure P. source Inlet pressure P of the conveying pipeline inlet , Pipeline outlet pressure P outlet Total system pressure difference ΔP section air volume flow rate Q air Fan operating frequency F fan Fan operating current I fan and feeder operating speed R feeder Since the monitoring process uses conventional techniques, it will not be described in detail here.

[0030] In step 2, the key state parameters of the pneumatic conveying system are predicted based on the VMD-improved AGNGO-Transformer-GRU hybrid model, specifically: Step 201: Construct a prediction model for key state parameters based on the VMD-improved AGNGO-Transformer-GRU hybrid model; Step 202: Train the prediction model for key state parameters based on the preset dataset; Step 203: Input the multi-sensor time series data vector into the key state parameter prediction model to obtain the key state parameters, including the predicted solid-gas ratio μ. hat and the probability value P for identifying pipeline deposition status deposition .

[0031] In step 201, a key state parameter prediction model is constructed based on the VMD-improved AGNGO-Transformer-GRU hybrid model, specifically as follows: This invention provides a detailed description of the VMD-improved AGNGO-Transformer-GRU hybrid model, which specifically includes: 1. Variational Mode Decomposition (VMD) Preprocessing Sensor signals (such as pressure and flow) in pneumatic conveying systems are typically non-stationary and nonlinear, and are mixed with noise. VMD adaptively decomposes the original signal into a series of relatively stationary intrinsic mode functions (IMFs) with a center frequency, thereby highlighting the physical processes at different time scales.

[0032] Input: Real-time time series data acquired from the sensor network in step 1, where each variable constitutes a separate time series; Output: Each original variable is decomposed into K IMF subsequences [IMF1(t), IMF2(t), ..., IMFK(t)] and a residual term r(t), which is equivalent to expanding a single feature channel into K+1 feature channels, greatly enriching the information dimension and quality of the model input; Key parameter: The number of modes K is determined by observing the changes in the center frequency under different K values ​​to avoid over-decomposition or under-decomposition.

[0033] 2. Transformer-GRU Feature Extraction and Temporal Modeling Core From the multi-channel IMF sequence after VMD decomposition, deep spatiotemporal features are extracted, and long-term dependencies of system dynamics are captured. Input: Multidimensional IMF sequences preprocessed and normalized by VMD; Structure as Figure 2 As shown, it specifically includes: (1) Transformer encoder: It utilizes its multi-head self-attention mechanism to calculate the relationship weights between all time steps in the sequence in parallel. This enables the model to simultaneously pay attention to events at different time points, such as the pressure change at the moment of wind turbine startup and the flow fluctuation during the stable delivery stage, thereby gaining a global understanding of the system state and overcoming the information decay problem of traditional RNN series models on long sequences. (2) GRU network: It receives high-dimensional features rich in global information extracted by Transformer and further performs serialization modeling. The update gate and reset gate mechanism of GRU enable it to selectively remember and forget information, and finely capture the temporal pattern of dynamic evolution of the system, such as the accumulation process of pipe blockage risk.

[0034] 3. Improved Adaptive Generalized Northern Eagle Algorithm (AGNGO) Optimizer Used to automatically and efficiently search for the optimal combination of hyperparameters in a Transformer-GRU model, ensuring optimal model performance; Optimization parameters include, but are not limited to, the number of Transformer layers, the number of attention heads, the number of hidden units in the GRU, the initial learning rate, the Dropout rate, and the training epochs. Its improvement strategy compared to the traditional AGNGO is as follows: First, Spatial Pyramid Matching (SPM) chaotic mapping is introduced for initialization in the initial stage of the original algorithm. This ensures that individuals are more evenly distributed throughout the search space as the algorithm iterates. Then, the follower position update formula from the Sparrow Search Algorithm (SSA) is improved, and a nonlinear parameter strategy using a sine function is applied to the prey identification stage position update formula in the NGO algorithm. This stage employs a dynamic judgment mechanism, based on the condition that the distance (d, m) between the eagle and the prey is greater than the safe distance (S). d (m) or the distance between the two is less than the alert value (d) v (m), alertness value refers to the distance threshold at which a goshawk becomes alert to its prey during a hunt (d). v =0.4m). When the prey is less than this value, the eagle will become more vigilant and adopt a more cautious hunting strategy to accelerate the convergence speed of the algorithm. Parameter v s For pursuit speed, this value plays a key role in the dynamic judgment mechanism. In addition, inspired by the Moth-Flame Optimization (MFO) algorithm, the spiral operator update strategy in MFO is applied to the pursuit and escape stages of the NGO algorithm to further improve the algorithm's ability to escape local optima.

[0035] In step 202, the key state parameter prediction model is trained based on a pre-set dataset, specifically as follows: 1. Data preparation: Collect a large number of data pairs (X, Y) containing the input time series and corresponding real labels from historical databases or previous system trial runs. Obtaining the true label Y: The true solid-to-gas ratio can be determined by accurately weighing the conveyed material during a specific experimental period and combining this with the air volumetric flow rate measured by a flow meter, using the formula μ=M. solid / (ρ air *Q air These data points were calculated. Although sparse, they are sufficient for supervised learning. Pipeline Deposition True Label: Data segments that occurred some time before a clearly marked pipe blockage event in the historical record are marked as deposition state; data segments that are running stably and efficiently in the system are marked as normal state.

[0036] 2. VMD-Improved AGNGO-Transformer-GRU Model Training Process: Step 1: Perform VMD decomposition on all input time series to construct an enhanced training dataset.

[0037] Step 2: Divide the dataset into training, validation, and test sets (e.g., 7:2:1).

[0038] Step 3: Start the improved AGNGO optimizer.

[0039] Each individual AGNGO (eagle) represents a set of hyperparameter configurations.

[0040] Fitness function: For each hyperparameter, construct a corresponding Transformer-GRU model, train it on the training set, and evaluate its performance on the validation set. Fitness value. Fitness = w1 * (1 - MSE) μ ) + w2 * Accuracy P +w3*(1-Loss) total MSE μ It is the mean square error of the solid-gas ratio prediction, Accuracy P It is the accuracy of sedimentary state classification, Loss total It is the overall loss function, w1, w2, and w3 are weights, and the goal is to maximize the fitness function; AGNGO iterates continuously through its improved strategies to find the hyperparameter combination that yields the highest fitness. Step 4: Fix the optimal hyperparameter configuration found by AGNGO, retrain the final VMD-Transformer-GRU model with all training data (training set + validation set), and evaluate the final performance on the test set.

[0041] 3. Online real-time diagnosis and adaptive learning Deployment: Deploy the best offline-trained model to the inference engine of the industrial control computer; Real-time inference: During system operation, sensor data from the most recent T time steps is collected in real time. After undergoing the same VMD preprocessing, the data is input into the model, and μ is output in real time. hat and P deposition .

[0042] Adaptive fine-tuning: To address the slow drift of system characteristics after long-term operation (such as equipment aging or changes in material properties), an online learning loop can be established. When the system is in a confidently stable operating condition, the current data and the corresponding operating variables (which can serve as approximate stable state labels) can be used to form new samples, and the model can be fine-tuned periodically to maintain high accuracy.

[0043] In step 3, based on the improved near-end strategy optimization PPO algorithm, an intelligent decision controller is constructed to generate optimal control commands according to the state of the pneumatic conveying system, specifically: Define the state space of reinforcement learning, whose state vector includes the fan outlet source pressure, the conveying pipeline inlet pressure, the conveying pipeline outlet pressure, the total system pressure difference, the air volume flow rate, the fan operating frequency, the fan operating current, the feeder operating speed, the solid-to-gas ratio prediction value, the pipeline deposition state identification probability value, and the control command history value of the previous moment. Define the dynamic space of reinforcement learning, whose dynamic vectors include the fan frequency adjustment amount and the feeder speed adjustment amount. Action execution: The industrial control computer adds the calculated action to the set value of the previous moment to obtain a new control command, and sends it to the actuator. Define the reward / punishment function for reinforcement learning, where the reward function R is... t At time t, it is defined as: R t =α*Efficiency t -β*Energy t -γ*Risk t -δ*Action t (1) In the formula, Efficiency t The incentive for efficient delivery is designed to encourage the system to maximize its delivery capacity while maintaining stability. Efficiency t =k1*(Rfeeder (t) / R feeder_max (2) In the formula, R feeder (t) is the current rotational speed of the feeder, which is positively correlated with the conveying capacity; R feeder_max This is the maximum safe speed of the feeder, used for normalization; Energy t Energy consumption penalty, used to penalize high-energy-consuming operation, is as follows: Energy t =k2*(I fan (t)*V*F fan (t) / F fan_rated ) / Power max (3) In the formula, F fan_rated For the rated frequency of the fan, Power max This is the maximum power, used for normalization. This parameter guides the agent to find the operating point with the highest energy efficiency. Risk t The penalty for pipe blockage risk is used to severely punish situations with a high risk of pipe blockage. Risk t =k3*(P deposition (t) + k4*H(ΔP) total (t)-ΔP threshold )) (4) In the formula, P deposition (t) The soft measurement from step 2 is the core risk indicator. A hard judgment based on pressure difference is added: H is a step function, and when the total system pressure difference ΔP... total (t) Exceeds the safety threshold ΔP threshold This provision takes effect at that time, providing additional risk penalties; Action t To control smooth action penalty, which penalizes excessively large or rapid control actions, ensures stable system operation, and reduces actuator wear, the absolute values ​​of the actions are normalized and then summed, resulting in: Action t =k5*(|ΔF fan | / ΔF max +|ΔR feeder | / ΔR max (5) In the formula, α, β, γ, and δ are the main weighting coefficients, and k1, k2, k3, k4, and k5 are proportional coefficients used for dimensional normalization and adjusting the degree of influence of sub-items. These coefficients need to be tuned through domain knowledge and simulation experiments. Due to its excellent stability and sample efficiency, the PPO algorithm is adopted, and both the policy network (Actor) and the value network (Critic) use deep neural networks. Network structure: Input layer: Dimensions and state vector s t Consistent.

[0044] Hidden layers: 3 fully connected layers, each with 128 neurons, using the ReLU activation function.

[0045] Output layer: Policy Network (Actor): Outputs two actions [ΔF] fan ΔR feeder The network outputs the mean (μ) of the logarithmic standard deviation (log) and an independent adaptive log-standard deviation. std ), used to construct a Gaussian distribution of actions in order to achieve a balance between exploration and exploitation; Value Network (Critic): Outputs a scalar value V(s) t ), representing the long-term expected return of the current state; PPO objective function: (1) Probability ratio: r t (θ) = π θ (a) t |s t ) / π θ_old (a) t |s t ); Where, π θ This is the current strategy (network parameter θ), π θ_old It's the old sampling strategy; (2) Dominance function estimation: t =R t +γ*V(s {t+1} )-V(s t ) Among them, R t It is the immediate reward, γ is the discount factor (e.g., 0.99), and V(s) is the output of the value network; (3) PPO pruning objective function: L t (θ) = E t [min(r t (θ)Â t ,clip(r t (θ), 1-ε, 1+ε)Â t (6) This function uses a clip operation to prevent excessively large step sizes in a single policy update, thus ensuring training stability. ε is a hyperparameter (e.g., 0.2).

[0046] The training and online learning process is described below: 1. Simulation Environment Construction: A high-fidelity pneumatic conveying system simulation environment is built on an industrial control computer for the early-stage safety trial and error learning of the intelligent agent; 2. Offline training: The intelligent agent interacts with the model in the simulation environment, collecting a large amount of empirical data (s t a t R t s {t+1} And store it in the experience replay buffer; Randomly sample small batches of data from the buffer, and use the Adam optimizer to alternately update the parameters of the policy network (Actor) and the value network (Critic) according to the PPO objective function described above; Through millions of iterations, the agent gradually learns to make optimal decisions in complex, coupled pneumatic transport systems. 3. Online deployment and adaptive learning: Deploy the offline-trained policy network onto the industrial control computer of the actual system; When the system is running, the agent determines the actual state s. t Output action a t To control physical systems.

[0047] The method further includes: Step 4: Establish a closed-loop control process, cyclically execute state perception, intelligent decision-making, instruction execution and limit protection, and experience data storage, establish an online adaptive learning mechanism, and periodically or triggerively use real experience data collected during system operation to fine-tune the policy network and value network.

[0048] This invention provides an embodiment in which a large tire manufacturing enterprise in Shandong Province introduces the intelligent pneumatic conveying system described in this invention into its intelligent compound rubber production workshop. This system is used for the automated and precise conveying of various powder materials such as carbon black, silica, and zinc oxide from the auxiliary equipment system to the internal mixer.

[0049] The company's original conveying system used traditional PLC control, which faced several core pain points: powders with different specific gravities and flowability were prone to sedimentation and blockage during the conveying process, leading to production interruptions; in order to ensure continuous material supply, the blower was often in a high-power operation state, resulting in huge energy consumption; the accuracy of small material proportions was greatly affected by the stability of the conveying process; and when changing material types, parameters needed to be manually adjusted repeatedly.

[0050] After applying this invention, a complete sensor network is first deployed at key nodes of the conveying pipeline to monitor the pressure, flow rate, and equipment operating status of each feed line in real time. The system utilizes the initially collected conveying data to specifically train the VMD-improved AGNGO-Transformer-GRU model, enabling it to accurately predict changes in the conveying status of different materials. In particular, it can provide early warning of pipe blockage risks for easily adsorbed and agglomerated materials such as carbon black.

[0051] In actual operation, when the system is conveying silica, the soft sensing module detects a decrease in flowability caused by changes in material moisture. The intelligent decision-making module immediately and dynamically adjusts the fan frequency and feeding speed of that pipeline, maintaining the conveying volume while mitigating the risk of pipe blockage. Multiple conveying pipelines throughout the internal mixing workshop are coordinated and scheduled through the same intelligent system, achieving autonomous optimization of conveying parameters for different materials.

[0052] After six months of continuous operation, statistics show that the system's overall energy consumption has decreased by 22%, the pipe blockage failure rate has decreased by 95%, and the timing and weight accuracy of small material delivery have improved by 30% and 15%, respectively. Through its online learning function, the system can quickly adapt when switching between different rubber compound formulations, reducing changeover and debugging time from several hours to minutes. This successful application provides key technical support for enterprises to build fully automated and intelligent rubber mixing workshops.

[0053] This embodiment fully demonstrates the effectiveness of the present invention in a real industrial environment. It not only solves the pain points of traditional pneumatic conveying systems, but also brings significant economic benefits and operational convenience to enterprises. The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0054] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A pneumatic conveying system for powdered materials, characterized in that, include: Physical execution unit, sensor network unit, and intelligent control unit; The physical execution unit includes an air supply device, a material supply device, and a gas-solid separation device connected in sequence via a conveying pipeline; The sensor network unit is deployed at key nodes of the physical execution unit to collect system operating parameters in real time. The intelligent control unit is communicatively connected to the sensor network unit and the physical execution unit. The intelligent control unit includes: The data acquisition and processing module is used to receive and process data from the sensor network unit; The critical state parameter prediction module is used to predict the critical state parameters of the pneumatic conveying system based on the processed data using a VMD-improved AGNGO-Transformer-GRU hybrid model. The intelligent decision-making module is used to generate optimal control commands based on the system status.

2. The system according to claim 1, characterized in that, The sensor network unit includes: A pressure sensor installed at the outlet of the gas supply device is used to monitor the outlet pressure of the blower. A pressure sensor is installed at the inlet of the conveying pipe after the feeding device to monitor the inlet pressure of the pipe; A pressure sensor is installed at the outlet of the delivery pipeline before the gas-solid separation device to monitor the pipeline outlet pressure; A gas flow meter installed at the outlet of the gas supply device is used to monitor the volumetric flow rate of air. A current sensor connected to the air supply unit is used to monitor the operating current of the fan; A frequency acquisition device connected to the gas supply unit is used to monitor the operating frequency of the fan; A speed sensor connected to the feeding device is used to monitor the operating speed of the feeder.

3. The system according to claim 2, characterized in that, The intelligent decision-making module includes: State, action, and reward / penalty construction unit, used to construct state vectors, action vectors, and reward / penalty functions; The strategy network unit uses the near-end strategy optimization PPO algorithm to output the fan frequency adjustment and feeder speed adjustment. Value network unit, used to evaluate the state value function.

4. A control method for a pneumatic conveying system for powdered materials, applied to the pneumatic conveying system for powdered materials as described in any one of claims 1-3, characterized in that, include: Step 1: Collect monitoring data from the pneumatic conveying system; Step 2: Based on the VMD-improved AGNGO-Transformer-GRU hybrid model, predict the key state parameters of the pneumatic conveying system; Step 3: Based on the improved near-end strategy optimization PPO algorithm, construct an intelligent decision controller to generate optimal control commands according to the state of the pneumatic conveying system.

5. The method according to claim 4, characterized in that, In step 1, monitoring data of the pneumatic conveying system is collected, specifically as follows: Based on the sensor network unit, the monitoring data of the pneumatic conveying system is collected to generate a multi-sensor time series data vector, including the blower outlet source pressure, the conveying pipeline inlet pressure, the conveying pipeline outlet pressure, the total system pressure difference, the air volume flow rate, the blower operating frequency, the blower operating current, and the feeder operating speed.

6. The method according to claim 5, characterized in that, In step 2, the key state parameters of the pneumatic conveying system are predicted based on the VMD-improved AGNGO-Transformer-GRU hybrid model, specifically: Construct a key state parameter prediction model based on the VMD-improved AGNGO-Transformer-GRU hybrid model; The key state parameter prediction model is trained based on a pre-set dataset; The multi-sensor time series data vector is input into the key state parameter prediction model to obtain the key state parameters, which include the predicted solid-gas ratio and the pipeline deposition state identification probability value.

7. The method according to claim 6, characterized in that, In step 3, based on the improved near-end strategy optimization PPO algorithm, an intelligent decision controller is constructed to generate optimal control commands according to the state of the pneumatic conveying system, specifically: Define the state space for reinforcement learning, whose state vector includes the fan outlet source pressure, the conveying pipeline inlet pressure, the conveying pipeline outlet pressure, the total system pressure difference, the air volume flow rate, the fan operating frequency, the fan operating current, the feeder operating speed, the solid-to-gas ratio prediction value, and the pipeline deposition state identification probability value. Define the dynamic space of reinforcement learning, whose dynamic vectors include the fan frequency adjustment and the feeder speed adjustment; Define the reward / punishment function for reinforcement learning, where the reward function R is... t At time t, it is defined as: R t =α*Efficiency t -β*Energy t -c*Risk t -d*Action t (1) In the formula, Efficiency t Energy t To incur energy consumption penalties, Risk t Action to punish the risk of pipe blockage t To control the smoothness penalty of actions, α, β, γ, and δ are the main weighting coefficients; A policy network and a value network based on the Proximal Policy Optimization (PPO) algorithm are constructed. Through training, the agent learns a control policy that maximizes the cumulative reward. The policy network outputs the mean of two action vectors and an independent adaptive log standard deviation to construct a Gaussian distribution of actions, achieving a balance between exploration and exploitation. The value network outputs a scalar value to represent the long-term expected reward of the current state.

8. The method according to claim 7, characterized in that, The method further includes: Step 4: Establish a closed-loop control process, cyclically execute state perception, intelligent decision-making, instruction execution and limit protection, and experience data storage, establish an online adaptive learning mechanism, and periodically or triggerively use real experience data collected during system operation to fine-tune the policy network and value network.

Images