An intelligent drug dosing system and method based on machine learning

Abstract

The present invention discloses an intelligent chemical dosing system and method based on machine learning, comprising the steps of: S1. Real-time collecting the operation parameters of the fluidized bed and water quality characteristic data, including hydrodynamic parameters, water quality index parameters and particle morphology parameters, and constructing a time-series feature database; S2. Using the data in the time-series feature database to construct a dynamic chemical dosing model based on the fluidized bed to jointly predict the chemical type, dosing amount and operation parameters; S3. According to the prediction results output by the model, controlling the chemical dosing amount, chemical concentration, water flow rate and aeration intensity; S4. Real-time monitoring the fluidized bed and chemical dosing control parameters, and dynamically correcting the fluidized bed parameters in combination with the turbidity of the fluidized bed effluent and the particle sedimentation efficiency, aiming to solve the problems of poor sewage treatment effect in the prior art, which may lead to waste of chemicals, unqualified treatment and secondary environmental pollution.

Description

Technical Field

[0001] The present invention relates to the technical field of water treatment, particularly to the technical field of sedimentation of suspended impurities, and specifically to an intelligent dosing system and method based on machine learning. Background Art

[0002] Currently, there are increasingly high requirements for the safety of drinking water quality, and more precise dosing methods are needed to ensure that the water quality meets the standards.

[0003] In the field of water treatment, traditional dosing methods have problems. For example, it is difficult to accurately control the dosage of the medicament, resulting in waste of the medicament or poor treatment effect; it is impossible to adjust the dosing strategy in real time according to water quality changes, and the adaptability is poor. The development of machine learning technology provides new ideas and methods for the research of intelligent dosing systems. By establishing a data-driven model, intelligent optimization and control of the dosing process can be achieved.

[0004] The intelligent dosing method can improve the treatment efficiency and ensure the continuity and stability of the water treatment process through real-time monitoring and automatic adjustment. Currently, the intelligent dosing technology mainly collects water quality data through sensors, and then uses intelligent algorithms to automatically adjust the dosage of the medicament according to these data to achieve intelligent control. However, the feedback system of the existing technology cannot comprehensively reflect various factors in the water treatment process, resulting in inaccurate and untimely adjustment of the dosing strategy.

[0005] Therefore, it is necessary to improve the intelligent dosing system and method based on machine learning in the existing technology to solve the above problems. Summary of the Invention

[0006] The present invention overcomes the deficiencies of the existing technology and provides an intelligent dosing system and method based on machine learning, aiming to solve the problems of poor sewage treatment effect, possible waste of medicament, unqualified treatment, and secondary environmental pollution in the existing technology.

[0007] To achieve the above object, the technical solution adopted by the present invention is: an intelligent dosing method based on machine learning, including:

[0008] S1. Real-time collect the operating parameters of the fluidized bed and water quality characteristic data, including: hydrodynamic parameters, water quality index parameters, and particle morphology parameters, and construct a time series feature database;

[0009] S2. Use the data in the time series feature database to construct a dynamic dosing model based on the fluidized bed, and jointly predict the type of medicament, dosage, and operating parameters;

[0010] S3. According to the prediction results output by the model, control the dosage of the medicament, the concentration of the medicament, the water flow rate, and the aeration intensity;

[0011] S4. Monitor the fluidized bed and dosing control parameters in real time, dynamically correct the fluidized bed parameters by combining the turbidity of the fluidized bed effluent and the particle sedimentation efficiency, and achieve optimal dosing regulation.

[0012] In a preferred embodiment of the present invention, a microporous injection device is arranged at the bottom of the fluidized bed to directly inject the medicament into the dense area of particle fluidization, and a multi-point dosing device is arranged at the water inlet of the fluidized bed and different positions of the bed body;

[0013] The medicament is a sustained-release medicament, and the sustained-release medicament is a coated particle, and the coating material is one of poly(lactic-co-glycolic acid), chitosan-sodium alginate composite film and mesoporous silica coating;

[0014] According to the fluid velocity distribution of the fluidized bed, uncoated medicament is added in the high-speed area at the bottom, and the sustained-release medicament is added in the low-speed area at the top.

[0015] In a preferred embodiment of the present invention, the control of the medicament concentration:

[0016] Calculate the particle growth rate by real-time monitoring of the particle size distribution, the roundness of the aggregates and other particle morphology parameters, where d2 and d1 are the median particle sizes of the particles at different times, and t is the time interval; when the particle growth rate is less than the set slow growth threshold r1, increase the medicament concentration, and the increased concentration is proportional to r1 - r; when the particle growth rate is greater than the set fast growth threshold r2, decrease the medicament concentration, and the decreased concentration is proportional to r - r2; according to the change of the particle growth rate, adjust the medicament concentration every 10 - 15 minutes.

[0017] In a preferred embodiment of the present invention, the coordinated regulation of the water flow velocity and the aeration intensity:

[0018] Determine the aeration intensity according to the dissolved oxygen concentration of the water quality index parameters and the local turbulence intensity of the water flow dynamics parameters; when the dissolved oxygen concentration DO is less than the dissolved oxygen lower limit DO1, and the local turbulence intensity I t is less than the turbulence intensity lower limit I1, increase the aeration intensity, A = A0 + k3×(DO1 - DO) + k4×(I1 - I t ), where A0 is the initial aeration intensity, and k3 and k4 are the corresponding weights; when the dissolved oxygen concentration DO is greater than the dissolved oxygen upper limit DO2, and the local turbulence intensity I t is greater than the turbulence intensity upper limit I2, decrease the aeration intensity, A = A0 - k5×(DO - DO2) - k6×(I t - I2), where k5 and k6 are the corresponding weights.

[0019] In a preferred embodiment of the present invention, in step S2, the specific construction steps of the dynamic dosing model:

[0020] S21. Integrate historical data into the time-series feature database as the data source for the model, and combine hydrodynamic parameters, water quality indicators, and particle morphology parameters into a 12-dimensional feature vector.

[0021] S22. Implement dynamic joint optimization of chemical dosing strategies and fluidized bed operations using the Actor-Critic framework.

[0022] S23. Conduct transfer learning by combining pre-training with fine-tuning of actual data.

[0023] In a preferred embodiment of the present invention, the coating layer has a radial pore gradient, and the porosity decreases from 50% to 10% from the inside to the outside.

[0024] In a preferred embodiment of the present invention, the said S4 includes: combining the real-time operating state of the fluidized bed, including turbidity and sedimentation efficiency, for dynamic optimization:

[0025] Determine the hydraulic retention time according to water quality indicators and particle sedimentation rate. The hydraulic retention time refers to the average time that sewage stays in the fluidized bed. When Tur is greater than the turbidity upper limit Tur1 and the particle sedimentation rate S is less than the sedimentation rate lower limit S1, increase the hydraulic retention time, HRT = HRT0 + k7×(Tur - Tur1) + k8×(S1 - S), where HRT0 is the initial hydraulic retention time, and k7, k8 are the corresponding weights; when Tur is less than the turbidity lower limit Tur2 and the particle sedimentation rate S is greater than the sedimentation rate upper limit S2, reduce the hydraulic retention time, HRT = HRT0 - k9×(Tur2 - Tur) - k 10 ×(S - S2), where HRT0 is the initial hydraulic retention time, and k9, k 10 are the corresponding weights.

[0026] In a preferred embodiment of the present invention, the carrier material of the particles in the fluidized bed is a mixture of magnetite Fe3O4 and pyrite FeS2, and the mass ratio is 3 - 5:1; constructing a hierarchical pore structure through chemical modification includes: constructing micropores by hydrothermal synthesis, mesopores by sol-gel method, and macropores by 3D printing technology; the micropore diameter is 0.3 - 0.4 nm, the mesopore diameter is 5 - 10 nm, and the macropore diameter is 50 - 200 μm.

[0027] In a preferred embodiment of the present invention, functional groups are loaded on the surface of the particles in the fluidized bed by surface grafting technology. Amino groups are grafted on the surface of the carrier material by chemical vapor deposition, and -SH groups are introduced by impregnating the particles with a thioacetic acid solution.

[0028] The present invention provides an intelligent chemical dosing system based on machine learning, including:

[0029] A multi-modal sensing network module for real-time collection of fluidized bed operation parameters and water quality characteristic data;

[0030] A data processing module for cleaning, synchronizing and feature extraction of raw data to construct a time series feature database;

[0031] A deep reinforcement learning decision module for dynamically optimizing chemical dosing strategies and fluidized bed operation parameters;

[0032] A chemical dosing control module for combinatorial control of chemical dosing amount, chemical concentration and water flow rate according to the prediction results output by the model;

[0033] A chemical dosing module for performing chemical dosing operations according to control parameters;

[0034] A closed-loop feedback and adaptive module for real-time collection of turbidity data at the outlet of the fluidized bed and sedimentation rate data of particles in the sedimentation tank and dynamically correcting operation parameters.

[0035] The present invention solves the defects existing in the background technology, and the present invention has the following beneficial effects:

[0036] (1) The present invention proposes an intelligent chemical dosing system and method based on machine learning. By using multi-modal sensors to collect fluidized bed operation parameters and water quality characteristic data in real time, a time series feature database is constructed. The deep reinforcement learning model is used to jointly predict chemical types, dosing amounts and operation parameters, and according to the prediction results output by the model, the chemical dosing amount, chemical concentration, water flow rate and aeration intensity are dynamically adjusted to achieve optimal chemical dosing control. The system dynamically corrects by real-time monitoring of the turbidity of the fluidized bed effluent and the particle sedimentation efficiency, combined with the Q-learning algorithm to ensure the treatment effect; through the coating layer design of the sustained-release chemical and the coordinated dosing of the flow field velocity distribution, combined with the multi-stage pore particle carrier, the present invention realizes the two-stage release and efficient reaction of the chemical, improves the adsorption and sedimentation efficiency of the particles, thereby improving the efficiency and quality of sewage treatment and reducing the risk of chemical waste and secondary pollution.

[0037] (2) Through the coordinated dosing of the coating layer design of the sustained-release agent and the flow field velocity distribution, and in combination with the multi-stage pore particle carrier, the present invention realizes the two-stage release of the agent and efficient reaction. Specifically, the uncoated agent is dosed in the high-speed area at the bottom to take effect quickly, and the sustained-release agent (the porosity of the coating layer decreases from 50% to 10% from the inside to the outside) is dosed in the low-speed area at the top. Combining the mass transfer advantages of the carrier micropores (0.3 - 0.4 nm), mesopores (5 - 10 nm) and macropores (50 - 200 μm), it promotes the diffusion of the agent and the adsorption of pollutants. Moreover, multi-point dosing can ensure the uniform distribution of the agent in the fluidized bed and enhance the adsorption and removal ability of the particles to the agent. The agent adheres to the surface of the fluidized particles, forming larger particle aggregates for sedimentation, and finally realizing the separation from water. Compared with the prior art, it solves the problem of instantaneous dilution of the agent, prolongs the action time, and further reduces the risk of agent waste and secondary pollution.

[0038] (3) Through the real-time monitoring of the particle growth rate and the dynamic adjustment of the agent concentration, and in combination with the closed-loop feedback of the Q-learning algorithm, the present invention realizes the adaptive optimization of the dosing strategy and outputs the optimal solution. Specifically, the growth rate is calculated based on the particle size distribution and roundness. When the rate is lower than the threshold, the concentration is increased proportionally, and vice versa. And the error function, the mean square error of turbidity and sedimentation rate, is used to drive the Q-learning to update the network parameters. Compared with the prior art, it overcomes the problem of lag in traditional system adjustment and further improves the dynamic response speed of agent concentration control and the stability of treatment effect.

[0039] (4) Through the coordinated regulation of the dissolved oxygen concentration and the local turbulence intensity, and in combination with the dynamic correction of the hydraulic retention time, the present invention optimizes the matching relationship between the aeration intensity and the water flow velocity. Specifically, when the dissolved oxygen is lower than the lower limit and the turbulence is insufficient, the aeration intensity is increased, and vice versa. At the same time, the hydraulic retention time is adjusted according to the turbidity and sedimentation rate. Compared with the prior art, it solves the contradiction that high turbulence makes it difficult for particles to settle, further improves the solid-liquid separation efficiency and reduces the energy consumption.

[0040] (5) Through the combination of surface grafting technology, the loading of amino groups and -SH groups on the multi-stage pore carrier material, the catalytic adsorption ability of the particles is strengthened. The amino groups are grafted by chemical vapor deposition to enhance the binding force of the flocculant, and -SH groups are introduced by thioacetic acid impregnation to capture heavy metals, combining the selective adsorption of micropores, the accelerated mass transfer of mesopores and the reduced resistance of macropores. Compared with the prior art, it breaks through the limitation of the single function of traditional carriers, further realizes the synergistic removal of composite pollutants, and significantly improves the versatility and adaptability of sewage treatment. Description of the Drawings

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments described in the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings;

[0042] Figure 1 is a flowchart of a preferred embodiment of the present invention;

[0043] Figure 2 is a structural diagram of a sustained-release agent of a preferred embodiment of the present invention;

[0044] Figure 3 is a schematic diagram of the device structure of a preferred embodiment of the invention;

[0045] As shown in the figure: 1. Coating layer; 2. Sustained-release layer; 3. Quick-release layer; 4. Fluidized bed; 5. Granules; 6. Chemical dosing device; 7. Water inlet. Detailed implementation manners

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0047] Many specific details are set forth in the following description in order to fully understand the present invention, but the present invention can also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited by the specific embodiments disclosed below.

[0048] Application overview:

[0049] Traditional intelligent chemical dosing systems mainly focus on adjusting the dosing amount and frequency of chemical agents according to water quality parameters. After the chemical agents are dosed, it takes a certain amount of time for mixing and reaction to make the chemical agents fully combine with the impurities in the water to form flocs that are easy to settle. After the mixing reaction, the impurities in the water combine with the chemical agents to form larger particles, and these particles settle in the sedimentation tank under the action of gravity to achieve solid-liquid separation;

[0050] However, the importance of the physico-chemical process is ignored, which limits its ability to optimize the sewage treatment effect, resulting in the chemical agents not fully reacting with the impurities to form flocs that are easy to settle, thus affecting the solid-liquid separation effect. In addition, insufficient reaction may also lead to chemical agent residues, causing secondary pollution to the environment.

[0051] Therefore, the present application proposes a sewage treatment solution for intelligent chemical dosing combined with a granulating fluidized bed, which can monitor the water flow condition and the state of particulate medium in real time, and use machine learning algorithms to find the optimal operating conditions, such as water flow rate, chemical types and concentrations, etc., so as to achieve precise dosing and uniform distribution of chemicals, promote the formation and growth of particulate aggregates, improve their sedimentation rate and effect, not only can shorten the treatment time, reduce the risk of secondary pollution caused by poor sedimentation, but also can further improve the efficiency and quality of sewage treatment.

[0052] Exemplary method:

[0053] Such as Figure 1 As shown, a machine learning-based intelligent chemical dosing method includes the steps of:

[0054] S1. Real-time collect the operating parameters of the fluidized bed and the water quality characteristic data, including: hydrodynamic parameters, water quality index parameters, and particulate morphology parameters, and construct a time series feature database;

[0055] S2. Use the data in the time series feature database to construct a dynamic chemical dosing model based on the fluidized bed, and jointly predict the chemical type, dosing amount, and operating parameters;

[0056] S3. According to the prediction results output by the model, control the chemical dosing amount, chemical concentration, water flow rate, and aeration intensity;

[0057] S4. Real-time monitor the fluidized bed and the chemical dosing control parameters, and dynamically correct the fluidized bed parameters in combination with the turbidity of the fluidized bed effluent and the particulate sedimentation efficiency.

[0058] The granulating fluidized bed is a device that forms a fluidized state of solid particles under the action of a fluid. In the fluidized state, the particles exhibit properties similar to those of a fluid, such as fluidity, compressibility, etc. This technology was originally mainly used in the chemical and pharmaceutical industries, but now it is transferred and applied to sewage treatment in the present application. This is because the particles in the fluidized bed are in a suspended state, increasing the contact opportunity between the chemicals and the impurities in the sewage, which is beneficial to the flocculation reaction; the limitation of the traditional chemical dosing system is that it cannot adapt to water quality changes in real time, while the fluidized state characteristics of the granulating fluidized bed directly affect the particle formation efficiency, thereby improving the efficiency of adsorption sedimentation and chemical reactions, and significantly improving the treatment effect.

[0059] In step S1, the operating parameters of the fluidized bed and the water quality characteristic data are collected in real time through a multi-modal sensor array integrated in the water treatment system, and the original data is preprocessed to construct a time series feature database;

[0060] In step S1, the multi-modal sensor array includes a water quality sensor group, an optical sensor array, and an acoustic sensor group integrated and installed in the granulation fluidized bed system; the sensors are arranged in a spatial grid distribution to ensure the data coverage density in the three-dimensional space of the fluidized bed. The limitations of single sensors are eliminated through multi-source data fusion, providing a comprehensive process characterization.

[0061] The raw data includes: hydrodynamic parameters, water quality index parameters, and particle morphology parameters; the hydrodynamic parameters include: water flow velocity, aeration intensity, bed expansion rate, and local turbulence intensity; the water quality index parameters include: real-time pH value, turbidity, dissolved oxygen, and hardness ion concentration; the particle morphology parameters include: particle size distribution, aggregate roundness, surface roughness, and sedimentation rate;

[0062] Outliers are removed, noise is filtered, and standardization preprocessing is performed on the raw data to avoid sensor signal drift easily caused by the high-turbulence environment in the fluidized bed;

[0063] Through the timestamp synchronization technology, sensor data with different sampling frequencies are unified to the same time reference to construct a time-aligned raw data matrix; the sampling frequency of the water quality sensor is 1 Hz, and the sampling frequency of the optical sensor is 10 Hz. The two are aligned to 10 Hz using the linear interpolation method; since the acoustic sensor has non-uniform sampling, it is resampled to 10 Hz through the sliding window averaging method; through the above-mentioned time-axis alignment of the raw data, it is integrated into a structured matrix;

[0064] The row dimension of the matrix is the timestamp T = {t0, t1, …, t n}, with a total of N time points; the column dimension is the raw data parameters, with a total of M parameters;

[0065] Multi-dimensional time series data matrix: Among them, each row corresponds to the multi-modal observation value at a time point, and each column corresponds to the time series of a sensor parameter.

[0066] For each sensor data stream, the statistic is calculated based on a sliding window length of 30 s, and the data points exceeding the threshold are marked as abnormal and removed; according to the operating principle of the fluidized bed, the hard constraint conditions are: the water flow velocity v ∈ [0.5, 5.0] m / s, pH value ∈ [6.0, 9.0], and the data that does not conform to the physical feasible region is directly removed;

[0067] The wavelet transform combined with the Kalman filter algorithm is used to separate the high-frequency noise signal and retain the effective characteristic frequency band. The wavelet transform can decompose the signal into components of different frequencies, thus effectively separating the high-frequency noise; the Kalman filter algorithm can estimate the state of the signal in real time based on the predicted value and observed value of the signal, further improving the filtering effect; Z-score standardization is performed on the multi-source data Among them, x is the original data, μ is the mean of the data, and σ is the standard deviation of the data, eliminating the difference in dimension.

[0068] Store the standardized multi-modal data in a structured database and construct a time-series feature database; the timestamp is the primary key, and the hydrodynamic parameters are a 4-dimensional vector [v, Q a , H b , I t , which represent the water flow velocity, aeration intensity, bed expansion rate, and local turbulence intensity respectively; the water quality index parameters are a 4-dimensional vector [pH, Tur, DO, Ha], which represent the real-time pH value, turbidity, dissolved oxygen, and hardness ion concentration respectively; the particle morphology parameters are a 4-dimensional vector [D 50 , C, R, S], which represent the median particle size, aggregate roundness, surface roughness, and sedimentation rate of the particle size respectively.

[0069] The preprocessing in step S1 eliminates sensor noise and dimension differences, and the time-aligned multi-dimensional data matrix can be used as the direct input of the neural network;

[0070] In step S2, the specific construction steps of the dynamic dosing model:

[0071] S21. Integrate historical data into the time-series feature database as the model data source, and merge the hydrodynamic parameters, water quality indicators, and particle morphology parameters into a 12-dimensional feature vector;

[0072] S22. Implement the dynamic joint optimization of the chemical dosing strategy and fluidized bed operation with the Actor-Critic framework; the policy network outputs: discrete actions (chemical selection) and continuous actions (dose and flow rate adjustment) value networks, evaluate the state value, and optimize the strategy through the reward function by integrating the turbidity error, sedimentation rate error, and chemical cost;

[0073] S23. Perform transfer learning by combining pre-training with fine-tuning of actual data;

[0074] In step S21, by integrating historical operation data and real-time preprocessing data, a high-dimensional, multi-modal time-series feature database is constructed to provide diverse samples for model training and enhance the generalization ability of the model to complex working conditions.

[0075] In step S22, the Actor-Critic framework is a core architecture in reinforcement learning, combining the dual mechanisms of policy optimization and value evaluation;

[0076] The input feature of the input layer is a 12-dimensional fusion feature vector;

[0077] The network structure of the policy network Actor is a three-layer fully connected network, i.e., input layer → 256 → 128 → output layer; the discrete action in the output action space is the selection of chemical agent type, and the continuous actions are the dosing concentration and the adjustment amount of water flow velocity;

[0078] The prediction of chemical agent type uses the Softmax function to generate a multi-class probability distribution where z k is the activation value of the k-th neuron in the output layer, corresponding to the k-th chemical agent, and the chemical agent type with the highest selection probability is used as the dosing decision at the current moment;

[0079] The prediction of dosing amount uses the Tanh activation function to generate a normalized dosing adjustment amount, and then maps it to the actual dosing range through a linear transformation, so as to dynamically adjust the dosing amount of the chemical agent to adapt to water quality fluctuations; Prediction of dosing amount where a dose ∈[-1, 1] is the normalized value output by the network, a dose is the predicted value of the chemical agent dosing amount, Δ dose is the dynamic adjustment amount of the chemical agent concentration, is the normalization factor, which is used to map the adjustment amount to the ratio within the high and low ranges, and High and Low are the lower and upper limits of the dose;

[0080] The prediction of operating parameters is output through the continuous action output layer of the Actor network. The predicted operating parameter a flow = BaseFlow × (1 + a offset ), where a offset ∈[-0.1, 0.1] is the adjustment ratio output by the network, and BaseFlow is the current reference flow velocity;

[0081] The Actor network simultaneously outputs the discrete action of chemical agent type and the continuous actions of dosing amount and flow velocity, and shares the underlying features through multi-task learning to capture the dependency relationship between parameters;

[0082] The network structure of the value network Critic shares the input layer with the Actor, and the output is the state value estimation V(s), which is used to evaluate the value of the current state and guide the update of the Actor policy;

[0083] The Critic network guides the Actor network to balance the influence of the three types of actions by evaluating the comprehensive reward. The reward function is R t = α·Δ sedimentation efficiency + β·(-Δ chemical agent consumption) + γ·turbidity penalty term; Δ is the difference, and α, β, and γ are the corresponding weight coefficients. The turbidity penalty term is an error weight term used to quantify the deviation of turbidity from the target range in the dynamic dosing model. Set the upper and lower limits of turbidity, and drive the model to adjust parameters by punishing the situation where turbidity exceeds the standard, so that the turbidity returns to the safe range.

[0084] In step S23, the flow field data generated by the CFD-PBM coupling model is used to update the Actor-Critic network parameters through actual data, correcting the deviation between the simulation and the actual situation.

[0085] Dynamically adjust the learning rate η according to the performance of the model on the validation set t+1 = η t ·e -λ·误差率 where η t is the current learning rate and e -λ·误差率 is the decay factor.

[0086] In step S2, the fusion feature library of historical data and real-time data covers diverse working conditions, reducing the model's dependence on a single data source. The Actor-Critic framework realizes the coordinated control of chemical dosing and fluidized bed operation.

[0087] In step S3, the dosing amount, chemical concentration, water flow velocity, and aeration intensity are combinedly controlled.

[0088] As Figure 3 shown, due to the high turbulence of the fluidized bed, the chemical may be rapidly diluted, making it difficult to form an effective concentration gradient on the particle surface. Therefore, multi-point chemical dosing devices are set at the inlet of the fluidized bed and different positions of the bed body. Utilize the turbulent state of the fluidized bed to achieve rapid chemical diffusion, ensuring that the chemical can be evenly dispersed in the fluidized bed and fully contact with the granular medium;

[0089] Specifically, a microporous injection device is set at the bottom of the fluidized bed to directly inject the chemical into the dense area of particle fluidization. Multi-point chemical dosing devices are set at the inlet of the fluidized bed and different positions of the bed body, increasing the contact opportunity between the chemical and the granular medium and facilitating the flocculation reaction;

[0090] Chemical dosing amount:

[0091] For a fluidized bed with a diameter of D meters, 6, 5, and 4 chemical dosing points are evenly set on the circumferences at heights of 0.1D, 0.3D, and 0.5D from the bottom. Determine the chemical dosing amounts at different positions according to the flow field simulation. The dosing amount ratio of the bottom microporous injection device is 0.3 - 0.5, the dosing amount ratio of the inlet is 0.1 - 0.2, and the dosing amount ratio of the chemical dosing points at other positions of the bed body is 0.3 - 0.6. And each point is proportionally distributed according to factors such as the height from the bottom and the flow field velocity;

[0092] As Figure 2As shown, the medicament is a sustained-release medicament. Specifically, the sustained-release medicament is a coated granule, which further realizes the sustained release of the medicament after combination and avoids the rapid dilution of the medicament. The coating material can be a polymer with good biocompatibility and slow release performance, specifically one of poly(lactic-co-glycolic acid), chitosan-sodium alginate composite film, and mesoporous silica coating as the coating material. The coating structure is designed as a core-shell structure, with the inner layer being the rapid release layer and the outer layer being the sustained release layer to achieve the two-stage release of the medicament. And the coating layer has a radial pore gradient, and the porosity decreases from 50% to 10% from the inside to the outside to match the turbulence intensity in different regions of the fluidized bed.

[0093] According to the flow field velocity distribution simulated by CFD, uncoated medicament is added in the high-speed area at the bottom to achieve rapid onset, and the sustained-release medicament is added in the low-speed area at the top to achieve the effect of prolonging the action time, prolonging the action time of the medicament and avoiding instantaneous dilution.

[0094] Medicament concentration control:

[0095] Specifically, the particle growth rate is calculated by real-time monitoring of particle size distribution, roundness of aggregates and particle morphology parameters. Among them, d2 and d1 are the median particle sizes of particles at different times, and t is the time interval; when the particle growth rate is less than the set slow growth threshold r1, the medicament concentration is increased, and the increased concentration is proportional to r1 - r; when the particle growth rate is greater than the set fast growth threshold r2, the medicament concentration is decreased, and the decreased concentration is proportional to r - r2; according to the change of the particle growth rate, the medicament concentration is adjusted every 10 - 15 minutes.

[0096] Water flow velocity and aeration intensity regulation:

[0097] The aeration intensity is determined according to the dissolved oxygen concentration of the water quality index parameters and the local turbulence intensity of the water flow dynamics parameters. The dissolved oxygen concentration is the amount of dissolved oxygen, and the local turbulence intensity refers to the degree of turbulence in the sewage fluidized bed, which is represented by the root mean square value of the water flow velocity fluctuation. High turbulence intensity helps the medicament and impurities in the water to be fully mixed and improves the reaction efficiency. However, too high turbulence intensity may cause particles to be difficult to settle, and it is necessary to adjust the water flow velocity and aeration intensity to optimize the sedimentation conditions; when the dissolved oxygen concentration DO is less than the dissolved oxygen lower limit DO1 and the local turbulence intensity I t is less than the turbulence intensity lower limit I1, the aeration intensity is increased, A = A0 + k3×(DO1 - DO) + k4×(I1 - I t ), where A0 is the initial aeration intensity, and k3 and k4 are the corresponding weights; when the dissolved oxygen concentration DO is greater than the dissolved oxygen upper limit DO2 and the local turbulence intensity I t is greater than the turbulence intensity upper limit I2, the aeration intensity is decreased, A = A0 - k5×(DO - DO2) - k6×(It -I2), where k5 and k6 are corresponding weights.

[0098] In step S4,

[0099] Online monitoring:

[0100] Real-time collect the turbidity data at the outlet of the fluidized bed and the sedimentation rate data of the particles in the sedimentation tank, compare the real-time collected turbidity and sedimentation rate data with the model prediction results, and calculate the error function;

[0101] Execution feedback:

[0102] For the predicted turbidity value Tur and the actual value Tur real , the error function can be expressed as the mean square error where i is at a certain time point; for the predicted sedimentation rate value S and the actual value S real Use the same mean square error to obtain Es;

[0103] The total error function comprehensively considers the errors of turbidity and sedimentation rate: E = α·Er + β·Es, where α and β are the weights corresponding to the errors;

[0104] Dynamic correction:

[0105] According to the feedback of the real-time monitoring data and the error function, use the Q-learning algorithm to update the value network parameters, learn the optimal strategy by executing different parameter regulations and according to the processing effects, and finally output the optimal regulation plan;

[0106] Define the state space Sp as a multi-dimensional time series data matrix, including hydrodynamic parameter, water quality index parameter, and particle morphology parameter. The action space A is a combination of chemical agent types, dosing doses, and operation parameters; Q(Sp t , A t ) ← Q(Sp t , A t ) + η[R t+1 + γmaxQ(Sp t+1 , a) - Q(Sp t , A t )], where η is the learning rate, γ is the discount factor, R t+1 is the immediate reward obtained after executing the action A t in the state Sp t , and maxQ(Sp t+1 , a) is the maximum Q value of all possible actions a in the next state Sp t+1 ; through continuous iterative updates, the system can learn the optimal chemical dosing strategy under different water quality conditions.

[0107] Perform dynamic optimization in combination with the real-time operating status of the fluidized bed, such as turbidity and sedimentation efficiency.

[0108] Determine the hydraulic retention time according to water quality indicators and particle sedimentation rate. The hydraulic retention time refers to the average time that sewage stays in the fluidized bed. When Tur is greater than the turbidity upper limit Tur1 and the particle sedimentation rate S is less than the sedimentation rate lower limit S1, increase the hydraulic retention time, HRT = HRT0 + k7×(Tur - Tur1) + k8×(S1 - S), where HRT0 is the initial hydraulic retention time, and k7, k8 are the corresponding weights; when Tur is less than the turbidity lower limit Tur2 and the particle sedimentation rate S is greater than the sedimentation rate upper limit S2, reduce the hydraulic retention time, HRT = HRT0 - k9×(Tur2 - Tur) - k 10 ×(S - S2), where HRT0 is the initial hydraulic retention time, and k9, k 10 are the corresponding weights.

[0109] The particles in the fluidized bed are composite carrier materials with surface catalytic activity, prepared by constructing a hierarchical pore structure through chemical modification and loading functional groups by combining surface grafting technology.

[0110] The carrier material of the particles is a mixture of magnetite Fe3O4 and pyrite FeS2, with a mass ratio of 3 - 5:1; magnetite has magnetic recovery characteristics, and pyrite contains sulfur active sites, which are suitable for oxidation-reduction synergistic reactions;

[0111] Construction of hierarchical pore structure:

[0112] Generate ZIF-8 metal-organic framework on the carrier surface by hydrothermal synthesis method. The pore diameter is 0.3 - 0.4nm. The micropore diameter is small, which can selectively adsorb small molecules. The hydrothermal method uses zinc nitrate and 2-methylimidazole as precursors, the temperature is 120°C, and the growth time is more than 12h;

[0113] Deposit mesopores with a pore diameter of 5 - 10nm by sol-gel method to form a mass transfer channel, which is beneficial to the rapid diffusion and transmission of reactants and products. The sol-gel method uses TEOS as the silicon source and CTAB as the template agent to deposit mesoporous SiO2 on the surface of ZIF-8;

[0114] Construct macropores with a pore diameter of 50 - 200μm by 3D printing. The macropores can significantly reduce the resistance of the fluid in the fluidized bed and improve the overall mass transfer efficiency. By selecting appropriate carrier materials and constructing a hierarchical pore structure, a more efficient and environmentally friendly catalytic, adsorption, and separation process can be achieved;

[0115] On the basis of constructing a hierarchical pore structure, combined with surface grafting technology, which is a method of grafting specific functional groups or molecular segments onto the material surface through chemical bonding. Through surface grafting, more active sites and the hydrophilicity and hydrophobicity towards impurities and chemicals in sewage are introduced onto the surface of the carrier material;

[0116] Using the chemical vapor deposition method, with 3-aminopropyltriethoxysilane as the modifier, amino groups are grafted at 150 °C under nitrogen for 2 - 4 h. Through the chemical vapor deposition method, APTES can be evenly distributed on the particle surface and chemically bond with the active sites on the surface of the carrier material to form a stable amino - NH2 grafting layer. Amino groups have strong electrostatic adsorption ability and can produce electrostatic adsorption with the negatively charged groups in anionic flocculants, enhancing the binding force between particles and flocculants and improving the flocculation effect, thereby further improving the efficiency and quality of sewage treatment.

[0117] By impregnating the particles with a 0.5 M thioacetic acid solution, -SH groups are introduced. -SH groups have strong affinity and reactivity and can specifically react with various pollutants in sewage, such as forming stable sulfide precipitates with heavy metal ions to achieve effective removal of heavy metal ions. At the same time, -SH groups can also undergo addition reactions or redox reactions with heavy metal ions to form sulfide precipitates, enhancing the multifunctionality and adaptability of the particles in sewage treatment.

[0118] Through the hierarchical pore structure and surface functionalization modification of the above particles, the intelligent chemical dosing system in this application can achieve a more efficient and environmentally friendly sewage treatment process, significantly improving the efficiency and quality of sewage treatment.

[0119] The synergistic effect between the hierarchical pores of the particles and the slow - release chemicals. Uncoated chemicals are added in the high - speed area at the bottom of the fluidized bed, and slow - release chemicals are added in the low - speed area at the top. Combining with the hierarchical pore structure of the carrier, two - stage release of chemicals is achieved. Macropores reduce fluid resistance, mesopores accelerate chemical diffusion, and micropores selectively adsorb small - molecule pollutants, increasing the contact probability between chemicals and impurities;

[0120] Precipitates such as CaCO3 formed by the reaction of chemicals like lime with calcium ions in water preferentially nucleate and grow at the amino sites on the carrier surface to form stable aggregates, achieving the synergistic removal of composite pollutants by combining the precipitates generated by the chemicals.

[0121] Exemplary system:

[0122] An intelligent chemical dosing system based on machine learning, comprising:

[0123] A multimodal sensing network module for real - time collecting the operating parameters of the fluidized bed and water quality characteristic data;

[0124] A data processing module, which is used to clean, synchronize and extract features from the original data, and construct a time-series feature database;

[0125] A deep reinforcement learning decision-making module, which is used to dynamically optimize the chemical dosing strategy and fluidized bed operation parameters;

[0126] A chemical dosing control module, which is used to combinatorially control the chemical dosing amount, chemical concentration and water flow rate according to the prediction results output by the model;

[0127] A chemical dosing module, which is used to perform the chemical dosing operation according to the control parameters;

[0128] A closed-loop feedback and adaptive module, which is used to collect the turbidity data of the fluidized bed outlet and the sedimentation rate data of the particles in the sedimentation tank in real time and dynamically correct the operation parameters.

[0129] Based on the ideal embodiments of the present invention as an inspiration, through the above description, relevant personnel can completely make various changes and modifications without departing from the technical idea of the present invention. The technical scope of the present invention is not limited to the content in the specification, and the technical scope must be determined according to the scope of the claims.

Claims

1. A machine learning-based intelligent drug dosing method, characterized in that, Including the steps: S1. Real-time collect the operation parameters of the fluidized bed and the water quality characteristic data, including hydrodynamic parameters, water quality index parameters and particle morphology parameters, and construct a time series characteristic database; S2. Use the data in the time series characteristic database to construct a dynamic dosing model based on the fluidized bed, and jointly predict the type of reagent, the dosing dose and the operation parameters; S3. According to the prediction results output by the model, control the dosing amount of the reagent, the reagent concentration, the water flow rate and the aeration intensity; S4. Real-time monitor the fluidized bed and the dosing control parameters, and dynamically correct the fluidized bed parameters in combination with the turbidity of the fluidized bed effluent and the particle sedimentation efficiency to achieve optimal dosing regulation; A microporous injection device is arranged at the bottom of the fluidized bed to directly inject the reagent into the dense area of particle fluidization, and a multi-point dosing device is arranged at the water inlet and different positions of the bed body of the fluidized bed; The reagent is a sustained-release reagent, and the sustained-release reagent is a coated particle, and the coating material is one of poly(lactic-co-glycolic acid), chitosan-sodium alginate composite film and mesoporous silica coating; According to the velocity distribution of the fluidized bed flow field, the uncoated reagent is added in the high-speed area at the bottom, and the sustained-release reagent is added in the low-speed area at the top; The coating layer has a radial pore gradient, and the porosity decreases from 50% to 10% from the inside to the outside; The carrier material of the particles in the fluidized bed is a mixture of magnetite Fe3O4 and pyrite FeS2, and the mass ratio is 3-5:1; a hierarchical pore structure is constructed on the carrier surface through chemical modification, including: micropores are constructed by hydrothermal synthesis, mesopores are constructed by sol-gel method, and macropores are constructed by 3D printing technology; the micropore diameter is 0.3-0.4nm, the mesopore diameter is 5-10nm, and the macropore diameter is 50-200μm.

2. The intelligent drug dosing method based on machine learning according to claim 1, characterized in that: Reagent concentration control: Calculate the particle growth rate by real-time monitoring of particle size distribution and particle shape parameters such as the roundness of aggregates. Among them, d2 and d1 are the median particle sizes at different times, and t is the time interval; when the particle growth rate is less than the set slow growth threshold r1, increase the chemical concentration, and the increased concentration is proportional to r1 - r; when the particle growth rate is greater than the set excessive growth threshold r2, decrease the chemical concentration, and the decreased concentration is proportional to r - r2; adjust the chemical concentration every 10 - 15 minutes according to the change in the particle growth rate.

3. The intelligent drug dosing method based on machine learning according to claim 1, wherein: Coordinated regulation of water flow rate and aeration intensity: Determine the aeration intensity based on the dissolved oxygen concentration of the water quality index parameters and the local turbulence intensity of the hydrodynamic parameters; when the dissolved oxygen concentration DO is less than the dissolved oxygen lower limit DO1 and the local turbulence intensity I t is less than the turbulence intensity lower limit I1, increase the aeration intensity, A = A0 + k3×(DO1 - DO) + k4×(I1 - I t ), where A0 is the initial aeration intensity, and k3, k4 are the corresponding weights; when the dissolved oxygen concentration DO is greater than the dissolved oxygen upper limit DO2 and the local turbulence intensity I t is greater than the turbulence intensity upper limit I2, decrease the aeration intensity, A = A0 - k5×(DO - DO2) - k6×(I t - I2), where k5, k6 are the corresponding weights.

4. A machine learning-based intelligent drug dosing method according to claim 1, characterized in that: In step S2, the specific construction steps of the dynamic dosing model: S21. Integrate the historical data into the time series characteristic database as the model data source, and merge the hydrodynamic parameters, water quality indexes and particle morphology parameters into a 12-dimensional feature vector; S22. Implement the dynamic joint optimization of the reagent dosing strategy and the fluidized bed operation with the Actor-Critic framework; S23. Through the combination of pre-training and fine-tuning of actual data, perform transfer learning.

5. A machine learning-based intelligent drug dosing method according to claim 1, characterized in that: The said S4 includes: combining the real-time operation state of the fluidized bed, including turbidity and sedimentation efficiency, for dynamic optimization: Determine the hydraulic retention time according to the water quality index and the particle sedimentation rate. The hydraulic retention time refers to the average time that the sewage stays in the fluidized bed. When Tur is greater than the turbidity upper limit Tur1 and the particle sedimentation rate S is less than the sedimentation rate lower limit S1, increase the hydraulic retention time, HRT = HRT0 + k7×(Tur - Tur1) + k8×(S1 - S), where HRT0 is the initial hydraulic retention time and k7, k8 are the corresponding weights; when Tur is less than the turbidity lower limit Tur2 and the particle sedimentation rate S is greater than the sedimentation rate upper limit S2, decrease the hydraulic retention time, HRT = HRT0 - k9×(Tur2 - Tur) - k 10 ×(S - S2), where HRT0 is the initial hydraulic retention time and k9, k 10 are the corresponding weights.

6. The intelligent drug dosing method based on machine learning according to claim 1, wherein: Functional groups are loaded on the surface of the particles in the fluidized bed by particle surface grafting technology. Amino groups are grafted on the surface of the carrier material by chemical vapor deposition, and -SH groups are introduced by impregnating the particles with thioacetic acid solution.

Images