Disc centrifuge liquid-liquid separation intelligent control system based on artificial intelligence

Abstract

The invention discloses a disc centrifuge liquid-liquid separation intelligent control system based on artificial intelligence, and particularly relates to the technical field of disc centrifuge control, the disc centrifuge liquid-liquid separation intelligent control system comprises a data acquisition module, an artificial intelligence module and an intelligent control module, the data acquisition module, the artificial intelligence module and the intelligent control module are in signal connection in sequence, and closed-loop control is formed. According to the system, the optimal feeding amount is automatically output through the pre-training model, dynamic prediction is carried out, dependence on professional experience of operators is greatly reduced, manual judgment errors are avoided, the effect that the separation effect of each batch of materials meets the preset requirement can be ensured by accurately predicting and restraining the discharging state through the artificial intelligence module and combining the stability of closed-loop control, and the separation efficiency is improved. Reworking and waste materials caused by substandard separation are reduced, the consistency of product quality is remarkably improved, the dual purposes that the separation effect reaches the standard and the treatment efficiency is the highest are achieved, the material handling capacity of equipment in unit time is improved, and energy consumption is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of disc centrifuge control, more particularly, the present application relates to a disc centrifuge liquid-liquid separation intelligent control system based on artificial intelligence. BACKGROUND

[0002] As a high-efficiency separation equipment, disc centrifuge is widely used in food, chemical, pharmaceutical and other industries, and its separation performance directly affects the production efficiency and product quality. During the operation of the disc centrifuge, the feed quantity is a key operating parameter, which has a complex nonlinear relationship with the separation effect and separation efficiency. On the one hand, under the condition of constant speed, the smaller the feed quantity, the more sufficient the separation process in the drum, and the better the separation effect. On the other hand, too low feed quantity will reduce the separation efficiency and increase the unit processing cost. Therefore, finding the optimal feed quantity to balance the separation effect and separation efficiency is of great significance to improve the overall performance of the disc centrifuge.

[0003] The control mode of the existing disc centrifuge mainly relies on the experience judgment of engineers and manual adjustment according to the equipment operation manual, and the degree of automation is very low. On the one hand, manual control cannot accurately capture the complex nonlinear relationship between feed quantity and separation effect and efficiency, which easily leads to fluctuation of separation effect, such as excessive turbidity of discharge or low discharge efficiency. On the other hand, different models such as drum structure, disc parameters or different materials such as viscosity, density and temperature difference have poor adaptability, and it is difficult to realize standardized and stable separation control, which seriously restricts the overall performance improvement of the disc centrifuge. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the present application provides a disc centrifuge liquid-liquid separation intelligent control system based on artificial intelligence to solve the problems raised in the background art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a disc centrifuge liquid-liquid separation intelligent control system based on artificial intelligence, comprising a data acquisition module, an artificial intelligence module and an intelligent control module, the data acquisition module, the artificial intelligence module and the intelligent control module are sequentially signal connected to form a closed loop control.

[0006] The data acquisition module is used for collecting real-time data of the centrifuge operation and constructing a centrifugal separation data set, the real-time data including operating parameters, centrifuge structure parameters, feed state parameters and discharge state parameters.

[0007] The artificial intelligence module includes a discharge state predictor and a feed quantity generator, supporting a training mode and a running mode: in the training mode, the discharge state predictor is trained with the centrifugal separation data set, and the pre-trained discharge state predictor is embedded in the feed quantity generator and trained; in the running mode, the optimal feed quantity q* is output according to the real-time data of the data acquisition module;

[0008] The intelligent control module is a negative feedback closed loop control module, which adjusts the feed pipe and discharge pipe valves of the centrifuge according to the optimal feed quantity q*, so that the actual feed quantity is consistent with q*.

[0009] Preferably, the data acquisition module includes multiple types of sensors, including:

[0010] A flow meter installed on the feed pipe for collecting the feed quantity q, with the unit of hL / hour;

[0011] A turbidimeter installed on the feed pipe, the light phase discharge pipe and the heavy phase discharge pipe for collecting the feed turbidity, the light phase discharge turbidity and the heavy phase discharge turbidity, with the unit of EBC;

[0012] A viscosity sensor installed on the feed pipe, the light phase discharge pipe and the heavy phase discharge pipe for collecting the feed viscosity, the light phase discharge viscosity and the heavy phase discharge viscosity, with the unit of Pa·s;

[0013] A temperature sensor installed on the feed pipe, the light phase discharge pipe and the heavy phase discharge pipe for collecting the feed temperature, the light phase discharge temperature and the heavy phase discharge temperature, with the unit of ℃;

[0014] A densimeter installed on the light phase discharge pipe and the heavy phase discharge pipe for collecting the light phase discharge density and the heavy phase discharge density, with the unit of g / cm 3 ;

[0015] A rotational speed sensor installed on the main shaft of the centrifuge for collecting the rotational speed n of the drum, with the unit of rpm.

[0016] Preferably, the operating parameters include the rotational speed n and the feed quantity q;

[0017] The centrifuge structure parameters S include the inner diameter of the drum, the total volume of the drum, the outer diameter of the disc, the inner diameter of the disc, the disc cone angle, the disc gap, the number of discs, the distance from the center hole to the center shaft, and the diameter of the center hole;

[0018] The feed state parameters M include the feed turbidity, the feed temperature, the feed viscosity, and the ratio of light and heavy phases;

[0019] The discharge state parameters V include light phase discharge turbidity, heavy phase discharge turbidity, light phase discharge viscosity, heavy phase discharge viscosity, light phase discharge temperature, heavy phase discharge temperature, light phase discharge density, and heavy phase discharge density.

[0020] Preferably, the discharge state predictor comprises an input feature encoding layer, a deep neural network, and a discharge parameter output layer.

[0021] Preferably, the input feature encoding layer processes the input parameter vector u = [n, q, S, M] to obtain [n, q, S, M, a, f], normalizes the extended u, and outputs the feature vector u' through a self-attention mechanism.

[0022] The deep neural network is a 4-layer fully connected neural network, which is used to learn the nonlinear coupling relationship of the input high-dimensional features and output the feature vector v'.

[0023] The discharge parameter output layer maps v' to the discharge state parameter vector v through a fully connected layer and uses an activation function to constrain the numerical range of v.

[0024] Preferably, the loss function of the discharge state predictor is mean square error (MSE), which is the square mean of the difference between the predicted discharge parameters and the measured discharge parameters.

[0025] Preferably, the feed amount generator comprises an input feature encoding layer, a feed amount generation network, and a pre-trained discharge state predictor.

[0026] Preferably, the input feature encoding layer processes the input parameter vector w = [n, S, M, V] to obtain [n, S, M, a, f], wherein the effective separation area a is calculated based on the centrifuge structure parameter S in w, and the separation factor f is calculated based on the centrifuge real-time running speed n and the structure parameter S in w.

[0027] Calculate the parameter difference Δ of the feed and the light phase and heavy phase discharge.

[0028] Δ = [ΔNTU l , ΔNTU h , ΔT l , ΔT h , Δη l , Δη hThe effective separation area a, the separation factor f, and the parameter difference Δ calculated above are supplemented to the initial input parameter vector w to form an extended input vector, and an extended input feature vector w = [n, S, M, V, Δ] is obtained. NTU is the turbidity value, T is the temperature value, and η is the viscosity value. The subscript l represents the light phase, and the subscript h represents the heavy phase. The separation requirements are quantified in this way, and a structured feature sequence is output after processing by the two-head self-attention mechanism.

[0029] The feed amount generation network is a 3-layer fully connected neural network, and the Softplus activation function is used in the output layer to ensure that the output feed amount q* ≥ 0.

[0030] The pre-trained discharge state predictor is embedded to verify the discharge effect corresponding to q*, and its parameters are frozen during the training process.

[0031] Preferably, the loss function of the feed amount generator is a double-objective loss function L = L constraint + γ·L incentive , wherein:

[0032] L constraint is a constraint term that forces the discharge parameters to meet the standards. Differentiated penalties are designed for different types of discharge requirements. A maximum penalty is given when the standards are not met, and the penalty is 0 when the standards are met, ensuring that the model prioritizes meeting the customer's separation requirements. L constraint = L NTU + L T + L η + L ρ .

[0033] L NTU is the turbidity constraint term: the turbidity cannot exceed the upper limit value set by the user.

[0034] NTU l,max is the upper limit of the turbidity of the light phase, and NTU h,max is the upper limit of the turbidity of the heavy phase.

[0035] L T is the temperature constraint term: the temperature must be within the upper and lower limit range set by the user.

[0036] T l,max is the upper limit of the temperature of the light phase, T l,min is the lower limit of the temperature of the light phase, T h,max is the upper limit of the temperature of the heavy phase, and T h,min is the lower limit of the temperature of the heavy phase.

[0037] L η is the viscosity constraint term: the viscosity must be within the upper and lower limit range set by the user.

[0038] η l,max η is the upper limit of the light phase viscosity l,min η is the lower limit of the light phase viscosity l,max η is the upper limit of the heavy phase viscosity h,min η is the lower limit of the heavy phase viscosity

[0039] L ρ ρ is the density constraint term: the density should be within the upper and lower limits set by the user;

[0040] ρ l,max ρ is the upper limit of the light phase density l,min ρ is the lower limit of the light phase density l,max ρ is the upper limit of the heavy phase density h,min ρ is the lower limit of the heavy phase density

[0041] L incentive is the incentive term, γ is the incentive coefficient, used to control the strength of the incentive for maximizing the feed amount.

[0042] Preferably, the execution element of the intelligent control module is a diaphragm valve of the feed pipe, light phase discharge pipe, and heavy phase discharge pipe, and the feed amount is adjusted by adjusting the valve pressure of the diaphragm valve.

[0043] Technical effects and advantages of the present application:

[0044] 1. By inputting the centrifuge structure parameters, operation parameters, material parameters, and discharge requirements by the user, the artificial intelligence module can automatically output the optimal feed amount through the pre-trained model, without the need for manual trial and error adjustment. The system supports real-time adjustment of separation strategies, combines real-time data feedback from the data acquisition module with dynamic prediction from the artificial intelligence model, realizes the transition from passive experience control to active intelligent decision-making, greatly reduces the dependence on the professional experience of the operator, avoids human judgment errors, and is especially suitable for large-scale, continuous production scenarios. The intelligent control module uses a diaphragm valve as the execution element, compares the actual feed amount with the optimal feed amount output by the artificial intelligence module in real time through a negative feedback closed loop, and adjusts the valve pressure of the feed pipe and discharge pipe immediately if there is a deviation, ensuring that the feed amount is stable at the optimal value and avoiding fluctuations in separation effects caused by fluctuations in material properties and minor equipment faults. Through the precise prediction and constraint of the discharge state by the artificial intelligence module, combined with the stability of the closed-loop control, the separation effect of each batch of material can be ensured to meet the preset requirements, reducing rework and waste caused by substandard separation, and significantly improving the consistency of product quality.

[0045] 2. Optimal balance is guaranteed by a double-target loss function. The loss function of the feed rate generator contains a constraint term and an incentive term. The constraint term gives a large penalty when the discharge parameters exceed the limit, forcing the turbidity, temperature, viscosity, and density of the light and heavy phase discharges to meet the user's requirements, ensuring that the separation effect meets the standard. The incentive term guides the model to output the maximum feed rate under the premise of meeting the discharge standard, solving the contradiction between low feed rate and good effect, and high feed rate and high efficiency. The discharge state predictor uses a 4-layer fully connected neural network to accurately learn the nonlinear coupling relationship between speed, feed rate, structure parameters, and material properties, and the prediction error is strictly controlled by mean square error (MSE) to provide a reliable basis for effect prediction for feed rate optimization. Finally, the dual goals of meeting the separation effect standard and achieving the highest processing efficiency are achieved.

[0046] 3. Generalization training is supported by full-dimensional data collection. The data collection module covers multiple types of sensors, and the collected parameters cover multiple operating parameters, structure parameters, feed state, and discharge state. A centrifugal separation data set for multiple models and materials can be constructed to provide comprehensive training samples for the artificial intelligence model. The physical prior and feature fusion enhance adaptability. The input feature encoding layer of the artificial intelligence module introduces the effective separation area a and the separation factor f. At the same time, the discharge predictor with self-attention mechanism and the feed generator with 2 heads of self-attention mechanism capture cross-dimensional parameter correlations, making the model adaptable to centrifuges with different structure parameters and materials with different characteristics. There is no need to develop repeatedly for a single working condition. Under the premise of meeting the separation effect standard, the incentive term guides the model to output the maximum feed rate, improving the material processing capacity of the equipment per unit time, avoiding high energy consumption and low output caused by long-term low load operation of the equipment, and precisely controlling to reduce energy waste caused by improper feed rate. BRIEF DESCRIPTION OF DRAWINGS

[0047] Fig. 1 The flowchart of the reasoning phase of the system of the present application.

[0048] Fig. 2 The flowchart of the training phase of the system of the present application. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0050] As shown in the accompanying Figs. 1-2The disc centrifuge liquid-liquid separation intelligent control system based on artificial intelligence shown comprises a data acquisition module, an artificial intelligence module and an intelligent control module, characterized in that the data acquisition module, the artificial intelligence module and the intelligent control module are sequentially signal connected to form a closed loop control.

[0051] The data acquisition module is used for acquiring real-time data of the centrifuge operation and constructing a centrifugal separation data set, and the real-time data includes operation parameters, centrifuge structure parameters, feed state parameters and discharge state parameters.

[0052] The artificial intelligence module includes a discharge state predictor and a feed amount generator, and supports a training mode and a running mode: in the training mode, the discharge state predictor is trained with the centrifugal separation data set, and then the pre-trained discharge state predictor is embedded into the feed amount generator and trained; in the running mode, the optimal feed amount q* is output according to the real-time data of the data acquisition module.

[0053] The intelligent control module is a negative feedback closed loop control module, which adjusts the feed pipe and discharge pipe valves of the centrifuge according to the optimal feed amount q* so that the actual feed amount is consistent with q*.

[0054] The data acquisition module includes multiple types of sensors, including:

[0055] A flow meter installed on the feed pipe for acquiring the feed amount q, with the unit of hL / hour;

[0056] A turbidimeter installed on the feed pipe, the light phase discharge pipe and the heavy phase discharge pipe for acquiring the feed turbidity, the light phase discharge turbidity and the heavy phase discharge turbidity, with the unit of EBC;

[0057] A viscosity sensor installed on the feed pipe, the light phase discharge pipe and the heavy phase discharge pipe for acquiring the feed viscosity, the light phase discharge viscosity and the heavy phase discharge viscosity, with the unit of Pa·s;

[0058] A temperature sensor installed on the feed pipe, the light phase discharge pipe and the heavy phase discharge pipe for acquiring the feed temperature, the light phase discharge temperature and the heavy phase discharge temperature, with the unit of ℃;

[0059] A densimeter installed on the light phase discharge pipe and the heavy phase discharge pipe for acquiring the light phase discharge density and the heavy phase discharge density, with the unit of g / cm 3 ;

[0060] A rotational speed sensor installed on the centrifuge main shaft for acquiring the drum rotational speed n, with the unit of rpm.

[0061] Solve the problem of low precision and poor real-time of manual collection, the sensor can collect in real time to ensure the timeliness and accuracy of the data, and the measurement unit of the sensor is clear to avoid control errors caused by parameter unit confusion, which can improve the reliability of system control.

[0062] The operating parameters include the rotation speed n and the feed amount q.

[0063] The centrifuge structure parameters S include the inner diameter of the rotating drum, the total volume of the rotating drum, the outer diameter of the disc, the inner diameter of the disc, the cone angle of the disc, the disc gap, the number of discs, the distance from the center hole to the center shaft, and the diameter of the center hole.

[0064] The feed state parameters M include the turbidity of the feed, the temperature of the feed, the viscosity of the feed, and the ratio of light and heavy phases.

[0065] The discharge state parameters V include the turbidity of the light phase discharge, the turbidity of the heavy phase discharge, the viscosity of the light phase discharge, the viscosity of the heavy phase discharge, the temperature of the light phase discharge, the temperature of the heavy phase discharge, the density of the light phase discharge, and the density of the heavy phase discharge.

[0066] The eight indicators of the discharge state parameters V are clear, which can fully cover the evaluation dimensions of the separation effect, ensure that the optimal feed amount can meet the requirements of multiple dimensions of the discharge, and avoid product quality defects caused by single index control.

[0067] The discharge state predictor includes an input feature encoding layer, a deep neural network, and a discharge parameter output layer.

[0068] The processing process of the input feature encoding layer is as follows: receiving the input parameter vector u = [n, q, S, M], calculating the effective separation area a and the separation factor f of the disc centrifuge, expanding u to obtain [n, q, S, M, a, f], normalizing the expanded u, and outputting the feature vector u' after processing by the self-attention mechanism.

[0069] The deep neural network is a 4-layer fully connected neural network, which is used to learn the nonlinear coupling relationship of the input high-dimensional features, and output the feature vector v'.

[0070] The discharge parameter output layer maps v' to the discharge state parameter vector v through a fully connected layer, and uses an activation function to constrain the numerical range of v.

[0071] By introducing physical parameters such as a and f, the adaptability of the model to different centrifuge models, such as centrifuges with different inner diameters of the rotating drum, is improved, and overfitting of the model is avoided. The self-attention mechanism can dynamically focus on features that significantly affect the separation effect, and the 4-layer fully connected network can deeply mine nonlinear relationships to improve the accuracy of discharge state prediction and provide a reliable basis for effect prediction for feed amount optimization.

[0072] The loss function of the discharge state predictor is mean square error (MSE), that is, the square mean of the difference between the predicted discharge parameter and the measured discharge parameter;

[0073] The training optimization target (MSE) of the discharge state predictor is defined to clearly judge the model training, so that the predicted discharge parameter is as close as possible to the measured value. A quantitative optimization direction is provided for model training to ensure that the training process is monitorable and convergent.

[0074] The feed amount generator includes an input feature encoding layer, a feed amount generation network, and a pre-trained discharge state predictor;

[0075] The processing process of the input feature encoding layer is: receiving an input parameter vector w = [n, S, M, V], calculating the effective separation area a and the separation factor f, wherein the effective separation area a is calculated according to the centrifuge structure parameter S in w, and the separation factor f is calculated according to the centrifuge real-time running speed n and the structure parameter S in w;

[0076] The parameter difference Δ of the feed and the light phase and the heavy phase discharge is calculated;

[0077] Δ = [ΔNTU l ,ΔNTU h ,ΔT l ,ΔT h ,Δη l ,Δη h ], the above calculated effective separation area a, separation factor f and parameter difference Δ are supplemented to the initial input parameter vector w to form an expanded input vector, and the expanded input feature vector w = [n, S, M, V, Δ] is output. The structured feature sequence is processed by the two-head self-attention mechanism;

[0078] The feed amount generation network is a 3-layer fully connected neural network, and the output layer adopts a Softplus activation function to ensure that the output feed amount q* ≥ 0;

[0079] The pre-trained discharge state predictor is embedded to verify the discharge effect corresponding to q*, and its parameters are frozen during the training process;

[0080] The problem of being unable to quantify the separation requirement and the problem of having no basis for feed amount optimization are solved. Δ can intuitively reflect the gap between the current feed and discharge requirements, so that the model optimization is more targeted.

[0081] The loss function of the feed amount generator is a double-objective loss function L = L constraint + γ·L incentive , wherein:

[0082] L constraintFor the constraint term, the discharge parameter is forced to meet the standard, and different types of discharge requirements are designed with different penalties. When it does not meet the standard, it gives a great penalty, that is, when each discharge parameter is over limit, it gives a great penalty, and when it meets the standard, the penalty is 0, which ensures that the model meets the customer's separation requirements first, L constraint = L NTU + L T + L η + L ρ ;

[0083] L NTU The turbidity constraint term: the turbidity cannot exceed the upper limit value set by the user;

[0084] The upper limit of the light phase turbidity is NTU h,max The upper limit of the heavy phase turbidity is NTU

[0085] L T The temperature constraint term: the temperature should be within the upper and lower limit range set by the user;

[0086] T l,ma The upper limit of the light phase temperature is T l,min The lower limit of the light phase temperature is T h,max The upper limit of the heavy phase temperature is T h,min The lower limit of the heavy phase temperature is T

[0087] L η The viscosity constraint term: the viscosity should be within the upper and lower limit range set by the user;

[0088] η l,max The upper limit of the light phase viscosity is η l,min The lower limit of the light phase viscosity is η l,max The upper limit of the heavy phase viscosity is η h,min The lower limit of the heavy phase viscosity is η

[0089] L ρ The density constraint term: the density should be within the upper and lower limit range set by the user;

[0090] ρ l,max The upper limit of the light phase density is ρ l,min The lower limit of the light phase density is ρ l,max The upper limit of the heavy phase density is ρ h,min The lower limit of the heavy phase density is ρ

[0091] L incentive The incentive term, γ is the incentive coefficient, used to control the incentive strength of the maximum feed amount;

[0092] Define the dual-objective optimization logic L of the feed amount generator constraint Ensure that the discharge meets the standard, L incentive Maximize the feed amount, specify the specific calculation method and penalty rules of the two types of constraint terms, and design different penalties for different discharge parameters such as turbidity, temperature, viscosity, and density. The turbidity is only constrained by the upper limit, the temperature is constrained by the upper and lower limits, and the model is adapted to the diversified discharge requirements in actual production, L constraint Prioritize ensuring that the discharge meets the standard to avoid product quality defects, L incentive Maximize the feed amount after meeting the standard to improve processing efficiency and reduce unit cost. Differentiated penalty rules can greatly punish non-compliance to force the model to prioritize correcting the over-standard parameters, avoiding sacrificing quality for efficiency. The incentive coefficient γ can be flexibly adjusted to improve the system's adaptability to different industry scenarios.

[0093] The execution element of the intelligent control module is the diaphragm valve of the feed pipe, the light phase discharge pipe, and the heavy phase discharge pipe. The feed amount is adjusted by adjusting the valve pressure of the diaphragm valve. The diaphragm valve can achieve precise control of the feed amount through pressure fine-tuning, ensuring that the actual feed amount quickly approaches q*, reducing separation fluctuations during the transition process, and avoiding pressure imbalance at the discharge end caused by single adjustment of the feed valve. Only adjusting the feed valve may cause heavy phase discharge to be difficult, ensuring the stability of the centrifuge operation and prolonging the service life of the equipment.

[0094] The working principle of the present application: the disc type centrifuge liquid-liquid separation intelligent control system based on artificial intelligence forms a closed loop control through the sequential signal connection of the data acquisition module, the artificial intelligence module, and the intelligent control module. First, the data acquisition module collects centrifuge operation data through multiple types of sensors and builds a centrifugal separation data set. The flow meter of the feed pipe collects the feed amount q, which is hL / hour. The turbidity meter of the feed pipe, the light phase, and the heavy phase discharge pipe collects the feed and light / heavy phase discharge turbidity, which is EBC. The viscosity sensor collects the feed and light / heavy phase discharge viscosity, which is Pa·s. The temperature sensor collects the feed and light / heavy phase discharge temperature, which is ℃. The density meter of the light phase and the heavy phase discharge pipe collects the light phase and the heavy phase discharge density, which is g / cm 3 The speed sensor on the main shaft collects the drum speed n, which is rpm. At the same time, the preset operating parameters speed n, feed amount q, centrifuge structure parameters S (drum inner diameter, drum total volume, disc outer diameter, disc inner diameter, disc cone angle, disc gap, disc number, center hole to center shaft distance, center hole diameter), feed state parameters M (feed turbidity, feed temperature, feed viscosity, light / heavy phase ratio), and discharge state parameters V (light / heavy phase discharge turbidity, viscosity, temperature, density) are integrated.

[0095] Then, the artificial intelligence module first enters the training mode to train the discharge state predictor, receives the input parameter vector u = [n, q, S, M], calculates the effective separation area a according to the structure parameter S and the separation factor f according to the rotation speed n and the structure parameter S, expands u to [n, q, S, M, a, f], processes the output feature vector u' through normalization and self-attention mechanism, then learns the nonlinear coupling relationship through a 4-layer fully connected neural network to output the feature vector v', and finally maps the discharge parameter output layer to the discharge state parameter vector v to predict the mean square error MSE of the discharge parameters and the measured discharge parameters.

[0096] Subsequently, the pre-trained discharge state predictor is embedded into the feed amount generator for training: receiving the input parameter vector w = [n, S, M, V], calculating the parameter difference Δ of the feed and the light phase and heavy phase discharge, Δ = [ΔNTU l , ΔNTU h , ΔT l , ΔT h , Δη l , Δη h ], supplementing the effective separation area a, the separation factor f and the parameter difference Δ calculated above to the initial input parameter vector w to form an expanded input vector, expanding the input feature vector w = [n, S, M, V, Δ], NTU is the turbidity value, T is the temperature value, η is the viscosity value, and the subscript l represents the light phase and the subscript h represents the heavy phase, thereby quantifying the separation requirement, outputting a structured feature sequence after processing by 2 heads of self-attention mechanism, inputting a 3-layer fully connected neural network after processing by 2 heads of self-attention mechanism, and using a Softplus activation function in the output layer to ensure that the optimal feed amount q* ≥ 0, while the embedded pre-trained discharge state predictor verifies the discharge effect corresponding to q*, the parameters are frozen during training, and a double-objective loss function L = L constraint + γ · L incentive is used to optimize the model, after training, the artificial intelligence module enters the running mode, and outputs the optimal feed amount q* according to the real-time data of the data acquisition module;

[0097] Finally, the intelligent control module serves as a negative feedback closed loop control module, taking the diaphragm valves of the feed pipe, the light phase discharge pipe and the heavy phase discharge pipe as the execution elements, adjusting the valve pressure of the diaphragm valve to make the actual feed amount consistent with the optimal feed amount q*, and at the same time, the data acquisition module continuously collects real-time running data and feeds back to the artificial intelligence module, forming a closed loop of collection, decision, control and feedback, and realizing intelligent control of liquid-liquid separation.

[0098] Finally should be explained a few points are: first, in the description of the present application, it should be pointed out that, unless otherwise specified and limited, the term "installation", "connected", "connection" should be broad, can be mechanical or electrical connection, but also can be two elements inside the communication, can be directly connected, "up", "down", "left", "right" and so on, only for indicating the relative position relationship, when the absolute position of the described object changes, the relative position relationship may change;

[0099] Second: the present application discloses the drawings in the embodiment, only involves the structure related to the present application, other structures can refer to the usual design, in the case of no conflict, the same embodiment and different embodiments of the present application can be combined with each other;

[0100] Finally: the above only for the preferred embodiment of the present application, and not for limiting the present application, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application, should be included in the protection scope of the present application.

Claims

1. An intelligent control system for liquid-liquid separation in a disc centrifuge based on artificial intelligence, comprising a data acquisition module, an artificial intelligence module, and an intelligent control module, characterized in that: The data acquisition module, artificial intelligence module, and intelligent control module are sequentially connected to form a closed-loop control. The data acquisition module is used to collect real-time data of the centrifuge operation and construct a centrifugation separation dataset. The real-time data includes operating parameters, centrifuge structural parameters, feed status parameters, and discharge status parameters. The artificial intelligence module includes a discharge status predictor and a feed rate generator, supporting training mode and running mode: In training mode, the discharge status predictor is trained with the centrifugal separation dataset, and then the pre-trained discharge status predictor is embedded into the feed rate generator and trained; In running mode, the optimal feed rate q* is output based on the real-time data from the data acquisition module. The intelligent control module is a negative feedback closed-loop control module, which adjusts the valves of the centrifuge's feed pipe and discharge pipe according to the optimal feed rate q*, so that the actual feed rate is consistent with q*.

2. The intelligent control system for liquid-liquid separation of a disc centrifuge based on artificial intelligence according to claim 1, characterized in that: The data acquisition module includes multiple types of sensors, including: The flow meter installed on the feed pipe is used to collect the feed rate q, with the unit being hL / hour; Turbidity meters installed in the feed pipe, light phase discharge pipe, and heavy phase discharge pipe are used to collect the turbidity of the feed, light phase discharge, and heavy phase discharge, with the unit being EBC. Viscosity sensors installed in the feed pipe, light phase discharge pipe, and heavy phase discharge pipe are used to collect the feed viscosity, light phase discharge viscosity, and heavy phase discharge viscosity, in Pa·s. Temperature sensors installed in the feed pipe, light phase discharge pipe, and heavy phase discharge pipe are used to collect the feed temperature, light phase discharge temperature, and heavy phase discharge temperature, in °C. Densitometers installed in the light phase discharge pipes and heavy phase discharge pipes are used to collect the discharge density of the light phase and heavy phase, with units of g / cm³. 3 ; The speed sensor installed on the centrifuge spindle is used to collect the drum speed n, which is measured in rpm.

3. The intelligent control system for liquid-liquid separation of a disc centrifuge based on artificial intelligence according to claim 2, characterized in that: The operating parameters include rotational speed n and feed rate q; The centrifuge structural parameters S include the inner diameter of the drum, the total volume of the drum, the outer diameter of the discs, the inner diameter of the discs, the disc cone angle, the disc gap, the number of discs, the distance from the center hole to the center shaft, and the diameter of the center hole; The feed state parameters M include feed turbidity, feed temperature, feed viscosity, and the ratio of light to heavy phases; The discharge state parameter V includes light phase discharge turbidity, heavy phase discharge turbidity, light phase discharge viscosity, heavy phase discharge viscosity, light phase discharge temperature, heavy phase discharge temperature, light phase discharge density, and heavy phase discharge density.

4. The intelligent control system for liquid-liquid separation of a disc centrifuge based on artificial intelligence according to claim 3, characterized in that: The discharge status predictor includes an input feature encoding layer, a deep neural network, and a discharge parameter output layer.

5. The intelligent control system for liquid-liquid separation of a disc centrifuge based on artificial intelligence according to claim 4, characterized in that: The processing procedure of the input feature coding layer is as follows: receive the input parameter vector u = [n, q, S, M], calculate the effective separation area a and separation factor f of the disc centrifuge to expand u to obtain [n, q, S, M, a, f], normalize the expanded u, and output the feature vector u' after processing through the self-attention mechanism; The deep neural network is a 4-layer fully connected neural network used to learn the nonlinear coupling relationship of the input high-dimensional features and output a feature vector v'. The output parameter layer maps v' to the output state parameter vector v through a fully connected layer, and uses an activation function to constrain the numerical range of v.

6. The intelligent control system for liquid-liquid separation of a disc centrifuge based on artificial intelligence according to claim 4, characterized in that: The loss function of the discharge state predictor is the mean square error (MSE), which is the squared mean of the difference between the predicted discharge parameters and the measured discharge parameters.

7. The intelligent control system for liquid-liquid separation of a disc centrifuge based on artificial intelligence according to claim 6, characterized in that: The feed rate generator includes an input feature encoding layer, a feed rate generation network, and a pre-trained discharge state predictor.

8. The intelligent control system for liquid-liquid separation of a disc centrifuge based on artificial intelligence according to claim 7, characterized in that: The processing procedure of the input feature coding layer is as follows: receive the input parameter vector w = [n, S, M, V], calculate the effective separation area a and the separation factor f, wherein the calculation of the effective separation area a is based on the centrifuge structural parameter S in w, and the calculation of the separation factor f is based on the centrifuge real-time operating speed n and structural parameter S in w; Calculate the parameter difference Δ between the feed and the light and heavy phases of the discharge; Δ=[ΔNTU l ,ΔNTU h ,ΔT l ,ΔT h ,Δη l ,Δη h The effective separation area a, separation factor f, and parameter difference Δ obtained above are added to the initial input parameter vector w to form an expanded input vector. The expanded input feature vector w = [n, S, M, V, Δ], where NTU is the turbidity value, T is the temperature value, η is the viscosity value, subscript l represents the light phase, and subscript h represents the heavy phase. The separation requirements are quantified in this way, and the structured feature sequence is output after processing by a two-head self-attention mechanism. The feed rate generation network is a 3-layer fully connected neural network, and the output layer uses the Softplus activation function to ensure that the output feed rate q*≥0; The pre-trained discharge state predictor is embedded to verify the discharge effect corresponding to q*, and its parameters are frozen during the training process.

9. The intelligent control system for liquid-liquid separation of a disc centrifuge based on artificial intelligence according to claim 8, characterized in that: The loss function of the feed rate generator is a biobjective loss function L = L constraint +γ·L incentive ,in: L constraint As a constraint, mandatory discharge parameters are enforced. Differentiated penalties are designed for different types of discharge requirements, with a large penalty for non-compliance and a zero penalty for compliance, ensuring that the model prioritizes meeting the customer's separation requirements. constraint =L NTU +L T +L η +L ρ ; L NTU Turbidity constraint: Turbidity cannot exceed the upper limit set by the user; NTU l,max The upper limit of light phase turbidity, NTU h,max This represents the upper limit of heavy phase turbidity. L T Temperature constraint: The temperature must be within the upper and lower limits set by the user; T l,max T is the upper limit of the light phase temperature. l,min T is the lower limit of the light phase temperature. h,max T is the upper limit of the rephase temperature. h,min This is the lower limit of the repetitive phase temperature; L η For viscosity constraints: the viscosity must be within the upper and lower limits set by the user; η l,max η represents the upper limit of viscosity of the light phase. l,min η is the lower limit of viscosity of the light phase. l,max η is the upper limit of the viscosity of the heavy phase. h,min This is the lower limit of the viscosity of the heavy phase; L ρ Density constraint: The density must be within the upper and lower limits set by the user; ρ l,max ρ is the upper limit of the light phase density. l,min As the lower limit of the light phase density, ρ l,max ρ is the upper limit of the heavy phase density. h,min This is the lower limit of the heavy phase density; L incentive As an incentive, γ is the excitation coefficient, used to control the excitation intensity to maximize the feed rate.

10. The intelligent control system for liquid-liquid separation of a disc centrifuge based on artificial intelligence according to claim 1, characterized in that: The actuators of the intelligent control module are diaphragm valves of the feed pipe, light phase discharge pipe, and heavy phase discharge pipe. The feed rate is adjusted by regulating the valve pressure of the diaphragm valves.

Images