Fluorine pump multi-connected refrigerant oil return and separation industrial control system based on industrial internet

Abstract

The industrial internet-based fluorine pump multi-connected refrigerant oil return and separation industrial control system belongs to the technical field of industrial automation control, and comprises: an edge computing collection unit, which is used for collecting and preprocessing data of a characteristic frequency band of freon pressure pulsation to generate preprocessed data; a multi-scale nested prediction control unit, which is used for receiving the preprocessed data and generating a control strategy instruction set; and a system execution unit, which is used for receiving and executing the control strategy instruction set. The application configures the control system with a sensor that can perceive the microscopic world, so that it can make decisions based on more real and complete physical information, thereby improving the accuracy of control.

Description

Technical Field

[0001] This invention relates to the field of industrial automation control, specifically to an industrial control system for multi-unit refrigerant return and separation of fluorine pumps based on the Industrial Internet. Background Technology

[0002] One of the core challenges in operating a multi-phase refrigerant pump refrigeration system lies in the precise control of the mixing and separation of refrigerant and lubricating oil. Traditional control methods simplify this mixture into an ideal two-phase flow, which is feasible when the system load is stable, but exhibits significant shortcomings under dynamic conditions with rapid load changes. Existing technologies fail to fully consider the drastic nonlinear abrupt changes in the miscibility of new environmentally friendly refrigerants, such as R410A / R32, with POE lubricating oil within specific temperature and pressure ranges. These abrupt changes can lead to a sharp drop in solubility in a very short time, resulting in oil miscibility instability. Simultaneously, traditional sensors cannot effectively detect nanoscale microfoam generated by the high-speed operation of the compressor, leading to signal acquisition distortion. The response speed of existing control logic is also insufficient, with a detection lag typically in the range of 200-300ms, making it impossible to provide early warning and effective response to rapid phase changes.

[0003] To address the aforementioned technical problems, this invention provides an industrial control system that deeply integrates the real-time data acquisition capabilities of the Industrial Internet with the unique phase change physical characteristics of Freon, thereby constructing an industrial control system for multi-unit refrigerant return and separation of Freon pumps based on the Industrial Internet. Summary of the Invention

[0004] The purpose of this invention is to provide an industrial control system for multi-unit refrigerant return and separation of fluorine pumps based on the Industrial Internet, so as to solve the problems mentioned in the background art.

[0005] The technical solution of the present invention includes:

[0006] The edge computing acquisition unit is used to acquire and preprocess data for the characteristic frequency bands of Freon pressure pulsation in order to generate preprocessed data.

[0007] A multi-scale nested predictive control unit is used to receive preprocessed data and generate a set of control strategy instructions;

[0008] The system execution unit is used to receive and execute the control strategy instruction set.

[0009] Preferably, the multi-scale nested prediction control unit generates a control strategy instruction set through a three-layer nested prediction architecture; the three-layer nested prediction architecture includes:

[0010] The molecular-level prediction module is used to process real-time high-frequency temperature and pressure data in the preprocessed data to output a three-dimensional local solubility distribution field.

[0011] The mesoscale prediction module is used to receive local solubility distribution field and macroscopic flow field data from preprocessed data to generate the probability density function of oil two-phase distribution.

[0012] The system-level prediction module is used to receive the probability density function of the two-phase distribution of oil and combine it with the preset system topology data to output a set of control strategy instructions.

[0013] Preferably, the molecular-level prediction module outputs a three-dimensional local solubility distribution field in the following manner:

[0014] Receive real-time high-frequency temperature data, pressure data, and concentration gradient derived from the previous calculation time.

[0015] The process is based on a pre-defined surrogate model of the Freon-oil interaction potential energy surface to output a three-dimensional local solubility distribution field in real time.

[0016] Preferably, the mesoscale prediction module generates the probability density function of the oil-liquid two-phase distribution in the following manner:

[0017] Determine the gradient of the chemical potential, and based on the gradient of the chemical potential, define the phase separation force term;

[0018] The phase separation force term is introduced into the fluid dynamics simulation of the lattice Boltzmann method to correct the fluid momentum;

[0019] The evolution equation of the method is solved and the particle distribution function is statistically averaged to generate the probability density function of the two-phase distribution of oil.

[0020] Preferably, the system-level prediction module outputs the control strategy instruction set in the following manner:

[0021] For each node in the system topology diagram data, calculate the statistical moments of the probability density function of the oil-liquid two-phase distribution within the physical volume corresponding to that node, and construct the initial feature vector of the node based on the statistical moments.

[0022] A graph neural network model is adopted, and the node state is updated through multi-layer information propagation. The final output layer of the graph neural network model is output as the control policy instruction set.

[0023] Preferably, the statistical moments include:

[0024] The average oil concentration within the physical volume corresponding to each node;

[0025] The concentration variance within the physical volume corresponding to the node;

[0026] The skewness of the concentration distribution within the physical volume corresponding to the node.

[0027] Preferably, the multi-scale nested prediction and control unit uses a semi-empirical prediction model to output dimensionless system key state prediction indicators.

[0028] By multiplying the fundamental thermodynamic function, the mixing enthalpy correction function, the foam Reynolds number correction term, and the phase transition time constant function, the system's critical state prediction index is obtained.

[0029] Preferably, the process for obtaining the foam Reynolds number corresponding to the foam Reynolds number correction term is as follows:

[0030] The equivalent diameter of the foam was determined by high-frequency pressure pulsation spectrum analysis;

[0031] Obtain the real-time density of the mixture, the real-time viscosity of the mixture, and the average flow velocity in the pipeline;

[0032] The intermediate product value is obtained by multiplying the real-time density of the mixture, the average flow velocity in the pipe, and the equivalent diameter of the foam.

[0033] Divide the intermediate product by the real-time viscosity of the mixture to obtain the foam Reynolds number.

[0034] Preferably, the process for obtaining the phase transition time constant function is as follows:

[0035] Monitor whether the system status undergoes a step change;

[0036] If the system state undergoes a step change, a first-order system response model is used to fit the trend of oil concentration change online in order to obtain the phase transition time constant.

[0037] This invention provides an improved industrial control system for multi-unit refrigerant return and separation of fluorine pumps based on the Industrial Internet. Compared with the prior art, it has the following improvements and advantages:

[0038] 1. This solution solves this problem by configuring an edge computing acquisition unit targeting the 50-200Hz characteristic frequency band of Freon pressure pulsation. By performing spectral analysis on the signal in this specific frequency band, the system can infer the equivalent diameter of the microscopic foam and calculate the foam Reynolds number, which can be used to correct the macroscopic fluid characteristics in the semi-empirical prediction model. This is equivalent to equipping the control system with a sensor that can perceive the microscopic world, enabling it to make decisions based on more realistic and complete physical information, thereby improving the accuracy of control.

[0039] 2. The entire architecture of this solution is designed specifically for prediction; the multi-scale nested predictive control unit can output probabilistic predictions of future states, while the semi-empirical prediction model can quickly output key state prediction indicators; in particular, by solving the phase transition time constant through online fitting, the system can adaptively learn its own response delay and incorporate this delay into the prediction model; this shift from post-event remediation to pre-event prediction is the core technological advancement of this solution; the resulting effect is that the system can actively induce controllable local phase separation in specific pipe sections, realize automatic stratification and redistribution of oil, achieve a self-healing effect, and even actively induce and maintain a supercritical oscillation effect when certain boundary conditions are met, enabling the system's cooling performance coefficient to achieve instantaneous leaps, all of which are unattainable by existing passive control technologies. Attached Figure Description

[0040] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0041] Figure 1 This is a flowchart of the industrial control system for multi-unit refrigerant return and separation of fluorine pumps based on the Industrial Internet of Things, which is based on the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0043] Example 1:

[0044] Please see Figure 1 This invention provides an industrial control system for multi-unit refrigerant return and oil separation of fluorine pumps based on the Industrial Internet, comprising:

[0045] The edge computing acquisition unit is used to acquire and preprocess data for the characteristic frequency bands of Freon pressure pulsation in order to generate preprocessed data.

[0046] A multi-scale nested predictive control unit is used to receive preprocessed data and generate a set of control strategy instructions;

[0047] The system execution unit is used to receive and execute the control strategy instruction set;

[0048] In this embodiment, an industrial control system for multi-unit refrigerant return and separation based on the Industrial Internet is provided. The system deeply integrates the real-time data acquisition capabilities of the Industrial Internet with the special physical properties of refrigerants, aiming to solve the problems of oil miscibility instability and response lag faced by traditional control methods under dynamic operating conditions. The system includes an edge computing acquisition unit, a multi-scale nested predictive control unit, and a system execution unit. The three work together to form a technical closed loop from data perception to predictive decision-making and then to physical execution.

[0049] The edge computing acquisition unit is designed to acquire and pre-process physical signals in the refrigeration system with high fidelity and low latency. In this embodiment, the unit is specifically configured to handle the weak pulsating signals generated by the Freon refrigerant during pressure changes. The underlying logic is that the unit performs data acquisition and filtering pre-processing operations on the 50-200Hz characteristic frequency band of Freon pressure pulsation. The frequency band is selected because it contains nanoscale micro-foam generated by the high-speed operation of the compressor and other key pre-phase change information, which traditional sensors usually cannot effectively capture. After acquisition and pre-processing, pre-processed data containing the micro-dynamic characteristics of the system is generated.

[0050] The multi-scale nested predictive control unit is designed to receive the preprocessed data and generate the optimal control strategy based on a combination of physical models and data-driven methods. The unit integrates a complex predictive model that can simulate the full-scale physical processes from molecular interactions to macroscopic pipeline flow.

[0051] The system execution unit is responsible for receiving and executing the control strategy instruction set generated by the multi-scale nested predictive control unit. The instruction set is a series of precise control commands, such as adjusting the opening of a specific electronic expansion valve, changing the operating frequency of the compressor, or starting a specific oil separation circuit. After receiving the instruction, the system execution unit immediately converts it into actions of the corresponding physical actuators in the refrigeration system, such as valves, pumps, and compressors, thereby completing the precise control of the system state.

[0052] By constructing a closed-loop control system that can actively predict and intervene in the oil-liquid separation process, the passive response mode of traditional control methods has been changed. The system can anticipate and avoid oil miscibility instability caused by drastic load changes, ensuring reliable recovery and circulation of lubricating oil, thereby significantly improving the operational stability, energy efficiency and equipment reliability of the multi-unit fluorine pump refrigeration system under dynamic operating conditions.

[0053] Example 2

[0054] The multi-scale nested predictive control unit generates a control strategy instruction set through a three-layer nested predictive architecture; the three-layer nested predictive architecture includes:

[0055] The molecular-level prediction module is used to process real-time high-frequency temperature and pressure data in the preprocessed data to output a three-dimensional local solubility distribution field.

[0056] The mesoscale prediction module is used to receive local solubility distribution field and macroscopic flow field data from preprocessed data to generate the probability density function of oil two-phase distribution.

[0057] The system-level prediction module is used to receive the probability density function of the two-phase distribution of oil and combine it with the preset system topology diagram data to output a set of control strategy instructions;

[0058] In this embodiment, the multi-scale nested predictive control unit generates a set of control strategy instructions through a three-layer nested predictive architecture, thereby realizing deep modeling and accurate prediction of complex physical processes. The architecture decomposes a complex system-level problem into three interrelated but functionally distinct modular sub-problems, forming a complete information transmission path from microscopic physical principles to macroscopic system control.

[0059] The three-layer nested prediction architecture includes: a molecular-level prediction module, a mesoscale prediction module, and a system-level prediction module.

[0060] The molecular-level prediction module is located at the bottom layer of the architecture and directly processes the real-time high-frequency temperature and pressure data in the preprocessed data provided by the edge computing acquisition unit. Its core task is to predict the local solubility of Freon and lubricating oil in any tiny area inside the pipeline, starting from the most basic physical level. The module finally outputs a three-dimensional local solubility distribution field, which provides key physical input for the calculation of the mesoscale prediction module.

[0061] The mesoscale prediction module receives the output of the molecular-level prediction module. It receives the three-dimensional local solubility distribution field and combines it with macroscopic flow field data extracted from the preprocessed data, such as flow velocity and pressure gradient, to simulate the actual movement, coalescence, and separation behavior of oil droplets in the pipeline. The module's output is a time-evolving probability density function of the oil-liquid two-phase distribution, which describes the distribution of the oil and liquid phases at any location and at any time in the pipeline in a probabilistic form.

[0062] The system-level prediction module is the highest layer of the entire prediction architecture. It receives the probability density function of the oil-liquid two-phase distribution from the mesoscale module and combines it with a pre-defined topology diagram of the entire refrigeration system represented by a graph data structure to perform global analysis and decision-making. The module's final output is a set of control strategy instructions that can be directly executed by the system execution unit.

[0063] The three-layer nested prediction architecture brings the technical benefit of solving the problem of cross-scale modeling in a highly structured and physically self-consistent way. By decomposing the problem into three levels: molecular, mesoscopic, and system, the system can not only know what will happen and predict that oil separation will occur, but also know why it will happen, and understand the microscopic driving force, mesoscopic evolution process, and macroscopic distribution results of oil separation. This makes the final control strategy more targeted, forward-looking, and globally optimal.

[0064] Existing technologies simplify the refrigerant and lubricating oil mixture in fluorine pump systems into an ideal two-phase flow. This simplification is acceptable under stable load conditions, but it leads to control failure under dynamic conditions due to the neglect of non-ideal effects. The essential difference between this solution and the previous one is that it uses a three-layer nested predictive architecture to provide a detailed characterization of the system across three scales: molecular, mesoscopic, and macroscopic.

[0065] The molecular-level prediction module is based on a pre-set surrogate model of the potential energy surface of the Freon-oil interaction derived from quantum chemical calculations. It can accurately capture the intense nonlinear miscibility between specific refrigerants and lubricating oils. This allows the control system to no longer respond blindly, but to anticipate and utilize special physical phenomena such as reverse dissolution, turning a passive response into an active one.

[0066] The mesoscale prediction module introduces a phase separation force term into the lattice Boltzmann method, transforming molecular-level solubility information into macroscopic driving forces, which can realistically simulate the formation and transport of oil droplets.

[0067] The system-level prediction module uses a graph neural network model to process global topological information, ensuring the overall optimality of control decisions;

[0068] This inside-out modeling approach enables the control system to achieve an unprecedented depth of understanding of the physical processes, thus maintaining stable system operation even under harsh dynamic conditions with rapid load switching.

[0069] Example 3

[0070] The molecular-level prediction module outputs a three-dimensional local solubility distribution field in the following manner:

[0071] Receive real-time high-frequency temperature data, pressure data, and concentration gradient derived from the previous calculation time.

[0072] The process is based on a pre-defined surrogate model of the potential energy surface of the Freon-oil interaction to output a three-dimensional local solubility distribution field in real time.

[0073] In this embodiment, the molecular-level prediction module outputs a three-dimensional local solubility distribution field in an efficient manner. The module's input includes not only real-time high-frequency temperature and pressure data, but also a concentration gradient derived from the previous calculation time. Introducing the concentration gradient as input allows the model to take into account the influence of historical states on the current solubility, thereby capturing the dynamic change process more accurately.

[0074] The core of the module is a pre-built surrogate model of the Freon-oil interaction potential energy surface. The pre-built model is completed before the system runs. As a simplified alternative to complex physical calculations, it works by obtaining detailed interaction energy data between Freon molecules and lubricating oil molecules under different temperatures, pressures and component concentrations through costly quantum chemical calculations. The data together constitute a high-dimensional potential energy surface. Then, to meet the real-time requirements of industrial control, this complex potential energy surface is solidified into a computationally efficient surrogate model, such as a high-dimensional lookup table or a set of multivariate polynomials.

[0075] This embodiment uses Gaussian software to perform quantum chemical calculations at the B3LYP / 6-31G(d) theoretical level, scanning the interaction energy in the temperature range of 220K-350K and the pressure range of 0.1MPa-2.5MPa. The more than 10,000 discrete data points obtained are fitted into a fourth-order multivariate polynomial of the following form using the least squares method:

[0076]

[0077] Among them, S l Let T be the local solubility at any point in space; the independent variable here is a dimensionless parameter: T. * =T / T ref P * =P / P ref ,and These are respectively composed of real-time high-frequency temperature data T, pressure data P, and the concentration gradient from the previous moment. Divide by their respective reference values ​​T ref P ref and We obtain; i, j, k: summation exponents in the polynomial; a ijk The fitting coefficients are used to meet the microsecond-level real-time computing requirements. This polynomial model can be stored in the memory of the edge computing acquisition unit for quick retrieval. During system operation, the molecular-level prediction module obtains real-time data from the edge computing acquisition unit and substitutes it into the proxy model for calculation, thereby outputting a three-dimensional local solubility distribution field in real time; (T * ) i The dimensionless temperature is raised to the power of i; (P) * ) j The j-th power represents the dimensionless pressure.

[0078] In actual system operation, the molecular-level prediction module only needs to input the real-time temperature, pressure, and concentration gradient as indices or variables into the surrogate model to quickly query or calculate the local solubility S at any point in space. l (x,y,z), thus outputting a complete three-dimensional local solubility distribution field in real time;

[0079] The system successfully applies computationally expensive first-principles calculations to industrial control applications where real-time performance is critical. By pre-constructing surrogate models, the system can achieve ultra-high-speed calculations of microscopic solubility without sacrificing the accuracy of the physical model, providing solid and timely microscopic physical information input for subsequent mesoscopic and macroscopic predictions.

[0080] Example 4

[0081] The mesoscale prediction module generates the probability density function of the oil two-phase distribution in the following manner:

[0082] Determine the gradient of the chemical potential, and based on the gradient of the chemical potential, define the phase separation force term;

[0083] The phase separation force term is introduced into the fluid dynamics simulation of the lattice Boltzmann method to correct the fluid momentum;

[0084] The evolution equation of the method is solved and the particle distribution function is statistically averaged to generate the probability density function of the two-phase distribution of oil.

[0085] In this embodiment, the mesoscale prediction module generates the probability density function of the two-phase distribution of oil through an improved hydrodynamic simulation method; the core of the module is to make an innovative improvement on the classic lattice Boltzmann method, so that it can simulate the phase separation process driven by solubility differences.

[0086] To perform the simulation, the gradient of the chemical potential needs to be determined; the chemical potential μ c The change in the system free energy caused by adding one unit of a component in a multi-component system is represented by the local solubility S. l The function of chemical potential; the gradient of chemical potential This describes the rate of change of chemical potential in space, representing the driving force of mass diffusion; based on the gradient of chemical potential, a key physical quantity—the phase separation force term F—is... s Defined here, force density, mathematically is:

[0087]

[0088] The physical significance of the force term lies in quantifying the microscopic forces that drive lubricating oil molecules to aggregate from high-solubility areas to low-solubility areas due to uneven solubility in space.

[0089] Based on the above definition, the phase separation force term is introduced into the fluid dynamics simulation of the lattice Boltzmann method; in the LBM evolution equation;

[0090] g k (x+c k Δt,t+Δt)-gk (x,t)=Ω k (g)

[0091] Among them, g k Let x be the particle distribution function along the discrete velocity direction k, and c be the spatial position. k For discrete velocity, Δt: time step, t: time, Ω k (g) In the collision operator, the calculated phase separation force term F is used... s The collision process is integrated to correct fluid momentum, k: direction of discrete velocity; the correction process applies an additional force derived from intermolecular interactions to the fluid micro-particles, thereby driving the diffusion, aggregation and separation of oil droplets on a macroscopic scale;

[0092] By iteratively solving the modified evolution equation and obtaining the final particle distribution function g... k By performing statistical averaging, the macroscopic probability density function P of the oil-liquid two-phase distribution over time can be obtained. d (x,y,z,t); where P d : Probability density function of oil-liquid two-phase distribution, which describes the distribution of oil and liquid phases at any location and time in the pipeline in the form of probability; x, y, z: spatial coordinates;

[0093] The gain effect of this implementation method is that it successfully couples molecular-level chemical potential information with solubility and mesoscopic-scale fluid dynamics simulation. By introducing a phase separation force term, the model can self-consistently simulate the complete physical process from the dissolved state to the two-phase separation within a unified framework, which greatly improves the accuracy of predicting oil droplet formation and transport processes.

[0094] Example 5

[0095] The system-level prediction module outputs the control strategy instruction set in the following manner:

[0096] For each node in the system topology diagram data, calculate the statistical moments of the probability density function of the oil-liquid two-phase distribution within the physical volume corresponding to the node, and construct the initial feature vector of the node based on the statistical moments.

[0097] A graph neural network model is adopted, and the node state is updated through multi-layer information propagation. The final output layer of the graph neural network model is output as the control policy instruction set.

[0098] In this embodiment, the system-level prediction module uses a graph neural network-based method to generate a global optimal control strategy from local state information.

[0099] The module needs to perform structured processing on the input information; for each node v in the system topology diagram data G representing the entire refrigeration system, where node v can represent a pipe segment, a heat exchanger, or a valve, the module will calculate the physical space volume V corresponding to the node. v The probability density function P of the two-phase distribution of oil in the liquid d The statistical moments; based on the statistical moments, the initial feature vector of a node. It is constructed; the initial feature vector is a vector that can describe the physical state of the nodes in a digital and structured form, and its role is to serve as the starting point for subsequent graph neural network calculations;

[0100] A model employing a graph neural network (GNN) is used to process the entire graph data; the core of GNN lies in updating the node state through multi-layer information propagation; its update rule can be expressed as:

[0101]

[0102] In the formula, The feature vector of node v in the (l+1)th layer; σ is the feature vector of node v at layer l, and the data source is the calculation result or initial feature vector of the previous layer; σ is the nonlinear activation function; N(v) is the set of neighboring nodes of node v in the graph, which is determined by the system topology graph data G. and These are the learnable weight matrices in the l-th layer of the network, obtained through machine learning training on massive amounts of historical operational data; u: neighboring nodes of node v; α vu : Attention coefficient, used to dynamically adjust the influence weight of neighbor u on v; v: a node representing a component such as a pipe, heat exchanger, or valve; l: the number of layers in the network;

[0103] In this embodiment, the training dataset contains operating data from the system running continuously for 1000 hours under various dynamic operating conditions, with a data point sampling frequency of 10Hz. The root mean square error is used as the loss function during training, i.e.:

[0104]

[0105] Among them, L RMSE : Root mean square error, used as the loss function during training; N: Number of data points in the training dataset; i: Summation exponent; C s,pred The control strategy instruction set output by the model, C s,true The training process uses the Adam optimizer for iterative optimization based on expert experience or offline optimization. The initial learning rate is set to 0.001 until the loss function converges.

[0106] α vuIt is an attention coefficient used to dynamically adjust the influence weight of neighbor u on v, and its value is based on h. v and h u The correlation is obtained through dynamic calculation of the network; σ is a non-linear activation function; this process is repeated multiple times, and each layer of propagation enables each node to integrate the information of its more distant neighbors, thereby gaining a perception of the global state of the system.

[0107] After multiple layers of information propagation, the final output layer of the graph neural network model is designed to directly generate specific control commands; the output layer serves as the control strategy instruction set C. s To output, C s It contains precise operating instructions for each actuator in the system;

[0108] The gain effect brought by the implementation method is that, by utilizing the powerful graph structure data processing capability of GNN, it can fully consider the physical connection and mutual influence between various components in the system. Compared with controlling each component in isolation, GNN can make globally coordinated and optimal decisions. For example, it can predict how a certain fluctuation in the upstream will affect multiple components in the downstream and generate a set of coordinated action instructions in advance to deal with it, thereby achieving intelligent, efficient and stable control of the entire pipeline network system.

[0109] Update equations of graph neural networks Among them, h v ,h u Features of nodes v and u, W1, W2: weights, α vu Attention coefficient; σ: Application of the activation function, the logical basis is to abstract the entire refrigeration network as a topological graph; initial feature vector h of the nodes. v The probability density function P of the two-phase distribution of oil fluid output from the mesoscale simulation. d The statistical moments, consisting of mean, variance, and skewness, complete the mapping from continuous physical fields to discrete graph features; the equations, through iterative aggregation of neighborhood information, enable the decisions of each node to contain global state information, thus achieving global optimization of the control strategy.

[0110] Compared with existing technologies, this solution has achieved substantial breakthroughs in control precision, system stability, and operational efficiency.

[0111] Example 6

[0112] Statistical moments include:

[0113] The average oil concentration within the physical volume corresponding to each node;

[0114] The concentration variance within the physical volume corresponding to the node;

[0115] Skewness of the concentration distribution within the physical volume corresponding to the node;

[0116] In this embodiment, the statistical moments used to construct the initial feature vectors of the nodes are defined; by using multiple statistical moments of different orders, the distribution of oil concentration in the physical space can be more comprehensively characterized from different dimensions.

[0117] In this embodiment, statistical moments include:

[0118] The average oil concentration μ within the physical volume corresponding to the node c It is a first moment, reflecting the overall level of oil content in the region;

[0119] Concentration variance within the physical volume corresponding to the node It is the second central moment, reflecting the uniformity of oil concentration distribution; a large variance indicates uneven oil distribution, which may indicate local enrichment or depletion.

[0120] Skewness γ of the concentration distribution within the physical volume corresponding to the node c It is the third central moment, reflecting the symmetry of the oil concentration distribution; positive skewness may mean that the oil concentration is low in most areas, but there are a few extremely high concentration points; negative skewness is the opposite.

[0121] The technical advantage gained by this implementation method lies in the fact that by constructing a multi-dimensional feature vector containing mean, variance, and skewness, the description of the node state is greatly enriched. To ensure dimensional consistency, the initial feature vector of the node is constructed using... Beforehand, each statistical moment needs to be dimensionless. For example, the following method can be used:

[0122] Dimensionless average oil concentration:

[0123] Dimensionless concentration variance:

[0124] Skewness of concentration distribution: γ c

[0125] in, Dimensionless average oil concentration is the value obtained after dimensionless processing of the original average oil concentration; μ c Average oil concentration refers to the average oil concentration within the physical space volume corresponding to a node. Dimensionless concentration variance is the value obtained after the original concentration variance has been dimensionlessized. Concentration variance refers to the variance of oil concentration within the physical volume corresponding to a node. The square of the reference concentration; C ref It is a preset reference concentration; therefore, the initial feature vector becomes in, The initial feature vector of a node, used to describe the initial physical state of each node in the graph neural network; γ c The skewness of the concentration distribution refers to the symmetry of the oil concentration distribution within the physical volume corresponding to the node, and it is itself a dimensionless value; all its components are dimensionless values. This more refined and physically consistent state representation provides GNN with higher quality input, enabling it to perform more detailed pattern recognition and more accurate reasoning decisions, thereby improving the effectiveness of the final control strategy. The feature vector enables GNN not only to identify the amount of oil in a region, but also to identify the uniformity and symmetry of its spatial distribution. This more refined and comprehensive state representation provides GNN with higher quality input, enabling it to perform more detailed pattern recognition and more accurate reasoning decisions, thereby improving the effectiveness of the final control strategy.

[0126] Example 7

[0127] The multi-scale nested predictive control unit uses a semi-empirical predictive model to output dimensionless system critical state prediction indicators.

[0128] By multiplying the basic thermodynamic function, the mixing enthalpy correction function, the foam Reynolds number correction term, and the phase transition time constant function, the system's key state prediction index is obtained.

[0129] In this embodiment, as a supplementary implementation of the multi-scale nested predictive control unit, the unit can adopt an efficient semi-empirical prediction model to quickly output a dimensionless system key state prediction index χ; index χ is a comprehensive dimensionless value that characterizes a certain key performance of the system at present or in the future, such as the predicted oil separation efficiency or system stability margin.

[0130] The model construction integrates classical physical theory with empirical modifications tailored to the specific characteristics of this system. By multiplying a fundamental thermodynamic function, a mixing enthalpy correction function, a foam Reynolds number correction term, and a phase transition time constant function, the system's key state prediction index χ is obtained; its mathematical form is expressed as:

[0131] χ=f b (p,T,q)·φ h (ΔH)·ψ r (R f )·θ t (τ)

[0132] In the formula: f b (p,T,q), where p: pressure, χ: dimensionless system critical state prediction index, f b φ is a dimensionless fundamental thermodynamic function. h: Dimensionless enthalpy of mixture correction function, ψ r : Dimensionless foam Reynolds number correction term, R f θ: Foam Reynolds number, used to capture the influence of microfoam on the macroscopic flow characteristics of fluids. t φ is a dimensionless function of the phase transition time constant, where τ is the phase transition time constant used to characterize the hysteresis effect of rapid phase transition kinetics of Freon, T is temperature, and q is mass flow rate, both acquired in real time by sensors. It is a dimensionless fundamental thermodynamic function derived from the well-known thermodynamic equation of state, describing the baseline behavior of the system in quasi-equilibrium. h (ΔH), where ΔH: enthalpy of mixing is a dimensionless correction function for quantifying the thermal effects of an imperfect mixture of Freon and lubricating oil; the numerical value of the enthalpy of mixing ΔH is based on the imperfect solution model ΔH=∑ i x i H i -H ideal Perform the calculation, where x i Let H be the mole fraction of component i, and its partial molar enthalpy H. i The interaction potential surface model output by the molecular-level prediction module is obtained in real time; ΔH: enthalpy of mixing; i: composition; x i : Mole fraction of component i; ψ r (R f ), where R f The foam Reynolds number is a dimensionless correction term used to capture the influence of microscopic foams on the macroscopic flow characteristics of fluids; H i Partial molar enthalpy of component i, H ideal Enthalpy of mixing in an ideal solution model; θ t (τ), where τ: phase transition time constant is a dimensionless phase transition time constant function used to characterize the hysteresis effect of rapid phase transition dynamics of Freon; the adjustable parameters in the model are not fixed values. The setting method is as follows: first, constrain the basic shape of the model through molecular dynamics simulation results, and then use historical data collected by edge computing units to optimize through deep learning regression training to minimize the error between the predicted value and the actual observed value.

[0133] The technical advantage of this implementation is that it provides a more computationally lightweight prediction approach. Compared to the complete three-layer nested architecture simulation, the semi-empirical model simplifies the physical details but captures several core physical factors that affect the system and optimizes the model parameters through data-driven methods. This enables it to run quickly on edge devices with low computing power requirements, achieving high-frequency prediction of critical system states, and is especially suitable for control scenarios that require rapid feedback and response.

[0134] The mathematical models introduced in this scheme all have clear physical correspondences and are logically interconnected, together forming a complete description of complex physical processes.

[0135] The core equation of the semi-empirical prediction model:

[0136] χ=f b (p,T,q)·φ h (ΔH)·ψ r (R f )·θ t (τ)

[0137] Where χ: system state index, f b : Fundamental thermodynamic function, φ h Enthalpy correction, ψ r Reynolds number correction, θ t The time constant function is not derived from first principles, but rather constructed based on physical insights; its logical starting point is the well-known quasi-equilibrium thermodynamic function f. b This serves as the benchmark for system behavior under ideal conditions; the subsequent three dimensionless product terms φ h ,ψ r ,θ t It is a layer-by-layer correction of non-ideal effects in the real world, corresponding to the non-ideal mixing thermal effect between molecules, the hydrodynamic disturbance caused by micro foam, and the dynamic hysteresis effect of system state changes. This construction ensures computational efficiency while preserving the completeness of physical meaning to the greatest extent, so that the generation of the prediction index χ is both fast and accurate.

[0138] Existing control logic typically suffers from a detection lag of 200-300ms, making it unable to provide early warnings and effective responses to rapid phase changes. The entire architecture of this solution is designed specifically for prediction. The multi-scale nested predictive control unit can output probabilistic predictions of future states, while the semi-empirical prediction model can quickly output the key state prediction index χ. In particular, by solving the phase change time constant τ through online fitting, the system can adaptively learn its own response delay and incorporate the delay into the prediction model. This shift from post-event remediation to pre-event prediction is the core technological advancement of this solution. The resulting effect is that the system can actively induce controllable local phase separation in specific pipe sections, achieving automatic stratification and redistribution of oil, resulting in a self-healing effect. Furthermore, under certain boundary conditions, it can actively induce and maintain a supercritical oscillation effect, enabling the system's cooling performance coefficient to achieve instantaneous leaps, all of which are unattainable by existing passive control technologies.

[0139] In one collaborative implementation approach, the semi-empirical prediction model can be used for high-frequency real-time state monitoring and rapid response, while the computationally more computationally intensive three-layer nested prediction architecture operates at a lower frequency to output a more refined physical state distribution and periodically calibrate key parameters in the semi-empirical model, thereby achieving complementary advantages.

[0140] The process for obtaining the bubble Reynolds number corresponding to the bubble Reynolds number correction term is as follows:

[0141] The equivalent diameter of the foam was determined by high-frequency pressure pulsation spectrum analysis;

[0142] Obtain the real-time density of the mixture, the real-time viscosity of the mixture, and the average flow velocity in the pipeline;

[0143] The intermediate product value is obtained by multiplying the real-time density of the mixture, the average flow velocity in the pipe, and the equivalent diameter of the foam.

[0144] Divide the intermediate product by the real-time viscosity of the mixture to obtain the foam Reynolds number;

[0145] In this embodiment, the core parameter corresponding to the foam Reynolds number correction term is the foam Reynolds number R. f The process of obtaining the data is specified to ensure the accuracy of its physical meaning and the feasibility of the calculation.

[0146] To calculate the parameters, pulsation spectrum analysis was performed on the high-frequency pressure signal acquired by the edge computing acquisition unit to identify and separate specific frequency components generated by the vibration and rupture of microbubbles. By analyzing the spectral characteristics, the equivalent diameter d of the microbubbles in the fluid can be deduced. b ;

[0147] The inversion process is based on the Minnaert resonance model in fluid acoustics, which establishes the resonant frequency f of the bubble in the liquid. r The relationship between it and its radius R; by performing a fast Fourier transform on the high-frequency pressure pulsation signal, the main energy peak frequency f in the characteristic frequency band of 50-200Hz was identified. peak And consider it as the average resonant frequency of the foam, according to the modified Minnaert formula:

[0148]

[0149] The equivalent diameter d of the foam can be calculated in real time. b Where R: bubble radius; f peak : The average resonant frequency of the foam, identified by a fast Fourier transform of the high-frequency pressure pulsation signal; γ is the adiabatic index of the Freon gas, a known physical property parameter; P0 is the average pressure inside the pipe, which can be obtained in real time by a pressure sensor; ρ mThis represents the real-time density of the mixture;

[0150] At the same time, the system needs to obtain the real-time density ρ of the mixture. m Real-time viscosity μ of the mixture m And the average flow velocity v in the pipeline; parameters can be calculated in real time by combining data collected from basic temperature, pressure, and flow sensors with a preset state equation describing the physical properties of the Freon-oil mixture; after obtaining all necessary parameters, the real-time density ρ of the mixture is calculated. m The average flow velocity v in the pipe and the equivalent diameter d of the foam b Perform a product operation, then divide the intermediate product value by the real-time viscosity μ of the mixture. m To obtain the final bubble Reynolds number R f The calculation formula is:

[0151]

[0152] Among them, R f : Foam Reynolds number; ρ m : Real-time density of the mixture; v: Average flow velocity in the pipe; d b : Foam equivalent diameter; μ m The real-time viscosity of the mixture; the gain effect brought by the implementation method is that it proposes a method that can quantify the influence of nanoscale micro foams that cannot be measured by traditional methods online and in real time; the equivalent diameter of the foam is estimated by spectral analysis and the information is integrated into the classical Reynolds number definition, so that the semi-empirical prediction model can more accurately correct the deviation of the macroscopic properties of the fluid caused by the presence of micro foams, thereby significantly improving the prediction accuracy of the model under complex two-phase flow conditions;

[0153] Among them, the bubble Reynolds number Among them, R f : Foam Reynolds number; ρ m : Real-time density of the mixture; v: Average flow velocity in the pipe; d b : Foam equivalent diameter; μ m The innovation of the real-time viscosity of the mixture lies not in the formula itself, but in the equivalent diameter d of the foam. b The method of obtaining the equivalent diameter of nanoscale foam by performing spectral analysis on high-frequency pressure pulsations is a manifestation of combining macroscopic fluid dynamics concepts with microscopic signal processing technology, which solves the problem that existing technologies cannot perceive the influence of microscopic foam.

[0154] Existing technologies using traditional sensors cannot detect nanoscale microfoam generated by the high-speed operation of compressors, leading to signal acquisition distortion and subsequent control misjudgments. This solution solves this problem by configuring an edge computing acquisition unit targeting the 50-200Hz characteristic frequency band of Freon pressure pulsation. By performing spectral analysis on signals in a specific frequency band, the system can deduce the equivalent diameter of the microfoam and calculate the foam Reynolds number, which is then used to correct the macroscopic fluid characteristics in a semi-empirical prediction model. This is equivalent to equipping the control system with sensors capable of observing the microscopic world, enabling it to make decisions based on more realistic and complete physical information, thus improving control accuracy.

[0155] The process of obtaining the phase transition time constant function is as follows:

[0156] Monitor whether the system status undergoes a step change;

[0157] If the system state undergoes a step change, a first-order system response model is used to fit the trend of oil concentration change online in order to obtain the phase transition time constant.

[0158] In this embodiment, the process of obtaining the core parameter on which the phase transition time constant function depends—the phase transition time constant τ—is specified in order to achieve adaptive capture of the dynamic response characteristics of the system.

[0159] The process begins with continuous monitoring of the system state, especially pressure p(t) or temperature T(t), to determine whether a step change has occurred. A step change refers to a significant, step-like jump in the operating parameters of the system in a short period of time under external disturbances, such as compressor start-up and shutdown or sudden load changes. If a step change in the system state is detected, the system will automatically trigger an online fitting program.

[0160] The threshold for judging a step change is the absolute value of the rate of change of pressure or temperature within a 1-second time window. Where X represents pressure or temperature exceeding a preset value, for example, when the pressure change rate exceeds 0.1 MPa / s, the system will continuously collect oil concentration C(t) data points for the next 30 seconds after triggering, and use the Levenberg-Marquardt algorithm to compare the data points with the model C(t) = C f (1-e -t / τ A nonlinear least squares fitting is performed to obtain the phase transition time constant τ under the current operating condition.

[0161] The program employs a first-order system response model to perform online fitting of the dynamic trend of oil concentration C(t) after a step change; the mathematical form of the model is:

[0162] C(t)=C f (1-e -t / τ )

[0163] Where C(t) is the oil concentration at time t obtained through indirect measurement or model estimation; C f The final stable oil concentration when the system reaches a new equilibrium state; t is the time calculated from the time the step occurs; τ is the characteristic parameter characterizing the speed of the response process, which is the target to be solved in the fitting process; by fitting the C(t) data points over a period of time with the model using optimization algorithms such as least squares, the phase transition time constant τ under the current operating conditions can be obtained.

[0164] The gain effect of this implementation method is that it makes the phase transition time constant τ in the model no longer a fixed, preset empirical value, but an adaptive parameter that can be updated in real time according to the actual dynamic response of the system each time. This enables the semi-empirical prediction model to accurately reflect the real response speed of the system under different operating conditions and different aging degrees, thereby greatly improving the model's ability to describe the dynamic characteristics of the system and the foresight of predictive control.

[0165] The phase transition time constant τ is determined by fitting a first-order system response model C(t) = C f (1-e -t / τ ), where C(t): concentration at time t, C f : final concentration, t: time, τ: time constant, realized from online measured data; this transforms τ from a fixed empirical parameter into an adaptive variable that can reflect the current real dynamic characteristics of the system, ensuring that the model accurately tracks the time-varying nature of the system.

Claims

1. An industrial control system for multi-unit refrigerant return and separation of fluorine pumps based on the Industrial Internet, characterized in that, include: The edge computing acquisition unit is used to acquire and preprocess data for the characteristic frequency bands of Freon pressure pulsation in order to generate preprocessed data. A multi-scale nested predictive control unit is used to receive preprocessed data and generate a set of control strategy instructions; The system execution unit is used to receive and execute the control strategy instruction set; The multi-scale nested predictive control unit generates a control strategy instruction set through a three-layer nested predictive architecture; the three-layer nested predictive architecture includes: The molecular-level prediction module is used to process real-time high-frequency temperature and pressure data in the preprocessed data to output a three-dimensional local solubility distribution field. The mesoscale prediction module is used to receive local solubility distribution field and macroscopic flow field data from preprocessed data to generate the probability density function of oil two-phase distribution. The system-level prediction module is used to receive the probability density function of the two-phase distribution of oil and combine it with the preset system topology data to output a set of control strategy instructions.

2. The industrial control system for multi-unit refrigerant return and separation of fluorine pumps based on the Industrial Internet as described in claim 1, characterized in that, The molecular-level prediction module outputs a three-dimensional local solubility distribution field in the following manner: Receive real-time high-frequency temperature data, pressure data, and concentration gradient derived from the previous calculation time. The process is based on a pre-defined surrogate model of the Freon-oil interaction potential energy surface to output a three-dimensional local solubility distribution field in real time.

3. The industrial control system for multi-unit refrigerant return and separation of fluorine pumps based on the Industrial Internet as described in claim 1, characterized in that, The mesoscale prediction module generates the probability density function of the oil two-phase distribution in the following manner: Determine the gradient of the chemical potential, and based on the gradient of the chemical potential, define the phase separation force term; The phase separation force term is introduced into the fluid dynamics simulation of the lattice Boltzmann method to correct the fluid momentum; The evolution equation of the method is solved and the particle distribution function is statistically averaged to generate the probability density function of the two-phase distribution of oil.

4. The industrial control system for multi-unit refrigerant return and separation of fluorine pumps based on the Industrial Internet as described in claim 1, characterized in that, The system-level prediction module outputs the control strategy instruction set in the following manner: For each node in the system topology diagram data, calculate the statistical moments of the probability density function of the oil-liquid two-phase distribution within the physical volume corresponding to that node, and construct the initial feature vector of the node based on the statistical moments. A graph neural network model is adopted, and the node state is updated through multi-layer information propagation. The final output layer of the graph neural network model is output as the control policy instruction set.

5. The industrial control system for multi-unit refrigerant return and separation of fluorine pumps based on the Industrial Internet as described in claim 4, characterized in that, The statistical moments include: The average oil concentration within the physical volume corresponding to each node; The concentration variance within the physical volume corresponding to the node; The skewness of the concentration distribution within the physical volume corresponding to the node.

6. An industrial control system for multi-unit refrigerant return and separation of fluorine pumps based on the Industrial Internet, characterized in that: include: The edge computing acquisition unit is used to acquire and preprocess data for the characteristic frequency bands of Freon pressure pulsation in order to generate preprocessed data. A multi-scale nested predictive control unit is used to receive preprocessed data and generate a set of control strategy instructions; The system execution unit is used to receive and execute the control strategy instruction set; The multi-scale nested prediction and control unit uses a semi-empirical prediction model to output dimensionless system key state prediction indicators. The semi-empirical prediction model obtains the system's key state prediction index by multiplying the basic thermodynamic function, the mixing enthalpy correction function, the foam Reynolds number correction term, and the phase transition time constant function. The enthalpy of mixing is a dimensionless correction function for quantifying the thermal effects of non-ideal mixing of Freon and lubricating oil; the foam Reynolds number is a dimensionless correction term for the foam Reynolds number, used to capture the influence of microscopic foam on the macroscopic flow characteristics of fluids; the phase change time constant is a dimensionless function for hysteresis of the rapid phase change kinetics of Freon.

7. The industrial control system for multi-unit refrigerant return and separation of fluorine pumps based on the Industrial Internet as described in claim 6, characterized in that, The process for obtaining the foam Reynolds number corresponding to the foam Reynolds number correction term is as follows: The equivalent diameter of the foam was determined by high-frequency pressure pulsation spectrum analysis; Obtain the real-time density of the mixture, the real-time viscosity of the mixture, and the average flow velocity in the pipeline; The intermediate product value is obtained by multiplying the real-time density of the mixture, the average flow velocity in the pipe, and the equivalent diameter of the foam. Divide the intermediate product by the real-time viscosity of the mixture to obtain the foam Reynolds number.

8. The industrial control system for multi-unit refrigerant return and separation of fluorine pumps based on the Industrial Internet as described in claim 6, characterized in that, The process for obtaining the phase transition time constant function is as follows: Monitor whether the system status undergoes a step change; If the system state undergoes a step change, a first-order system response model is used to fit the trend of oil concentration change online in order to obtain the phase transition time constant.

Images